Features

you-won’t-believe-the-excuses-lawyers-have-after-getting-busted-for-using-ai

You won’t believe the excuses lawyers have after getting busted for using AI

AI, AI Citations, ai hallucinations, ai lawyer, Artificial Intelligence, chatbot, chatgpt, Features, Microsoft Copilot, Policy / /

I got hacked; I lost my login; it was a rough draft; toggling windows is hard.

Credit: Aurich Lawson | Getty Images

Credit: Aurich Lawson | Getty Images

Amid what one judge called an “epidemic” of fake AI-generated case citations bogging down courts, some common excuses are emerging from lawyers hoping to dodge the most severe sanctions for filings deemed misleading.

Using a database compiled by French lawyer and AI researcher Damien Charlotin, Ars reviewed 23 cases where lawyers were sanctioned for AI hallucinations. In many, judges noted that the simplest path to avoid or diminish sanctions was to admit that AI was used as soon as it’s detected, act humble, self-report the error to relevant legal associations, and voluntarily take classes on AI and law. But not every lawyer takes the path of least resistance, Ars’ review found, with many instead offering excuses that no judge found credible. Some even lie about their AI use, judges concluded.

Since 2023—when fake AI citations started being publicized—the most popular excuse has been that the lawyer didn’t know AI was used to draft a filing.

Sometimes that means arguing that you didn’t realize you were using AI, as in the case of a California lawyer who got stung by Google’s AI Overviews, which he claimed he took for typical Google search results. Most often, lawyers using this excuse tend to blame an underling, but clients have been blamed, too. A Texas lawyer this month was sanctioned after deflecting so much that the court had to eventually put his client on the stand after he revealed she played a significant role in drafting the aberrant filing.

“Is your client an attorney?” the court asked.

“No, not at all your Honor, just was essentially helping me with the theories of the case,” the lawyer said.

Another popular dodge comes from lawyers who feign ignorance that chatbots are prone to hallucinating facts.

Recent cases suggest this excuse may be mutating into variants. Last month, a sanctioned Oklahoma lawyer admitted that he didn’t expect ChatGPT to add new citations when all he asked the bot to do was “make his writing more persuasive.” And in September, a California lawyer got in a similar bind—and was sanctioned a whopping $10,000, a fine the judge called “conservative.” That lawyer had asked ChatGPT to “enhance” his briefs, “then ran the ‘enhanced’ briefs through other AI platforms to check for errors,” neglecting to ever read the “enhanced” briefs.

Neither of those tired old excuses hold much weight today, especially in courts that have drawn up guidance to address AI hallucinations. But rather than quickly acknowledge their missteps, as courts are begging lawyers to do, several lawyers appear to have gotten desperate. Ars found a bunch citing common tech issues as the reason for citing fake cases.

When in doubt, blame hackers?

For an extreme case, look to a New York City civil court, where a lawyer, Innocent Chinweze, first admitted to using Microsoft Copilot to draft an errant filing, then bizarrely pivoted to claim that the AI citations were due to malware found on his computer.

Chinweze said he had created a draft with correct citations but then got hacked, allowing bad actors “unauthorized remote access” to supposedly add the errors in his filing.

The judge was skeptical, describing the excuse as an “incredible and unsupported statement,” particularly since there was no evidence of the prior draft existing. Instead, Chinweze asked to bring in an expert to testify that the hack had occurred, requesting to end the proceedings on sanctions until after the court weighed the expert’s analysis.

The judge, Kimon C. Thermos, didn’t have to weigh this argument, however, because after the court broke for lunch, the lawyer once again “dramatically” changed his position.

“He no longer wished to adjourn for an expert to testify regarding malware or unauthorized access to his computer,” Thermos wrote in an order issuing sanctions. “He retreated” to “his original position that he used Copilot to aid in his research and didn’t realize that it could generate fake cases.”

Possibly more galling to Thermos than the lawyer’s weird malware argument, though, was a document that Chinweze filed on the day of his sanctions hearing. That document included multiple summaries preceded by this text, the judge noted:

Some case metadata and case summaries were written with the help of AI, which can produce inaccuracies. You should read the full case before relying on it for legal research purposes.

Thermos admonished Chinweze for continuing to use AI recklessly. He blasted the filing as “an incoherent document that is eighty-eight pages long, has no structure, contains the full text of most of the cases cited,” and “shows distinct indications that parts of the discussion/analysis of the cited cases were written by artificial intelligence.”

Ultimately, Thermos ordered Chinweze to pay $1,000, the most typical fine lawyers received in the cases Ars reviewed. The judge then took an extra non-monetary step to sanction Chinweze, referring the lawyer to a grievance committee, “given that his misconduct was substantial and seriously implicated his honesty, trustworthiness, and fitness to practice law.”

Ars could not immediately reach Chinweze for comment.

Toggling windows on a laptop is hard

In Alabama, an attorney named James A. Johnson made an “embarrassing mistake,” he said, primarily because toggling windows on a laptop is hard, US District Judge Terry F. Moorer noted in an October order on sanctions.

Johnson explained that he had accidentally used an AI tool that he didn’t realize could hallucinate. It happened while he was “at an out-of-state hospital attending to the care of a family member recovering from surgery.” He rushed to draft the filing, he said, because he got a notice that his client’s conference had suddenly been “moved up on the court’s schedule.”

“Under time pressure and difficult personal circumstance,” Johnson explained, he decided against using Fastcase, a research tool provided by the Alabama State Bar, to research the filing. Working on his laptop, he opted instead to use “a Microsoft Word plug-in called Ghostwriter Legal” because “it appeared automatically in the sidebar of Word while Fastcase required opening a separate browser to access through the Alabama State Bar website.”

To Johnson, it felt “tedious to toggle back and forth between programs on [his] laptop with the touchpad,” and that meant he “unfortunately fell victim to the allure of a new program that was open and available.”

Moorer seemed unimpressed by Johnson’s claim that he understood tools like ChatGPT were unreliable but didn’t expect the same from other AI legal tools—particularly since “information from Ghostwriter Legal made it clear that it used ChatGPT as its default AI program,” Moorer wrote.

The lawyer’s client was similarly put off, deciding to drop Johnson on the spot, even though that risked “a significant delay of trial.” Moorer noted that Johnson seemed shaken by his client’s abrupt decision, evidenced by “his look of shock, dismay, and display of emotion.”

And switching to a new lawyer could eat up more of that money. Moorer further noted that Johnson seemingly let AI do his homework while working on behalf of the government. But as the judge noted, “public funds for appointed counsel are not a bottomless well and are limited resource.”

“It has become clear that basic reprimands and small fines are not sufficient to deter this type of misconduct because if it were, we would not be here,” Moorer concluded.

Ruling that Johnson’s reliance on AI was “tantamount to bad faith,” Moorer imposed a $5,000 fine. The judge also would have “considered potential disqualification, but that was rendered moot” since Johnson’s client had already dismissed him.

Asked for comment, Johnson told Ars that “the court made plainly erroneous findings of fact and the sanctions are on appeal.”

Plagued by login issues

As a lawyer in Georgia tells it, sometimes fake AI citations may be filed because a lawyer accidentally filed a rough draft instead of the final version.

Other lawyers claim they turn to AI as needed when they have trouble accessing legal tools like Westlaw or LexisNexis.

For example, in Iowa, a lawyer told an appeals court that she regretted relying on “secondary AI-driven research tools” after experiencing “login issues her with her Westlaw subscription.” Although the court was “sympathetic to issues with technology, such as login issues,” the lawyer was sanctioned, primarily because she only admitted to using AI after the court ordered her to explain her mistakes. In her case, however, she got to choose between paying a minimal $150 fine or attending “two hours of legal ethics training particular to AI.”

Less sympathetic was a lawyer who got caught lying about the AI tool she blamed for inaccuracies, a Louisiana case suggested. In that case, a judge demanded to see the research history after a lawyer claimed that AI hallucinations came from “using Westlaw Precision, an AI-assisted research tool, rather than Westlaw’s standalone legal database.”

It turned out that the lawyer had outsourced the research, relying on a “currently suspended” lawyer’s AI citations, and had only “assumed” the lawyer’s mistakes were from Westlaw’s AI tool. It’s unclear what tool was actually used by the suspended lawyer, who likely lost access to a Westlaw login, but the judge ordered a $1,000 penalty after the lawyer who signed the filing “agreed that Westlaw did not generate the fabricated citations.”

Judge warned of “serial hallucinators”

Another lawyer, William T. Panichi in Illinois, has been sanctioned at least three times, Ars’ review found.

In response to his initial penalties ordered in July, he admitted to being tempted by AI while he was “between research software.”

In that case, the court was frustrated to find that the lawyer had contradicted himself, and it ordered more severe sanctions as a result.

Panichi “simultaneously admitted to using AI to generate the briefs, not doing any of his own independent research, and even that he ‘barely did any personal work [him]self on this appeal,’” the court order said, while also defending charging a higher fee—supposedly because this case “was out of the ordinary in terms of time spent” and his office “did some exceptional work” getting information.

The court deemed this AI misuse so bad that Panichi was ordered to disgorge a “payment of $6,925.62 that he received” in addition to a $1,000 penalty.

“If I’m lucky enough to be able to continue practicing before the appellate court, I’m not going to do it again,” Panichi told the court in July, just before getting hit with two more rounds of sanctions in August.

Panichi did not immediately respond to Ars’ request for comment.

When AI-generated hallucinations are found, penalties are often paid to the court, the other parties’ lawyers, or both, depending on whose time and resources were wasted fact-checking fake cases.

Lawyers seem more likely to argue against paying sanctions to the other parties’ attorneys, hoping to keep sanctions as low as possible. One lawyer even argued that “it only takes 7.6 seconds, not hours, to type citations into LexisNexis or Westlaw,” while seemingly neglecting the fact that she did not take those precious seconds to check her own citations.

The judge in the case, Nancy Miller, was clear that “such statements display an astounding lack of awareness of counsel’s obligations,” noting that “the responsibility for correcting erroneous and fake citations never shifts to opposing counsel or the court, even if they are the first to notice the errors.”

“The duty to mitigate the harms caused by such errors remains with the signor,” Miller said. “The sooner such errors are properly corrected, either by withdrawing or amending and supplementing the offending pleadings, the less time is wasted by everyone involved, and fewer costs are incurred.”

Texas US District Judge Marina Garcia Marmolejo agreed, explaining that even more time is wasted determining how other judges have responded to fake AI-generated citations.

“At one of the busiest court dockets in the nation, there are scant resources to spare ferreting out erroneous AI citations in the first place, let alone surveying the burgeoning caselaw on this subject,” she said.

At least one Florida court was “shocked, shocked” to find that a lawyer was refusing to pay what the other party’s attorneys said they were owed after misusing AI. The lawyer in that case, James Martin Paul, asked to pay less than a quarter of the fees and costs owed, arguing that Charlotin’s database showed he might otherwise owe penalties that “would be the largest sanctions paid out for the use of AI generative case law to date.”

But caving to Paul’s arguments “would only benefit serial hallucinators,” the Florida court found. Ultimately, Paul was sanctioned more than $85,000 for what the court said was “far more egregious” conduct than other offenders in the database, chastising him for “repeated, abusive, bad-faith conduct that cannot be recognized as legitimate legal practice and must be deterred.”

Paul did not immediately respond to Ars’ request to comment.

Michael B. Slade, a US bankruptcy judge in Illinois, seems to be done weighing excuses, calling on all lawyers to stop taking AI shortcuts that are burdening courts.

“At this point, to be blunt, any lawyer unaware that using generative AI platforms to do legal research is playing with fire is living in a cloud,” Slade wrote.

This story was updated on November 11 to clarify a judge’s comments on misuse of public funds.

Photo of Ashley Belanger

Ashley is a senior policy reporter for Ars Technica, dedicated to tracking social impacts of emerging policies and new technologies. She is a Chicago-based journalist with 20 years of experience.

You won’t believe the excuses lawyers have after getting busted for using AI Read More »

nasa-is-kind-of-a-mess:-here-are-the-top-priorities-for-a-new-administrator

NASA is kind of a mess: Here are the top priorities for a new administrator

challenges, Features, isaacman, NASA, Space / /

“He inevitably will have to make tough calls.”

Jared Isaacman, right, led the crew of Polaris Dawn, which performed the first private spacewalk. Credit: Polaris Dawn

Jared Isaacman, right, led the crew of Polaris Dawn, which performed the first private spacewalk. Credit: Polaris Dawn

After a long summer and fall of uncertainty, private astronaut Jared Isaacman has been renominated to lead NASA, and there appears to be momentum behind getting him confirmed quickly as the space agency’s 15th administrator. It is possible, although far from a lock, the Senate could finalize his nomination before the end of this year.

It cannot happen soon enough.

The National Aeronautics and Space Administration is, to put it bluntly, kind of a mess. This is not meant to disparage the many fine people who work at NASA. But years of neglect, changing priorities, mismanagement, creeping bureaucracy, meeting bloat, and other factors have taken their toll. NASA is still capable of doing great things. It still inspires. But it needs a fresh start.

“Jared has already garnered tremendous support from nearly everyone in the space community,” said Lori Garver, who served as NASA’s deputy administrator under President Obama. “This should give him a tail wind as he inevitably will have to make tough calls.”

Garver worked for a Democratic administration, and it’s notable that Isaacman has admirers from across the political spectrum, from left-leaning space advocates to right-wing influencers. A decade and a half ago, Garver led efforts to get NASA to more fully embrace commercial space. In some ways, Isaacman will seek to further this legacy, and Garver knows all too well how difficult it is to change the sprawling space agency and beat back entrenched contractors.

“Expectations are high, yet the challenge of marrying outsized goals to greatly reduced budget guidance from his administration remains,” Garver said. “It will be difficult to deliver on accelerating Artemis, transitioning to commercial LEO destinations, starting a serious nuclear electric propulsion program for Mars transportation, and attracting non-government funding for science missions. He’s coming in with a lot of support, which he will need in the current divisive political environment.”

Here’s a rundown of some of the challenges Isaacman must overcome to be a successful administrator.

A shrunken NASA

At the beginning of this year, the civil servant workforce at the space agency numbered about 18,000 people. NASA said that about 3,870 employees exited this year under various deferred resignation, early retirement, or buyout programs. After subtracting another 500 employees who left through normal attrition, NASA’s headcount will be down by 20 to 25 percent by the end of this year.

The question is how impactful these losses are. A number of the departures were from senior positions, leaving important divisions—such as Astrophysics—with acting directors and interim people in key positions. Some people who left were nearing retirement, and this may ultimately benefit the space agency by allowing younger people to bring new energy to the mission.

Yet there are very real concerns about NASA’s ability to retain its best people. As the commercial space industry grows around some of its key centers, including Alabama, Florida, and Texas, these companies cherry-pick the best NASA engineers by offering higher salaries and stock options. These engineers, in turn, know who to hire at the local field centers who are most promising.

This brain drain diminishes the engineering excellence at NASA. Can Isaacman do more with less?

Very low morale

Isaacman also arrives after what has essentially been a lost year for NASA.

Imagine you’re a NASA employee. You came to the agency to lead exploration of the Solar System and beyond. Then the second Trump administration shows up and demands widespread workforce cuts. The White House subsequently also proposes a 25 percent hit to the space agency’s budget and draconian cuts for NASA’s science programs.

Then, to cap off the spring of 2025, Isaacman’s nomination was pulled for purely political reasons. Not everyone at NASA liked Isaacman. There was genuine concern that he would shake things up and rattle cages. But Isaacman was also perceived as young, dynamic, and well-liked by the broader space community. He genuinely wanted to see NASA succeed. And then—poof—he’s gone. This only exacerbated uncertainty about the agency’s future.

Interim NASA Administrator Sean Duffy provides remarks at a briefing prior to the Crew 11 launch in August.

Credit: NASA

Interim NASA Administrator Sean Duffy provides remarks at a briefing prior to the Crew 11 launch in August. Credit: NASA

Isaacman’s de-nomination was followed by the appointment of Sean Duffy, a former reality TV star serving as the Secretary of Transportation, to lead NASA on an interim basis. Duffy was a wild card, but it soon became clear he saw NASA as a vehicle to further his political career. And even if Duffy had been focused on solutions, he knew little about space and already had a full-time job leading the Department of Transportation. NASA employees are not fools. They saw this and understood this move’s implications.

Finally, in a coup de grâce, the government shut down on October 1. The majority of NASA’s civil servant workforce has been sitting at home for six weeks, not getting paid, not exploring, and wondering just what the hell they’re doing working for NASA.

Arte-miss?

As NASA has struggled this year, China has made demonstrable progress in its lunar program. It is now probable that China’s Lanyue lander will put humans on the lunar surface by or before the year 2030, likely beating NASA in its return to the Moon with the Artemis Program.

NASA’s lunar program was created during the first Trump administration, but then NASA leader Jim Bridenstine was unable to secure enough funding (remember the whole Pell Grant fiasco?) before he left office in early 2021. This left NASA without the resources it needed to build a management team to lead the program and support key elements, including a lander and lunar spacesuits.

These problems more or less persisted under President Joe Biden and his NASA Administrator, Bill Nelson. From 2021 to 2024, the leaders of NASA essentially said everything was fine and that a lunar landing by 2026 was on track. When reporters, including myself, would ask the leaders of the Artemis Program, we were effectively shouted down.

For example, in January 2024, I pressed NASA’s chief of deep space exploration, Jim Free, about the non-viability of a 2026 human landing date.

“It’s interesting because we have 11 people in industry on here that have signed contracts to meet those dates,” Free replied during a teleconference, which included representatives from SpaceX, Axiom, and the other companies. “So from my perspective, the people in industry are here today saying we support it. We’ve signed contracts to those dates on the government side based on the technical details that they’ve given us, that our technical teams have come forward with.”

A shorter version of that might be: “Shut up, we know what we’re doing.”

NASA has already delayed the lunar landing officially to 2027. And no one believes that date is real. One of Isaacman’s first jobs will be to conduct an honest assessment of where the Artemis Program truly is and to rapidly take steps to get it on track. I think we can be confident he will do so with eyes wide open.

Human Landing System

So what will he do about this? The biggest challenge involves the Human Landing System (HLS), a necessary component to get humans to the surface from lunar orbit and back.

Ars explored how NASA found itself in this predicament in a long article published in early October. As for what to do now, NASA basically has two realistic options going forward. It can light a fire under SpaceX to prioritize the HLS component of its Starship program, and possibly adopt a simplified architecture. Or it can work with Blue Origin to develop to a human system using its Blue Moon Mk. 1 lander (originally intended for cargo) and a modified Mk. 1 lander for ascent purposes. (Blue says it is game). Beyond that, there is no hardware in work that could possibly accommodate a landing before 2030.

Duffy initially blustered about American capabilities. Repeatedly, he said, “We are going to beat the Chinese to the Moon.” It sounded good, but it underlined his inexperience with spaceflight because it was just not true.

Less than a month ago, Duffy changed his tune. He blamed SpaceX and its Starship vehicle for delays to Artemis, and he said he was “opening up” the lander competition. The problem is that Duffy’s solution was to raise the prospect of a “government option” lunar lander. He had been having discussions with Lockheed Martin, Northrop Grumman, and others about the possibility of issuing a cost-plus contract to build a smaller lunar lander in 30 months.

An artist’s illustration of multiple Starships on the lunar surface, with a Moon base in the background.

Credit: SpaceX

An artist’s illustration of multiple Starships on the lunar surface, with a Moon base in the background. Credit: SpaceX

Duffy should have known that this timeline was completely unrealistic. Moreover, a rapidly built lunar lander (think five years, at a bare minimum) would likely cost on the order of $20 billion, which NASA did not have. But no one in his inner circle, including Amit Kshatriya, NASA’s associate administrator, was telling him that. They were encouraging him.

Isaacman is not going to be snowed under by this kind of (preposterous) proposal. Most likely, he will push SpaceX to prioritize HLS and be eager to work with Blue Origin to develop a human lander based on Mk. 1 technology.

His first call as administrator may well be to Blue Origin founder Jeff Bezos.

Commercial LEO Destinations

Another looming problem involves commercial space stations in low-Earth orbit, which are supposed to be flying before the end of 2030 when the International Space Station is due to be retired.

There is much uncertainty over whether the primary companies involved in this effort—be it for financial, technical, regulatory, or other reasons—will be able to launch and test space stations by 2030 in order to allow NASA to maintain a continuous presence in low-Earth orbit. The main contractors are Axiom Space, Voyager Technologies, Blue Origin, and Vast Space.

This is one area in which Duffy took action. In August, he signed a document that implemented major changes to the Commercial LEO Destinations program. One of the biggest shifts was a lowering of the minimum requirements. Instead of fully operational stations, the new directive required only the capability to support four astronauts for 1-month increments in low-Earth orbit.

However, it is unclear that Duffy fully understood what he was signing, because there was an immediate pushback. Moreover, prior to the government shutdown, there was a lot of discussion about ripping up the directive and reverting to the old rules for commercial space stations. Everyone in the industry is scratching their heads about what comes next.

In the meantime, the space station companies are trying to raise funds, design stations for uncertain requirements, and prepare for competition for the next phase of NASA awards. This program needs more funding, clarity, and urgency for it to be successful.

Earth science

In recent days, there has been some excellent reporting about the fate of Earth science at NASA, which is part of the space agency’s core mission. Space.com published a long feature article about the Trump administration’s efforts to undermine Maryland’s Goddard Space Flight Center, which is NASA’s oldest field center.

Goddard houses the largest Earth science workforce at the agency, and its study of climate change is at odds with the policy positions of the Trump administration and many members of a Republican-controlled Congress. The result has been steep funding cuts, canceled missions, and closed buildings.

One of Isaacman’s most challenging jobs will be to balance support for Earth science while also placating an administration that frankly does not want to publish reports about how human activity is warming the planet.

In remarks on the social media site X, Isaacman recently said he wanted to expand commercial partnerships to science missions. “Better to have 10 x $100 million missions and a few fail than a single overdue and costly $1B+ mission,” he wrote. Isaacman said NASA should also buy more Earth data from providers like Planet and BlackSky, which already have satellites in orbit.

“Why build bespoke satellites at greater cost and delay when you could pay for the data as needed from existing providers?” he asked.

Planetary science

Another area of concern is planetary science. When one picks apart Trump’s budget priorities, there are two clear and disturbing trends.

The first is that there are no significant planetary science missions in the pipeline after the ambitious Dragonfly mission, which is scheduled to launch to Titan in July 2028. It becomes difficult to escape the reality that this administration is not prioritizing any mission that launches after Trump leaves office in January 2029. As a result, after Dragonfly, the planetary pipeline is running low.

Another major concern is the fate of the famed Jet Propulsion Laboratory in California. The lab laid off 550 people last month, which followed previous cuts. The center director, Laurie Leshin, stepped down on June 1. With the Mars Sample Return mission on hold, and quite possibly canceled, the future of NASA’s premier planetary science mission center is cloudy.

A view of the control room at NASA’s Jet Propulsion Laboratory in California.

Credit: NASA

A view of the control room at NASA’s Jet Propulsion Laboratory in California. Credit: NASA

Isaacman has said he has never “remotely suggested” that NASA could do without the Jet Propulsion Laboratory.

“Personally, I have publicly defended programs like the Chandra X-ray Observatory, offered to fund a Hubble reboost mission, and anything suggesting that I am anti-science or want to outsource that responsibility is simply untrue,” he wrote on X.

That is likely true, but charting a bright course for the future of planetary science, on a limited budget, will be a major challenge for the new administrator.

New initiatives

All of the above concerns NASA’s existing challenges. But Isaacman will certainly want to make his own mark. This is likely to involve a spaceflight technology he considers to be the missing link in charting a course for humans to explore the Solar System beyond the Moon: nuclear electric propulsion.

As he explained to Ars earlier this year, Isaacman’s signature issue was going to be a full-bore push into nuclear electric propulsion.

“We would have gone right to a 100-kilowatt test vehicle that we would send somewhere inspiring with some great cameras,” he said. “Then we are going right to megawatt class, inside of four years, something you could dock a human-rated spaceship to, or drag a telescope to a Lagrange point and then return, big stuff like that. The goal was to get America underway in space on nuclear power.”

Another key element of this plan is that it would give some of NASA’s field centers, including Marshall Space Flight Center, important work to do after the seemingly inevitable cancellation of the Space Launch System rocket.

Standing up new programs, and battling against existing programs that have strong backing in Congress and industry, will require all of the diplomatic skill and force of personality Isaacman can muster.

We will soon find out if he has the right stuff.

Photo of Eric Berger

Eric Berger is the senior space editor at Ars Technica, covering everything from astronomy to private space to NASA policy, and author of two books: Liftoff, about the rise of SpaceX; and Reentry, on the development of the Falcon 9 rocket and Dragon. A certified meteorologist, Eric lives in Houston.

NASA is kind of a mess: Here are the top priorities for a new administrator Read More »

mark-zuckerberg’s-illegal-school-drove-his-neighbors-crazy

Mark Zuckerberg’s illegal school drove his neighbors crazy

Features, mark zuckerberg, Meta, Policy, syndication, ZOning / /

Neighbors complained about noise, security guards, and hordes of traffic.

An entrance to Mark Zuckerberg’s compound in Palo Alto, California. Credit: Loren Elliott/Redux

The Crescent Park neighborhood of Palo Alto, California, has some of the best real estate in the country, with a charming hodgepodge of homes ranging in style from Tudor revival to modern farmhouse and contemporary Mediterranean. It also has a gigantic compound that is home to Mark Zuckerberg, his wife Priscilla Chan, and their daughters Maxima, August, and Aurelia. Their land has expanded to include 11 previously separate properties, five of which are connected by at least one property line.

The Zuckerberg compound’s expansion first became a concern for Crescent Park neighbors as early as 2016, due to fears that his purchases were driving up the market. Then, about five years later, neighbors noticed that a school appeared to be operating out of the Zuckerberg compound. This would be illegal under the area’s residential zoning code without a permit. They began a crusade to shut it down that did not end until summer 2025.

WIRED obtained 1,665 pages of documents about the neighborhood dispute—including 311 records, legal filings, construction plans, and emails—through a public record request filed to the Palo Alto Department of Planning and Development Services. (Mentions of “Zuckerberg” or “the Zuckerbergs” appear to have been redacted. However, neighbors and separate public records confirm that the property in question belongs to the family. The names of the neighbors who were in touch with the city were also redacted.)

The documents reveal that the school may have been operating as early as 2021 without a permit to operate in the city of Palo Alto. As many as 30 students might have enrolled, according to observations from neighbors. These documents also reveal a wider problem: For almost a decade, the Zuckerbergs’ neighbors have been complaining to the city about noisy construction work, the intrusive presence of private security, and the hordes of staffers and business associates causing traffic and taking up street parking.

Over time, neighbors became fed up with what they argued was the city’s lack of action, particularly with respect to the school. Some believed that the delay was because of preferential treatment to the Zuckerbergs. “We find it quite remarkable that you are working so hard to meet the needs of a single billionaire family while keeping the rest of the neighborhood in the dark,” reads one email sent to the city’s Planning and Development Services Department in February. “Just as you have not earned our trust, this property owner has broken many promises over the years, and any solution which depends on good faith behavioral changes from them is a failure from the beginning.”

Palo Alto spokesperson Meghan Horrigan-Taylor told WIRED that the city “enforces zoning, building, and life safety rules consistently, without regard to who owns a property.” She also refuted the claim that neighbors were kept in the dark, claiming that the city’s approval of construction projects at the Zuckerberg properties “were processed the same way they are for any property owner.” She added that, though some neighbors told the city they believe the Zuckerbergs received “special treatment,” that is not accurate.

“Staff met with residents, conducted site visits, and provided updates by phone and email while engaging the owner’s representative to address concerns,” Horrigan-Taylor said. “These actions were measured and appropriate to abate the unpermitted use and responsive to neighborhood issues within the limits of local and state law.”

According to The New York Times, which first reported on the school’s existence, it was called “Bicken Ben School” and shared a name with one of the Zuckerbergs’ chickens. The listing for Bicken Ben School, or BBS for short, in a California Department of Education directory claims the school opened on October 5, 2022. This, however, is the year after neighbors claim to have first seen it operating. It’s also two and a half years after Sara Berge—the school’s point of contact, per documents WIRED obtained from the state via public record request—claims to have started her role as “head of school” for a “Montessori pod” at a “private family office” according to her LinkedIn profile, which WIRED viewed in September and October. Berge did not respond to a request to comment.

Between 2022 and 2025, according to the documents Bicken Ben filed to the state, the school grew from nine to 14 students ranging from 5 to 10 years old. Neighbors, however, estimated that they observed 15 to 30 students. Berge similarly claimed on her LinkedIn profile to have overseen “25 children” in her job. In a June 2025 job listing for “BBS,” the school had a “current enrollment of 35–40 students and plans for continued growth,” which the listing says includes a middle school.

In order for the Zuckerbergs to run a private school on their land, which is in a residential zone, they need a “conditional use” permit from the city. However, based on the documents WIRED obtained, and Palo Alto’s public database of planning applications, the Zuckerbergs do not appear to have ever applied for or received this permit.

Per emails obtained by WIRED, Palo Alto authorities told a lawyer working with the Zuckerbergs in March 2025 that the family had to shut down the school on its compound by June 30. A state directory lists BBS, the abbreviation for Bicken Ben School, as having operated until August 18, and three of Zuckerberg’s neighbors—who all requested anonymity due to the high-profile nature of the family—confirmed to WIRED in late September that they had not seen or heard students being dropped off and picked up on weekdays in recent weeks.

However, Zuckerberg family spokesperson Brian Baker tells WIRED that the school didn’t close, per se. It simply moved. It’s not clear where it is now located, or whether the school is operating under a different name.

In response to a detailed request for comment, Baker provided WIRED with an emailed statement on behalf of the Zuckerbergs. “Mark, Priscilla and their children have made Palo Alto their home for more than a decade,” he said. “They value being members of the community and have taken a number of steps above and beyond any local requirements to avoid disruption in the neighborhood.”

“Serious and untenable”

By the fall of 2024, Zuckerberg’s neighbors were at their breaking point. At some point in mid-2024, according to an email from then mayor Greer Stone, a group of neighbors had met with Stone to air their grievances about the Zuckerberg compound and the illegal school they claimed it was operating. They didn’t arrive at an immediate resolution.

In the years prior, the city had received several rounds of complaints about the Zuckerberg compound. Complaints for the address of the school were filed to 311, the nationwide number for reporting local non-emergency issues, in February 2019, September 2021, January 2022, and April 2023. They all alleged that the property was operating illegally under city code. Both were closed by the planning department, which found no rule violations. An unknown number of additional complaints, mentioned in emails among city workers, were also made between 2020 and 2024—presumably delivered via phone calls, in person, or to city departments not included in WIRED’s public record request.

In December 2020, building inspection manager Korwyn Peck wrote to code enforcement officer Brian Reynolds about an inspection he attempted to conduct around the Zuckerberg compound, in response to several noise and traffic complaints from neighbors. He described that several men in SUVs had gathered to watch him, and a tense conversation with one of them had ensued. “This appears to be a site that we will need to pay attention to,” Peck wrote to Reynolds.

“We have all been accused of ‘not caring,’ which of course is not true,” Peck added. “It does appear, however, with the activity I observed tonight, that we are dealing with more than four simple dwellings. This appears to be more than a homeowner with a security fetish.”

In a September 11, 2024, email to Jonathan Lait, Palo Alto’s director of planning and development services and Palo Alto city attorney Molly Stump, one of Zuckerberg’s neighbors alleged that since 2021, “despite numerous neighborhood complaints” to the city of Palo Alto, including “multiple code violation reports,” the school had continued to grow. They claimed that a garage at the property had been converted into another classroom, and that an increasing number of children were arriving each day. Lait and Stump did not respond to a request to comment.

“The addition of daily traffic from the teachers and parents at the school has only exacerbated an already difficult situation,” they said in the email, noting that the neighborhood has been dealing with an “untenable traffic” situation for more than eight years.

They asked the city to conduct a formal investigation into the school on Zuckerberg’s property, adding that their neighbors are also “extremely concerned” about the school, and “are willing to provide eyewitness accounts in support of this complaint.”

Over the next week, another neighbor forwarded this note to all six Palo Alto city council members, as well as then mayor Stone. One of these emails described the situation as “serious” and “untenable.”

“We believe the investigation should be swift and should yield a cease and desist order,” the neighbor wrote.

Lait responded to the neighbor who sent the original complaint on October 15, claiming that he’d had an “initial call” with a “representative” of the property owners and that he was directing the city’s code enforcement staff to reexamine the property.

On December 11, 2024, the neighbor claimed that since one of their fellow neighbors had spoken to a Zuckerberg representative, and the representative had allegedly admitted that there was a school on the property, “it seems like an open and shut case.”

“Our hope is that there is an equal process in place for all residents of Palo Alto regardless of wealth or stature,” the neighbor wrote. “It is hard to imagine that this kind of behavior would be ignored in any other circumstance.”

That same day, Lait told Christine Wade, a partner at SSL Law Firm—who, in an August 2024 email thread, said she was “still working with” the Zuckerberg family—that the Zuckerbergs lacked the required permit to run a school in a residential zone.

“Based on our review of local and state law, we believe this use constitutes a private school use in a residential zone requiring a conditional use permit,” Lait wrote in an email to Wade. “We also have not found any state preemptions that would exclude a use like this from local zoning requirements.” Lait added that a “next step,” if a permit was not obtained, would be sending a cease and desist to the property owner.

According to several emails, Wade, Lait, and Mark Legaspi, CEO of the Zuckerberg family office called West 10, went on to arrange an in-person meeting at City Hall on January 9. (This is the first time that the current name of the Zuckerberg family office, West 10, has been publicly disclosed. The office was previously called West Street.) Although WIRED did not obtain notes from the meeting, Lait informed the neighbor on January 10 that he had told the Zuckerbergs’ “representative” that the school would need to shut down if it didn’t get a conditional use permit or apply for that specific permit.

Lait added that the representative would clarify what the family planned to do in about a week; however, he noted that if the school were to close, the city may give the school a “transition period” to wind things down. Wade did not respond to a request for comment.

“At a minimum, give us extended breaks”

There was another increasingly heated conversation happening behind the scenes. On February 3 of this year, at least one neighbor met with Jordan Fox, an employee of West 10.

It’s unclear exactly what happened at this meeting, or if the neighbor who sent the September 11 complaint was in attendance. But a day after the meeting with Fox, two additional neighbors added their names to the September 11 complaint, per an email to Lait.

On February 12, a neighbor began an email chain with Fox. This email was forwarded to Planning Department officials two months later. The neighbor, who seemingly attended the meeting, said they had “connected” with fellow neighbors “to review and revise” an earlier list of 14 requests that had been reportedly submitted to the Zuckerbergs at some previous point. The note does not specify the contents of this original list of requests, but of the 19 neighbors who originally contributed to it, they claimed that 15 had contributed to the revised list.

The email notes that the Zuckerbergs had been “a part of our neighborhood for many years,” and that they “hope that this message will start an open and respectful dialogue,” built upon the “premise of how we all wish to be treated as neighbors.”

“Our top requests are to minimize future disruption to the neighborhood and proactively manage the impact of the many people who are affiliated with you,” the email says. This includes restricting parking by “security guards, contractors, staff, teachers, landscapers, visitors, etc.” In the event of major demolitions, concrete pours, or large parties, the email asks for advance notice, and for dedicated efforts to “monitor and mitigate noise.”

The email also asks the Zuckerbergs to, “ideally stop—but at a minimum give us extended breaks from—the acquisition, demolition and construction cycle to let the neighborhood recover from the last eight years of disruption.”

At this point, the email requests that the family “abide by both the letter and the spirit of Palo Alto” by complying with city code about residential buildings.

Specifically, it asks the Zuckerbergs to get a use permit for the compound’s school and to hold “a public hearing for transparency.” It also asks the family to not expand its compound any further. “We hope this will help us get back the quiet, attractive residential neighborhood that we all loved so much when we chose to move here.”

In a follow-up on March 4, Fox acknowledged the “unusual” effects that come with being neighbors with Mark Zuckerberg and his family.

“I recognize and understand that the nature of our residence is unique given the profile and visibility of the family,” she wrote. “I hope that as we continue to grow our relationship with you over time, you will increasingly enjoy the benefits of our proximity—e.g., enhanced safety and security, shared improvements, and increased property values.”

Fox said that the Zuckerbergs instituted “a revised parking policy late last year” that should address their concerns, and promised to double down on efforts to give advanced notice about construction, parties, and other potential disruptions.

However, Fox did not directly address the unpermitted school and other nonresidential activities happening at the compound. She acknowledged that the compound has “residential support staff” including “childcare, culinary, personal assistants, property management, and security,” but said that they have “policies in place to minimize their impact on the neighborhood.”

It’s unclear if the neighbor responded to Fox.

“You have not earned our trust”

While these conversations were happening between Fox and Zuckerberg’s neighbors, Lait and others at the city Planning Department were scrambling to find a solution for the neighbor who complained on September 11, and a few other neighbors who endorsed the complaint in September and February.

Starting in February, one of these neighbors took the lead on following up with Lait. They asked him for an update on February 11, and heard back a few days later. He didn’t have any major updates, “but after conversations with the family’s representatives, he said he was exploring whether a “subset of children” could continue to come to the school sometimes for “ancillary” uses.

“I also believe a more nuanced solution is warranted in this case,” Lait added. Ideally, such a solution would respond to the neighbors’ complaints, but allow the Zuckerbergs to “reasonably be authorized by the zoning code.”

The neighbor wasn’t thrilled. The next day, they replied and called the city’s plan “unsatisfactory.”

“The city’s ‘nuanced solution’ in dealing with this serial violator has led to the current predicament,” they said (referring to the nuanced solution Lait mentioned in his last email.)

Horrigan-Taylor, the Palo Alto spokesperson, told WIRED that Lait’s mention of a “nuanced” solution referred to “resolving, to the extent permissible by law, neighborhood impacts and otherwise permitted use established by state law and local zoning.”

“Would I, or any other homeowner, be given the courtesy of a ‘nuanced solution’ if we were in violation of city code for over four years?” they added.

“Please know that you have not earned our trust and that we will take every opportunity to hold the city accountable if your solution satisfies a single [redacted] property owner over the interests of an entire neighborhood,” they continued.

“If you somehow craft a ‘nuanced solution’ based on promises,” the neighbor said, “the city will no doubt once again simply disappear and the damage to the neighborhood will continue.”

Lait did not respond right away. The neighbor followed up on March 13, asking if he had “reconsidered” his plan to offer a “‘nuanced solution’ for resolution of these ongoing issues by a serial code violator.” They asked when the neighborhood could “expect relief from the almost decade long disruptions.”

Behind the scenes, Zuckerberg’s lawyers were fighting to make sure the school could continue to operate. In a document dated March 14, Wade argues that she believed the activities at “the Property” “represent an appropriate residential use based on established state law as well as constitutional principles.”

Wade said that “the Family” was in the process of obtaining a “Large Family Daycare” license for the property, which is legal for a cohort of 14 or fewer children all under the age of 10.

“We consistently remind our vendors, guests, etc. to minimize noise, not loiter anywhere other than within the Family properties, and to keep areas clean,” Wade added in the letter. Wade also attached an adjusted lease corresponding with the address of the illicit school, which promises that the property will be used for only one purpose. The exact purpose is redacted.

On March 25, Lait told the neighbor that the city’s June 30 deadline for the Zuckerbergs to shut down the school had not changed. However, the family’s representative said that they were pursuing a daycare license. These licenses are granted by the state, not the city of Palo Alto.

The subtext of this email was that if the state gave them a daycare licence, there wasn’t much the city could do. Horrigan-Taylor confirmed with WIRED that “state licensed large family day care homes” do not require city approval, adding that the city also “does not regulate homeschooling.”

“Thanks for this rather surprising information,” the neighbor replied about a week later. “We have repeatedly presented ideas to the family over the past 8 years with very little to show for it, so from our perspective, we need to understand the city’s willingness to act or not to act.”

Baker told WIRED that the Zuckerbergs never ended up applying for a daycare license, a claim that corresponds with California’s public registry of daycare centers. (There are only two registered daycare centers in Palo Alto, and neither belongs to the Zuckerbergs. The Zuckerbergs’ oldest child, Maxima, will also turn 10 in December and consequently age out of any daycare legally operating in California.)

Horrigan-Taylor said that a representative for the Zuckerbergs told the city that the family wanted to move the school to “another location where private schools are permitted by right.”

In a school administrator job listing posted to the Association Montessori International website in July 2022 for “BBS,” Bicken Ben head of school Berge claims that the school had four distinct locations, and that applicants must be prepared to travel six to eight weeks per year. The June 2025 job listing also says that the “year-round” school spans “across multiple campuses,” but the main location of the job is listed as Palo Alto. It’s unclear where the other sites are located.

Most of the Zuckerbergs’ neighbors did not respond to WIRED’s request for comment. However, the ones that did clearly indicated that they would not be forgetting the Bicken Ben saga, or the past decade of disruption, anytime soon.

“Frankly I’m not sure what’s going on,” one neighbor said, when reached by WIRED via landline. “Except for noise and construction debris.”

This story originally appeared on wired.com.

Photo of WIRED

Wired.com is your essential daily guide to what’s next, delivering the most original and complete take you’ll find anywhere on innovation’s impact on technology, science, business and culture.

Mark Zuckerberg’s illegal school drove his neighbors crazy Read More »

how-to-declutter,-quiet-down,-and-take-the-ai-out-of-windows-11-25h2

How to declutter, quiet down, and take the AI out of Windows 11 25H2

Evergreen, Features, install, Tech, Windows 11 / /

A new major Windows 11 release means a new guide for cleaning up the OS.

Credit: Aurich Lawson | Getty Images

Credit: Aurich Lawson | Getty Images

It’s that time of year again—temperatures are dropping, leaves are changing color, and Microsoft is gradually rolling out another major yearly update to Windows 11.

The Windows 11 25H2 update is relatively minor compared to last year’s 24H2 update (the “25” here is a reference to the year the update was released, while the “H2” denotes that it was released in the second half of the year, a vestigial suffix from when Microsoft would release two major Windows updates per year). The 24H2 update came with some major under-the-hood overhauls of core Windows components and significant performance improvements for the Arm version; 25H2 is largely 24H2, but with a rolled-over version number to keep it in line with Microsoft’s timeline for security updates and tech support.

But Microsoft’s continuous update cadence for Windows 11 means that even the 24H2 version as it currently exists isn’t the same one Microsoft released a year ago.

To keep things current, we’ve combed through our Windows cleanup guide, updating it for the current build of Windows 11 25H2 (26200.7019) to help anyone who needs a fresh Windows install or who is finally updating from Windows 10 now that Microsoft is winding down support for it. We’ll outline dozens of individual steps you can take to clean up a “clean install” of Windows 11, which has taken an especially user-hostile attitude toward advertising and forcing the use of other Microsoft products.

As before, this is not a guide about creating an extremely stripped-down, telemetry-free version of Windows; we stick to the things that Microsoft officially supports turning off and removing. There are plenty of experimental hacks and scripts that take it a few steps farther, and/or automate some of the steps we outline here—NTDev’s Tiny11 project is one—but removing built-in Windows components can cause unexpected compatibility and security problems, and Tiny11 has historically had issues with basic table-stakes stuff like “installing security updates.”

These guides capture moments in time, and regular monthly Windows patches, app updates downloaded through the Microsoft Store, and other factors all can and will cause small variations from our directions. You may also see apps or drivers specific to your PC’s manufacturer. This guide also doesn’t cover the additional bloatware that may come out of the box with a new PC, starting instead with a freshly installed copy of Windows from a USB drive.

Table of Contents

Starting with Setup: Avoiding Microsoft account sign-in

The most contentious part of Windows 11’s setup process relative to earlier Windows versions is that it mandates a Microsoft account sign-in, with none of the readily apparent “limited account” fallbacks that existed in Windows 10. As of Windows 11 22H2, that’s true of both the Home and Pro editions.

There are two reasons I can think of not to sign in with a Microsoft account. The first is that you want nothing to do with a Microsoft account, thank you very much. Signing in makes Windows bombard you with more Microsoft 365, OneDrive, and Game Pass subscription upsells since all you need to do is add them to an account that already exists, and Windows setup will offer subscriptions to each if you sign in first.

The second—which describes my situation—is that you do use a Microsoft account because it offers some handy benefits like automated encryption of your local drive (having those encryption keys tied to my account has saved me a couple of times) or syncing of browser info and some preferences. But you don’t want to sign in at setup, either because you don’t want to be bothered with the extra upsells or you prefer your user folder to be located at “C:UsersAndrew” rather than “C:Users.”

Regardless of your reasoning, if you don’t want to bother with sign-in at setup, you have a few different options:

Use the command line

During Windows 11 Setup, after selecting a language and keyboard layout but before connecting to a network, hit Shift+F10 to open the command prompt (depending on your keyboard, you may also need to hit the Fn key before pressing F10). Type OOBEBYPASSNRO, hit Enter, and wait for the PC to reboot.

When it comes back, click “I don’t have Internet” on the network setup screen, and you’ll have recovered the option to use “limited setup” (aka a local account) again, like older versions of Windows 10 and 11 offered.

This option has been removed from some Windows 11 testing builds, but it still works as of this writing in 25H2. We may see this option removed in a future update to Windows.

For Windows 11 Pro

For Windows 11 Pro users, there’s a command-line-free workaround you can take advantage of.

Proceed through the Windows 11 setup as you normally would, including connecting to a network and allowing the system to check for updates. Eventually, you’ll be asked whether you’re setting your PC up for personal use or for “work or school.”

Select the “work or school” option, then “sign-in options,” at which point you’ll finally be given a button that says “domain join instead.” Click this to indicate you’re planning to join the PC to a corporate domain (even though you aren’t), and you’ll see the normal workflow for creating a “limited” local account.

The downside is that you’re starting your relationship with your new Windows install by lying to it. But hey, if you’re using the AI features, your computer is probably going to lie to you, too. It all balances out.

Using the Rufus tool

Credit: Andrew Cunningham

The Rufus tool can streamline a few of the more popular tweaks and workarounds for Windows 11 install media. Rufus is a venerable open source app for creating bootable USB media for both Windows and Linux. If you find yourself doing a lot of Windows 11 installs and don’t want to deal with Microsoft accounts, Rufus lets you tweak the install media itself so that the “limited setup” options always appear, no matter which edition of Windows you’re using.

To start, grab Rufus and then a fresh Windows 11 ISO file from Microsoft. You’ll also want an 8GB or larger USB drive; I’d recommend a 16GB or larger drive that supports USB 3.0 speeds, both to make things go a little faster and to leave yourself extra room for drivers, app installers, and anything else you might want to set a new PC up for the first time. (I also like this SanDisk drive that has a USB-C connector on one end and a USB-A connector on the other to ensure compatibility with all kinds of PCs.)

Fire up Rufus, select your USB drive and the Windows ISO, and hit Start to copy over all of the Windows files. After you hit Start, you’ll be asked if you want to disable some system requirements checks, remove the Microsoft account requirement, or turn off all the data collection settings that Windows asks you about the first time you set it up. What you do here is up to you; I usually turn off the sign-in requirement, but disabling the Secure Boot and TPM checks doesn’t stop those features from working once Windows is installed and running.

The rest of Windows 11 setup

The main thing I do here, other than declining any and all Microsoft 365 or Game Pass offers, is turn all the toggles on the privacy settings screen to “no.” This covers location services, the Find My Device feature, and four toggles that collectively send a small pile of usage and browsing data to Microsoft that it uses “to enhance your Microsoft experiences.” Pro tip: Use the Tab key and spacebar to quickly toggle these without clicking or scrolling.

Of these, I can imagine enabling Find My Device if you’re worried about theft or location services if you want Windows and apps to be able to access your location. But I tend not to send any extra telemetry or browsing data other than the basics (the only exception being on machines I enroll in the Windows Insider Preview program for testing, since Microsoft requires you to send more detailed usage data from those machines to help it test its beta software). If you want to change any of these settings after setup, they’re all in the Settings app under Privacy & Security.

If you have signed in with a Microsoft account during setup, you can expect to see several additional setup screens that aren’t offered when you’re signing in with a local account, including attempts to sell Microsoft 365, OneDrive, and Xbox Game Pass subscriptions. Accept or decline these offers as desired.

Cleaning up Windows 11

Reboot once this is done, and you’ll be at the Windows desktop. Start by installing any drivers you need, plus Windows updates.

When you first connect to the Internet, Windows may or may not decide to automatically pull down a few extraneous third-party apps and app shortcuts, things like Spotify or Grammarly—this has happened to me consistently in most Windows 11 installs I’ve done over the years, though it hasn’t generally happened on the 24H2 and 25H2 PCs I’ve set up.

Open the Start menu and right-click each of the apps you don’t want to remove the icons for and/or uninstall. Some of these third-party apps are just stubs that won’t actually be installed to your computer until you try to run them, so removing them directly from the Start menu will get rid of them entirely.

Right-clicking and uninstalling the unwanted apps that are pinned to the Start menu is the fastest (and, for some, the only) way to get rid of them.

Credit: Andrew Cunningham

Right-clicking and uninstalling the unwanted apps that are pinned to the Start menu is the fastest (and, for some, the only) way to get rid of them. Credit: Andrew Cunningham

The other apps and services included in a fresh Windows install generally at least have the excuse of being first-party software, though their usefulness will be highly user-specific: Xbox, the new Outlook app, Clipchamp, and LinkedIn are the ones that stand out, plus the ad-driven free-to-play version of the Solitaire suite that replaced the simple built-in version during the Windows 8 era.

Rather than tell you what I remove, I’ll tell you everything that can be removed from the Installed Apps section of the Settings app (also quickly accessible by right-clicking the Start button in the taskbar). You can make your own decisions here; I generally leave the in-box versions of classic Windows apps like Sound Recorder and Calculator while removing things I don’t use, like To Do or Clipchamp.

This list should be current for a fresh, fully updated install of Windows 11 25H2, at least in the US, but it doesn’t include any apps that might be specific to your hardware, like audio or GPU settings apps. Some individual apps may or may not appear as part of your Windows install.

  • Calculator
  • Camera
  • Clock (may also appear as Windows Clock)
  • Copilot
  • Family
  • Feedback Hub
  • Game Assist
  • Media Player
  • Microsoft 365 Copilot
  • Microsoft Clipchamp
  • Microsoft OneDrive: Removing this, if you don’t use it, should also get rid of notifications about OneDrive and turning on Windows Backup.
  • Microsoft Teams
  • Microsoft To Do
  • News
  • Notepad
  • Outlook for Windows
  • Paint
  • Photos
  • Power Automate
  • Quick Assist
  • Remote Desktop Connection
  • Snipping Tool
  • Solitaire & Casual Games
  • Sound Recorder
  • Sticky Notes
  • Terminal
  • Weather
  • Web Media Extensions
  • Xbox
  • Xbox Live

In Windows 11 23H2, Microsoft moved almost all of Windows’ non-removable apps to a System Components section, where they can be configured but not removed; this is where things like Phone Link, the Microsoft Store, Dev Home, and the Game Bar have ended up. The exception is Edge and its associated updater and WebView components; these are not removable, but they aren’t listed as “system components” for some reason, either.

Start, Search, Taskbar, and lock screen decluttering

Microsoft has been on a yearslong crusade against unused space in the Start menu and taskbar, which means there’s plenty here to turn off.

  • Right-click an empty space on the desktop, click Personalize, and click any of the other built-in Windows themes to turn off the Windows Spotlight dynamic wallpapers and the “Learn about this picture” icon.
  • Right-click the Taskbar and click Taskbar settings. I usually disable the Widgets board; you can leave this if you want to keep the little local weather icon in the lower-left corner of your screen, but this space is also sometimes used to present junky news articles from the Microsoft Start service.
    • If you want to keep Widgets enabled but clean it up a bit, open the Widgets menu, click the Settings gear in the top-right corner, scroll to “Show or hide feeds,” and turn the feed off. This will keep the weather, local sports scores, stocks, and a few other widgets, but it will get rid of the spammy news articles.
  • Also in the Taskbar settings, I usually change the Search field to “search icon only” to get rid of the picture in the search field and reduce the amount of space it takes up. Toggle the different settings until you find one you like.
  • Open Settings > Privacy & Security > Recommendations & offers and disable “Personalized offers,” “Improve Start and search results,” “Show notifications in Settings,” “Recommendations and offers in Settings,” and “Advertising ID” (some of these may already be turned off). These settings mostly either send data to Microsoft or clutter up the Settings app with various recommendations and ads.
  • Open Settings > Privacy & Security > Diagnostics & feedback, scroll down to “Feedback frequency,” and select “Never” to turn off all notifications requesting feedback about various Windows features.
  • Open Settings > Privacy & Security, click Search and disable “Show search highlights.” This cleans up the Search menu quite a bit, focusing it on searches you’ve done yourself and locally installed apps.

  • Open Settings > Personalization > Lock screen. Under “Personalize your lock screen,” switch from “Windows spotlight” to either Picture or Slideshow to use local images for your lock screen, and then uncheck the “get fun facts, tips, tricks, and more” box that appears. This will hide the other text boxes and clickable elements that Windows automatically adds to the lock screen in Spotlight mode. Under “Lock screen status,” select “none” to hide the weather widget and other stocks and news widgets from your lock screen.
  • If you own a newer Windows PC with a dedicated Copilot key, you can navigate to Settings > Personalization > Text input and scroll down to remap the key. Unfortunately, its usefulness is still limited—you can reassign it to the Search function or to the built-in Microsoft 365 app, but by default, Windows doesn’t give you the option to reassign it to open any old app.

Credit: Andrew Cunningham

By default, the Start menu will occasionally make “helpful” suggestions about third-party Microsoft Store apps to grab. These can and should be turned off.

  • Open Settings > Personalization > Start. Turn off “Show recommendations for tips, shortcuts, new apps, and more.” This will disable a feature where Microsoft Store apps you haven’t installed can show up in Recommendations along with your other files. You can also decide whether you want to be able to see more pinned apps or more recent/recommended apps and files on the Start menu, depending on what you find more useful.
  • On the same page, disable “show account-related notifications” to reduce the number of reminders and upsell notifications you see related to your Microsoft account.

Credit: Andrew Cunningham

  • Open Settings > System > Notifications, scroll down, and expand the additional settings section. Uncheck all three boxes here, which should get rid of all the “finish setting up your PC” prompts, among other things.
  • Also feel free to disable notifications from any specific apps you don’t want to hear from.

In-app AI features

Microsoft has steadily been adding image and text generation capabilities to some of the bedrock in-box Windows apps, from Paint and Photos to Notepad.

Exactly which AI features you’re offered will depend on whether you’ve signed in with a Microsoft account or not or whether you’re using a Copilot+ PC with access to more AI features that are executed locally on your PC rather than in the cloud (more on those in a minute).

But the short version is that it’s usually not possible to turn off or remove these AI features without uninstalling the entire app. Apps like Notepad and Edge do have toggles for shutting off Copilot and other related features, but no such toggles exist in Paint, for example.

Even if you can find some Registry key or another backdoor way to shut these things off, there’s no guarantee the settings will stick as these apps are updated; it’s probably easier to just try to ignore any AI features within these apps that you don’t plan to use.

Removing Recall, and other extra steps for Copilot+ PCs

So far, everything we’ve covered has been applicable to any PC that can run Windows 11. But new PCs with the Copilot+ branding—anything with a Qualcomm Snapdragon X chip in it or things with certain Intel Core Ultra or AMD Ryzen AI CPUs—get extra features that other Windows 11 PCs don’t have. Given that these are their own unique subclass of PCs, it’s worth exploring what’s included and what can be turned off.

Removing Recall will be possible, though it’s done through a relatively obscure legacy UI rather than the Settings app. Credit: Andrew Cunningham

One Copilot+ feature that can be fully removed, in part because of the backlash it initially caused, is the data-scraping Recall feature. Recall won’t be enabled on your Copilot+ system unless you’re signed in with a Microsoft account and you explicitly opt in. But if fully removing the feature gives you extra peace of mind, then by all means, remove it.

  • If you just want to make sure Recall isn’t active, navigate to Settings > Privacy & security > Recall & snapshots. This is where you adjust Recall’s settings and verify whether it’s turned on or off.
  • To fully remove Recall, open Settings > System > Optional Features, scroll down to the bottom of this screen, and click More Windows features. This will open the old “Turn Windows features on or off” Control Panel applet used to turn on or remove some legacy or power-user-centric components, like old versions of the .NET Framework or Hyper-V. It’s arranged alphabetically.
  • In Settings > Privacy & security > Click to Do, you’ll also find a toggle to disable Click to Do, a Copilot+ feature that takes a screenshot of your desktop and tries to make recommendations or suggest actions you might perform (copying and pasting text or an image, for example).

Apps like Paint or Photos may also prompt you to install an extension for AI-powered image generation from the Microsoft Store. This extension—which weighs in at well over a gigabyte as of this writing—is not installed by default. If you have installed it, you can remove it by opening Settings > Apps > Installed apps and removing “ImageCreationHostApp.”

Bonus: Cleaning up Microsoft Edge

I use Edge out of pragmatism rather than love—”the speed, compatibility, and extensions ecosystem of Chrome, backed by the resources of a large company that isn’t Google” is still a decent pitch. But Edge has become steadily less appealing as Microsoft has begun pushing its own services more aggressively and stuffing the browser with AI features. In a vacuum, Firefox aligns better with what I want from a browser, but it just doesn’t respond well to my normal tab-monster habits despite several earnest attempts to switch—things bog down and RAM runs out. I’ve also had mixed experience with the less-prominent Chromium clones, like Opera, Vivaldi, and Brave. So Edge it is, at least for now.

The main problem with Edge on a new install of Windows is that even more than Windows, it exists in a universe where no one would ever want to switch search engines or shut off any of Microsoft’s “value-added features” except by accident. Case in point: Signing in with a Microsoft account will happily sync your bookmarks, extensions, and many kinds of personal data. But many settings for search engine changes or for opting out of Microsoft services do not sync between systems and require a fresh setup each time.

Below are the Edge settings I change to maximize the browser’s usefulness (and usable screen space) while minimizing annoying distractions; it involves turning off most of the stuff Microsoft has added to the Chromium version of Edge since it entered public preview many years ago. Here’s a list of things to tweak, whether you sign in with a Microsoft account or not.

  • On the Start page when you first open the browser, hit the Settings gear in the upper-right corner. Turn off “Quick links” (or if you leave them on, turn off “Show sponsored links”) and then turn off “show content.” Whether you leave the custom background or the weather widget is up to you.
  • Click the “your privacy choices” link at the bottom of the menu and turn off the “share my data with third parties for personalized ads” toggle.

Edge has scattered some of the settings we change over the last year, but the browser is still full of toggles we prefer to keep turned off. Andrew Cunningham

  • In the Edge UI, click the ellipsis icon near the upper-right corner of the screen and click Settings.
  • Click Profiles in the left Settings sidebar. Click Microsoft Rewards, and then turn it off.
  • Click Privacy, Search, & Services in the Settings sidebar.
    • In Tracking prevention, I set tracking prevention to “strict,” though if you use some other kind of content blocker, this may be redundant; it can also occasionally prompt “it looks like you’re using an ad-blocker” pop-up from sites even if you aren’t.
    • In Privacy, if they’re enabled, disable the toggles under “Optional diagnostic data,” “Help improve Microsoft products,” and “Allow Microsoft to save your browsing activity.”
    • In Search and connected experiences, disable the “Suggest similar sites when a website can’t be found,” “Save time and money with Shopping in Microsoft Edge,” and “Organize your tabs” toggles.
      • If you want to switch from Bing, click “Address bar and search” and switch to your preferred engine, whether that’s Google, DuckDuckGo, or something else. Then click “Search suggestions and filters” and disable “Show me search and site suggestions using my typed characters.”

These settings retain basic spellcheck without any of the AI-related additions. Credit: Andrew Cunningham

  • Click Appearance in the left-hand Settings sidebar, and scroll down to Copilot and sidebar
    • Turn the sidebar off, and turn off the “Personalize my top sites in customize sidebar” and “Allow sidebar apps to show notifications” toggles.
    • Click Copilot under App specific settings. Turn off “Show Copilot button on the toolbar.” Then, back in the Copilot and sidebar settings, turn off the “Show sidebar button” toggle that has just appeared.
  • Click Languages in the left-hand navigation. Disable “Use Copilot for writing on the web.” Turn off “use text prediction” if you want to prevent things you type from being sent to Microsoft, and switch the spellchecker from Microsoft Editor to Basic. (I don’t actually mind Microsoft Editor, but it’s worth remembering if you’re trying to minimize the amount of data Edge sends back to the company.)

Windows-as-a-nuisance

The most time-consuming part of installing a fresh, direct-from-Microsoft copy of Windows XP or Windows 7 was usually reinstalling all the apps you wanted to run on your PC, from your preferred browser to Office, Adobe Reader, Photoshop, and the VLC player. You still need to do all of that in a new Windows 11 installation. But now more than ever, most people will want to go through the OS and turn off a bunch of stuff to make the day-to-day experience of using the operating system less annoying.

That’s more relevant now that Microsoft has formally ended support for Windows 10. Yes, Windows 10 users can get an extra year of security updates relatively easily, but many who have been putting off the Windows 11 upgrade will be taking the plunge this year.

The settings changes we’ve recommended here may not fix everything, but they can at least give you some peace, shoving Microsoft into the background and allowing you to do what you want with your PC without as much hassle. Ideally, Microsoft would insist on respectful, user-friendly defaults itself. But until that happens, these changes are the best you can do.

Photo of Andrew Cunningham

Andrew is a Senior Technology Reporter at Ars Technica, with a focus on consumer tech including computer hardware and in-depth reviews of operating systems like Windows and macOS. Andrew lives in Philadelphia and co-hosts a weekly book podcast called Overdue.

How to declutter, quiet down, and take the AI out of Windows 11 25H2 Read More »

internet-archive’s-legal-fights-are-over,-but-its-founder-mourns-what-was-lost

Internet Archive’s legal fights are over, but its founder mourns what was lost

copyright act, copyright law, digital library, Features, Internet Archive, online archive, online library, Policy, wayback machine / /

“We survived, but it wiped out the library,” Internet Archive’s founder says.

Internet Archive founder Brewster Kahle celebrates 1 trillion web pages on stage with staff. Credit: via the Internet Archive

Last month, the Internet Archive’s Wayback Machine archived its trillionth webpage, and the nonprofit invited its more than 1,200 library partners and 800,000 daily users to join a celebration of the moment. To honor “three decades of safeguarding the world’s online heritage,” the city of San Francisco declared October 22 to be “Internet Archive Day.” The Archive was also recently designated a federal depository library by Sen. Alex Padilla (D-Calif.), who proclaimed the organization a “perfect fit” to expand “access to federal government publications amid an increasingly digital landscape.”

The Internet Archive might sound like a thriving organization, but it only recently emerged from years of bruising copyright battles that threatened to bankrupt the beloved library project. In the end, the fight led to more than 500,000 books being removed from the Archive’s “Open Library.”

“We survived,” Internet Archive founder Brewster Kahle told Ars. “But it wiped out the Library.”

An Internet Archive spokesperson confirmed to Ars that the archive currently faces no major lawsuits and no active threats to its collections. Kahle thinks “the world became stupider” when the Open Library was gutted—but he’s moving forward with new ideas.

History of the Internet Archive

Kahle has been striving since 1996 to transform the Internet Archive into a digital Library of Alexandria—but “with a better fire protection plan,” joked Kyle Courtney, a copyright lawyer and librarian who leads the nonprofit eBook Study Group, which helps states update laws to protect libraries.

When the Wayback Machine was born in 2001 as a way to take snapshots of the web, Kahle told The New York Times that building free archives was “worth it.” He was also excited that the Wayback Machine had drawn renewed media attention to libraries.

At the time, law professor Lawrence Lessig predicted that the Internet Archive would face copyright battles, but he also believed that the Wayback Machine would change the way the public understood copyright fights.

”We finally have a clear and tangible example of what’s at stake,” Lessig told the Times. He insisted that Kahle was “defining the public domain” online, which would allow Internet users to see ”how easy and important” the Wayback Machine “would be in keeping us sane and honest about where we’ve been and where we’re going.”

Kahle suggested that IA’s legal battles weren’t with creators or publishers so much as with large media companies that he thinks aren’t “satisfied with the restriction you get from copyright.”

“They want that and more,” Kahle said, pointing to e-book licenses that expire as proof that libraries increasingly aren’t allowed to own their collections. He also suspects that such companies wanted the Wayback Machine dead—but the Wayback Machine has survived and proved itself to be a unique and useful resource.

The Internet Archive also began archiving—and then lending—e-books. For a decade, the Archive had loaned out individual e-books to one user at a time without triggering any lawsuits. That changed when IA decided to temporarily lift the cap on loans from its Open Library project to create a “National Emergency Library” as libraries across the world shut down during the early days of the COVID-19 pandemic. The project eventually grew to 1.4 million titles.

But lifting the lending restrictions also brought more scrutiny from copyright holders, who eventually sued the Archive. Litigation went on for years. In 2024, IA lost its final appeal in a lawsuit brought by book publishers over the Archive’s Open Library project, which used a novel e-book lending model to bypass publishers’ licensing fees and checkout limitations. Damages could have topped $400 million, but publishers ultimately announced a “confidential agreement on a monetary payment” that did not bankrupt the Archive.

Litigation has continued, though. More recently, the Archive settled another suit over its Great 78 Project after music publishers sought damages of up to $700 million. A settlement in that case, reached last month, was similarly confidential. In both cases, IA’s experts challenged publishers’ estimates of their losses as massively inflated.

For Internet Archive fans, a group that includes longtime Internet users, researchers, students, historians, lawyers, and the US government, the end of the lawsuits brought a sigh of relief. The Archive can continue—but it can’t run one of its major programs in the same way.

What the Internet Archive lost

To Kahle, the suits have been an immense setback to IA’s mission.

Publishers had argued that the Open Library’s lending harmed the e-book market, but IA says its vision for the project was not to frustrate e-book sales (which it denied its library does) but to make it easier for researchers to reference e-books by allowing Wikipedia to link to book scans. Wikipedia has long been one of the most visited websites in the world, and the Archive wanted to deepen its authority as a research tool.

“One of the real purposes of libraries is not just access to information by borrowing a book that you might buy in a bookstore,” Kahle said. “In fact, that’s actually the minority. Usually, you’re comparing and contrasting things. You’re quoting. You’re checking. You’re standing on the shoulders of giants.”

Meredith Rose, senior policy counsel for Public Knowledge, told Ars that the Internet Archive’s Wikipedia enhancements could have served to surface information that’s often buried in books, giving researchers a streamlined path to source accurate information online.

But Kahle said the lawsuits against IA showed that “massive multibillion-dollar media conglomerates” have their own interests in controlling the flow of information. “That’s what they really succeeded at—to make sure that Wikipedia readers don’t get access to books,” Kahle said.

At the heart of the Open Library lawsuit was publishers’ market for e-book licenses, which libraries complain provide only temporary access for a limited number of patrons and cost substantially more than the acquisition of physical books. Some states are crafting laws to restrict e-book licensing, with the aim of preserving library functions.

“We don’t want libraries to become Hulu or Netflix,” said Courtney of the eBook Study Group, posting warnings to patrons like “last day to check out this book, August 31st, then it goes away forever.”

He, like Kahle, is concerned that libraries will become unable to fulfill their longtime role—preserving culture and providing equal access to knowledge. Remote access, Courtney noted, benefits people who can’t easily get to libraries, like the elderly, people with disabilities, rural communities, and foreign-deployed troops.

Before the Internet Archive cases, libraries had won some important legal fights, according to Brandon Butler, a copyright lawyer and executive director of Re:Create, a coalition of “libraries, civil libertarians, online rights advocates, start-ups, consumers, and technology companies” that is “dedicated to balanced copyright and a free and open Internet.”

But the Internet Archive’s e-book fight didn’t set back libraries, Butler said, because the loss didn’t reverse any prior court wins. Instead, IA had been “exploring another frontier” beyond the Google Books ruling, which deemed Google’s searchable book excerpts a transformative fair use, hoping that linking to books from Wikipedia would also be deemed fair use. But IA “hit the edge” of what courts would allow, Butler said.

IA basically asked, “Could fair use go this much farther?” Butler said. “And the courts said, ‘No, this is as far as you go.’”

To Kahle, the cards feel stacked against the Internet Archive, with courts, lawmakers, and lobbyists backing corporations seeking “hyper levels of control.” He said IA has always served as a research library—an online destination where people can cross-reference texts and verify facts, just like perusing books at a local library.

“We’re just trying to be a library,” Kahle said. “A library in a traditional sense. And it’s getting hard.”

Fears of big fines may delay digitization projects

President Donald Trump’s cuts to the federal Institute of Museum and Library Services have put America’s public libraries at risk, and reduced funding will continue to challenge libraries in the coming years, ALA has warned. Butler has also suggested that under-resourced libraries may delay digitization efforts for preservation purposes if they worry that publishers may threaten costly litigation.

He told Ars he thinks courts are getting it right on recent fair use rulings. But he noted that libraries have fewer resources for legal fights because copyright law “has this provision that says, well, if you’re a copyright holder, you really don’t have to prove that you suffered any harm at all.”

“You can just elect [to receive] a massive payout based purely on the fact that you hold a copyright and somebody infringed,” Butler said. “And that’s really unique. Almost no other country in the world has that sort of a system.”

So while companies like AI firms may be able to afford legal fights with rights holders, libraries must be careful, even when they launch projects that seem “completely harmless and innocuous,” Butler said. Consider the Internet Archive’s Great 78 Project, which digitized 400,000 old shellac records, known as 78s, that were originally pressed from 1898 to the 1950s.

“The idea that somebody’s going to stream a 78 of an Elvis song instead of firing it up on their $10-a-month Spotify subscription is silly, right?” Butler said. “It doesn’t pass the laugh test, but given the scale of the project—and multiply that by the statutory damages—and that makes this an extremely dangerous project all of a sudden.”

Butler suggested that statutory damages could disrupt the balance that ensures the public has access to knowledge, creators get paid, and human creativity thrives, as AI advances and libraries’ growth potentially stalls.

“It sets the risk so high that it may force deals in situations where it would be better if people relied on fair use. Or it may scare people from trying new things because of the stakes of a copyright lawsuit,” Butler said.

Courtney, who co-wrote a whitepaper detailing the legal basis for different forms of “controlled digital lending” like the Open Library project uses, suggested that Kahle may be the person who’s best prepared to push the envelope on copyright.

When asked how the Internet Archive managed to avoid financial ruin, Courtney said it survived “only because their leader” is “very smart and capable.” Of all the “flavors” of controlled digital lending (CDL) that his paper outlined, Kahle’s methodology for the Open Library Project was the most “revolutionary,” Courtney said.

Importantly, IA’s loss did not doom other kinds of CDL that other archives use, he noted, nor did it prevent libraries from trying new things.

“Fair use is a case-by-case determination” that will be made as urgent preservation needs arise, Courtney told Ars, and “libraries have a ton of stuff that aren’t going to make the jump to digital unless we digitize them. No one will have access to them.”

What’s next for the Internet Archive?

The lawsuits haven’t dampened Kahle’s resolve to expand IA’s digitization efforts, though. Moving forward, the group will be growing a project called Democracy’s Library, which is “a free, open, online compendium of government research and publications from around the world” that will be conveniently linked in Wikipedia articles to help researchers discover them.

The Archive is also collecting as many physical materials as possible to help preserve knowledge, even as “the library system is largely contracting,” Kahle said. He noted that libraries historically tend to grow in societies that prioritize education and decline in societies where power is being concentrated, and he’s worried about where the US is headed. That makes it hard to predict if IA—or any library project—will be supported in the long term.

With governments globally partnering with the biggest tech companies to try to win the artificial intelligence race, critics have warned of threats to US democracy, while the White House has escalated its attack on libraries, universities, and science over the past year.

Meanwhile, AI firms face dozens of lawsuits from creators and publishers, which Kahle thinks only the biggest tech companies can likely afford to outlast. The momentum behind AI risks giving corporations even more control over information, Kahle said, and it’s uncertain if archives dedicated to preserving the public memory will survive attacks from multiple fronts.

“Societies that are [growing] are the ones that need to educate people” and therefore promote libraries, Kahle said. But when societies are “going down,” such as in times of war, conflict, and social upheaval, libraries “tend to get destroyed by the powerful. It used to be king and church, and it’s now corporations and governments.” (He recommended The Library: A Fragile History as a must-read to understand the challenges libraries have always faced.)

Kahle told Ars he’s not “black and white” on AI, and he even sees some potential for AI to enhance library services.

He’s more concerned that libraries in the US are losing support and may soon cease to perform classic functions that have always benefited civilizations—like buying books from small publishers and local authors, supporting intellectual endeavors, and partnering with other libraries to expand access to diverse collections.

To prevent these cultural and intellectual losses, he plans to position IA as a refuge for displaced collections, with hopes to digitize as much as possible while defending the early dream that the Internet could equalize access to information and supercharge progress.

“We want everyone [to be] a reader,” Kahle said, and that means “we want lots of publishers, we want lots of vendors, booksellers, lots of libraries.”

But, he asked, “Are we going that way? No.”

To turn things around, Kahle suggested that copyright laws be “re-architected” to ensure “we have a game with many winners”—where authors, publishers, and booksellers get paid, library missions are respected, and progress thrives. Then society can figure out “what do we do with this new set of AI tools” to keep the engine of human creativity humming.

Photo of Ashley Belanger

Ashley is a senior policy reporter for Ars Technica, dedicated to tracking social impacts of emerging policies and new technologies. She is a Chicago-based journalist with 20 years of experience.

Internet Archive’s legal fights are over, but its founder mourns what was lost Read More »

“unexpectedly,-a-deer-briefly-entered-the-family-room”:-living-with-gemini-home

“Unexpectedly, a deer briefly entered the family room”: Living with Gemini Home

AI, Features, Google, Google Gemini, Google Home, Tech / /

60 percent of the time, it works every time

Gemini for Home unleashes gen AI on your Nest camera footage, but it gets a lot wrong.

Google Home with Gemini

The Google Home app has Gemini integration for paying customers. Credit: Ryan Whitwam

The Google Home app has Gemini integration for paying customers. Credit: Ryan Whitwam

You just can’t ignore the effects of the generative AI boom.

Even if you don’t go looking for AI bots, they’re being integrated into virtually every product and service. And for what? There’s a lot of hand-wavey chatter about agentic this and AGI that, but what can “gen AI” do for you right now? Gemini for Home is Google’s latest attempt to make this technology useful, integrating Gemini with the smart home devices people already have. Anyone paying for extended video history in the Home app is about to get a heaping helping of AI, including daily summaries, AI-labeled notifications, and more.

Given the supposed power of AI models like Gemini, recognizing events in a couple of videos and answering questions about them doesn’t seem like a bridge too far. And yet Gemini for Home has demonstrated a tenuous grasp of the truth, which can lead to some disquieting interactions, like periodic warnings of home invasion, both human and animal.

It can do some neat things, but is it worth the price—and the headaches?

Does your smart home need a premium AI subscription?

Simply using the Google Home app to control your devices does not turn your smart home over to Gemini. This is part of Google’s higher-tier paid service, which comes with extended camera history and Gemini features for $20 per month. That subscription pipes your video into a Gemini AI model that generates summaries for notifications, as well as a “Daily Brief” that offers a rundown of everything that happened on a given day. The cheaper $10 plan provides less video history and no AI-assisted summaries or notifications. Both plans enable Gemini Live on smart speakers.

According to Google, it doesn’t send all of your video to Gemini. That would be a huge waste of compute cycles, so Gemini only sees (and summarizes) event clips. Those summaries are then distilled at the end of the day to create the Daily Brief, which usually results in a rather boring list of people entering and leaving rooms, dropping off packages, and so on.

Importantly, the Gemini model powering this experience is not multimodal—it only processes visual elements of videos and does not integrate audio from your recordings. So unusual noises or conversations captured by your cameras will not be searchable or reflected in AI summaries. This may be intentional to ensure your conversations are not regurgitated by an AI.

Gemini smart home plans

Credit: Google

Paying for Google’s AI-infused subscription also adds Ask Home, a conversational chatbot that can answer questions about what has happened in your home based on the status of smart home devices and your video footage. You can ask questions about events, retrieve video clips, and create automations.

There are definitely some issues with Gemini’s understanding of video, but Ask Home is quite good at creating automations. It was possible to set up automations in the old Home app, but the updated AI is able to piece together automations based on your natural language request. Perhaps thanks to the limited set of possible automation elements, the AI gets this right most of the time. Ask Home is also usually able to dig up past event clips, as long as you are specific about what you want.

The Advanced plan for Gemini Home keeps your videos for 60 days, so you can only query the robot on clips from that time period. Google also says it does not retain any of that video for training. The only instance in which Google will use security camera footage for training is if you choose to “lend” it to Google via an obscure option in the Home app. Google says it will keep these videos for up to 18 months or until you revoke access. However, your interactions with Gemini (like your typed prompts and ratings of outputs) are used to refine the model.

The unexpected deer

Every generative AI bot makes the occasional mistake, but you’ll probably not notice every one. When the AI hallucinates about your daily life, however, it’s more noticeable. There’s no reason Google should be confused by my smart home setup, which features a couple of outdoor cameras and one indoor camera—all Nest-branded with all the default AI features enabled—to keep an eye on my dogs. So the AI is seeing a lot of dogs lounging around and staring out the window. One would hope that it could reliably summarize something so straightforward.

One may be disappointed, though.

In my first Daily Brief, I was fascinated to see that Google spotted some indoor wildlife. “Unexpectedly, a deer briefly entered the family room,” Gemini said.

Home Brief with deer

Dogs and deer are pretty much the same thing, right? Credit: Ryan Whitwam

Gemini does deserve some credit for recognizing that the appearance of a deer in the family room would be unexpected. But the “deer” was, naturally, a dog. This was not a one-time occurrence, either. Gemini sometimes identifies my dogs correctly, but many event clips and summaries still tell me about the notable but brief appearance of deer around the house and yard.

This deer situation serves as a keen reminder that this new type of AI doesn’t “think,” although the industry’s use of that term to describe simulated reasoning could lead you to believe otherwise. A person looking at this video wouldn’t even entertain the possibility that they were seeing a deer after they’ve already seen the dogs loping around in other videos. Gemini doesn’t have that base of common sense, though. If the tokens say deer, it’s a deer. I will say, though, Gemini is great at recognizing car models and brand logos. Make of that what you will.

The animal mix-up is not ideal, but it’s not a major hurdle to usability. I didn’t seriously entertain the possibility that a deer had wandered into the house, and it’s a little funny the way the daily report continues to express amazement that wildlife is invading. It’s a pretty harmless screw-up.

“Overall identification accuracy depends on several factors, including the visual details available in the camera clip for Gemini to process,” explains a Google spokesperson. “As a large language model, Gemini can sometimes make inferential mistakes, which leads to these misidentifications, such as confusing your dog with a cat or deer.”

Google also says that you can tune the AI by correcting it when it screws up. This works sometimes, but the system still doesn’t truly understand anything—that’s beyond the capabilities of a generative AI model. After telling Gemini that it’s seeing dogs rather than deer, it sees wildlife less often. However, it doesn’t seem to trust me all the time, causing it to report the appearance of a deer that is “probably” just a dog.

A perfect fit for spooky season

Gemini’s smart home hallucinations also have a less comedic side. When Gemini mislabels an event clip, you can end up with some pretty distressing alerts. Imagine that you’re out and about when your Gemini assistant hits you with a notification telling you, “A person was seen in the family room.”

A person roaming around the house you believed to be empty? That’s alarming. Is it an intruder, a hallucination, a ghost? So naturally, you check the camera feed to find… nothing. An Ars Technica investigation confirms AI cannot detect ghosts. So a ghost in the machine?

Oops, we made you think someone broke into your house.

Credit: Ryan Whitwam

Oops, we made you think someone broke into your house. Credit: Ryan Whitwam

On several occasions, I’ve seen Gemini mistake dogs and totally empty rooms (or maybe a shadow?) for a person. It may be alarming at first, but after a few false positives, you grow to distrust the robot. Now, even if Gemini correctly identified a random person in the house, I’d probably ignore it. Unfortunately, this is the only notification experience for Gemini Home Advanced.

“You cannot turn off the AI description while keeping the base notification,” a Google spokesperson told me. They noted, however, that you can disable person alerts in the app. Those are enabled when you turn on Google’s familiar faces detection.

Gemini often twists reality just a bit instead of creating it from whole cloth. A person holding anything in the backyard is doing yardwork. One person anywhere, doing anything, becomes several people. A dog toy becomes a cat lying in the sun. A couple of birds become a raccoon. Gemini likes to ignore things, too, like denying there was a package delivery even when there’s a video tagged as “person delivers package.”

Gemini misses package

Gemini still refused to admit it was wrong.

Credit: Ryan Whitwam

Gemini still refused to admit it was wrong. Credit: Ryan Whitwam

At the end of the day, Gemini is labeling most clips correctly and therefore produces mostly accurate, if sometimes unhelpful, notifications. The problem is the flip side of “mostly,” which is still a lot of mistakes. Some of these mistakes compel you to check your cameras—at least, before you grow weary of Gemini’s confabulations. Instead of saving time and keeping you apprised of what’s happening at home, it wastes your time. For this thing to be useful, inferential errors cannot be a daily occurrence.

Learning as it goes

Google says its goal is to make Gemini for Home better for everyone. The team is “investing heavily in improving accurate identification” to cut down on erroneous notifications. The company also believes that having people add custom instructions is a critical piece of the puzzle. Maybe in the future, Gemini for Home will be more honest, but it currently takes a lot of hand-holding to move it in the right direction.

With careful tuning, you can indeed address some of Gemini for Home’s flights of fancy. I see fewer deer identifications after tinkering, and a couple of custom instructions have made the Home Brief waste less space telling me when people walk into and out of rooms that don’t exist. But I still don’t know how to prompt my way out of Gemini seeing people in an empty room.

Nest Cam 2025

Gemini AI features work on all Nest cams, but the new 2025 models are “designed for Gemini.”

Credit: Ryan Whitwam

Gemini AI features work on all Nest cams, but the new 2025 models are “designed for Gemini.” Credit: Ryan Whitwam

Despite its intention to improve Gemini for Home, Google is releasing a product that just doesn’t work very well out of the box, and it misbehaves in ways that are genuinely off-putting. Security cameras shouldn’t lie about seeing intruders, nor should they tell me I’m lying when they fail to recognize an event. The Ask Home bot has the standard disclaimer recommending that you verify what the AI says. You have to take that warning seriously with Gemini for Home.

At launch, it’s hard to justify paying for the $20 Advanced Gemini subscription. If you’re already paying because you want the 60-day event history, you’re stuck with the AI notifications. You can ignore the existence of Daily Brief, though. Stepping down to the $10 per month subscription gets you just 30 days of event history with the old non-generative notifications and event labeling. Maybe that’s the smarter smart home bet right now.

Gemini for Home is widely available for those who opted into early access in the Home app. So you can avoid Gemini for the time being, but it’s only a matter of time before Google flips the switch for everyone.

Hopefully it works better by then.

Photo of Ryan Whitwam

Ryan Whitwam is a senior technology reporter at Ars Technica, covering the ways Google, AI, and mobile technology continue to change the world. Over his 20-year career, he’s written for Android Police, ExtremeTech, Wirecutter, NY Times, and more. He has reviewed more phones than most people will ever own. You can follow him on Bluesky, where you will see photos of his dozens of mechanical keyboards.

“Unexpectedly, a deer briefly entered the family room”: Living with Gemini Home Read More »

10m-people-watched-a-youtuber-shim-a-lock;-the-lock-company-sued-him-bad-idea.

10M people watched a YouTuber shim a lock; the lock company sued him. Bad idea.

Features, lawsuit, lock picking, locks, McNally, Policy, Prove Locks / /

It’s still legal to pick locks, even when you swing your legs.

“Opening locks” might not sound like scintillating social media content, but Trevor McNally has turned lock-busting into online gold. A former US Marine Staff Sergeant, McNally today has more than 7 million followers and has amassed more than 2 billion views just by showing how easy it is to open many common locks by slapping, picking, or shimming them.

This does not always endear him to the companies that make the locks.

On March 3, 2025, a Florida lock company called Proven Industries released a social media promo video just begging for the McNally treatment. The video was called, somewhat improbably, “YOU GUYS KEEP SAYING YOU CAN EASILY BREAK OFF OUR LATCH PIN LOCK.” In it, an enthusiastic man in a ball cap says he will “prove a lot of you haters wrong.” He then goes hard at Proven’s $130 model 651 trailer hitch lock with a sledgehammer, bolt cutters, and a crowbar.

Naturally, the lock hangs tough.

An Instagram user brought the lock to McNally’s attention by commenting, “Let’s introduce it to the @mcnallyofficial poke.” Someone from Proven responded, saying that McNally only likes “the cheap locks lol because they are easy and fast.” Proven locks were said to be made of sterner stuff.

But on April 3, McNally posted a saucy little video to social media platforms. In it, he watches the Proven promo video while swinging his legs and drinking a Juicy Juice. He then hops down from his seat, goes over to a Proven trailer hitch lock, and opens it in a matter of seconds using nothing but a shim cut from a can of Liquid Death. He says nothing during the entire video, which has been viewed nearly 10 million times on YouTube alone.

Despite practically begging people to attempt this, Proven Industries owner Ron Lee contacted McNally on Instagram. “Just wanted to say thanks and be prepared!” he wrote. McNally took this as a threat.

(Oddly enough, Proven’s own homepage features a video in which the company trashes competing locks and shows just how easy it is to defeat them. And its news pages contain articles and videos on “The Hidden Flaws of Master Locks” and other brands. Why it got so upset about McNally’s video is unclear.)

The next day, Lee texted McNally’s wife. The message itself was apparently Lee’s attempt to de-escalate things; he says he thought the number belonged to McNally, and the message itself was unobjectionable. But after the “be prepared!” notice of the day before, and given the fact that Lee already knew how to contact him on Instagram, McNally saw the text as a way “to intimidate me and my family.” That feeling was cemented when McNally found out that Lee was a triple felon—and that in one case, Lee had hired someone “to throw a brick through the window of his ex-wife.”

Concerned about losing business, Lee kept trying to shut McNally down. Proven posted a “response video” on April 6 and engaged with numerous social media commenters, telling them that things were “going to get really personal” for McNally. Proven employees alleged publicly that McNally was deceiving people about all the prep work he had done to make a “perfectly cut out” shim. Without extensive experience, long prep work, and precise measurements, it was said, Proven’s locks were in little danger of being opened by rogue actors trying to steal your RV.

“Sucks to see how many people take everything they see online for face value,” one Proven employee wrote. “Sounds like a bunch of liberals lol.”

Proven also had its lawyers file “multiple” DMCA takedown notices against the McNally video, claiming that its use of Proven’s promo video was copyright infringement.

McNally didn’t bow to the pressure, though, instead uploading several more videos showing him opening Proven locks. In one of them, he takes aim at Proven’s claims about his prep work by retrieving a new lock from an Amazon delivery kiosk, taking it outside—and popping it in seconds using a shim he cuts right on camera, with no measurements, from an aluminum can.

Help us write more stories like this—while ditching ads

Ars subscribers support our independent journalism, which they can read ad-free and with enhanced privacy protections. And it’s only a few bucks a month.

Ars Pro

$5 / month

Subscribe

  • No ads
  • No tracking
  • Enhanced experience

58.3333% off!

Ars Pro

$25 / year

Subscribe

  • Best value
  • Still no ads
  • Still no tracking

Ars Pro++

$50 / year

Subscribe

  • All Ars Pro features
  • Support journalism
  • Special ++ badge

On May 1, Proven filed a federal lawsuit against McNally in the Middle District of Florida, charging him with a huge array of offenses: (1) copyright infringement, (2) defamation by implication, (3) false advertising, (4) violating the Florida Deceptive and Unfair Trade Practices Act, (5) tortious interference with business relationships, (6) unjust enrichment, (7) civil conspiracy, and (8) trade libel. Remarkably, the claims stemmed from a video that all sides admit was accurate and in which McNally himself said nothing.

Screenshot of a social media exchange.

In retrospect, this was probably not a great idea.

Don’t mock me, bro

How can you defame someone without even speaking? Proven claimed “defamation by implication,” arguing that the whole setup of McNally’s videos was unfair to the company and its product. McNally does not show his prep work, which (Proven argued) conveys to the public the false idea that Proven’s locks are easy to bypass. While the shimming does work, Proven argued that it would be difficult for an untrained user to perform.

But what Proven really, really didn’t like was being mocked. McNally’s decision to drink—and shake!—a juice box on video comes up in court papers a mind-boggling number of times. Here’s a sample:

McNally appears swinging his legs and sipping from an apple juice box, conveying to the purchasing public that bypassing Plaintiff’s lock is simple, trivial, and even comical…

…showing McNally drinking from, and shaking, a juice box, all while swinging his legs, and displaying the Proven Video on a mobile device…

The tone, posture, and use of the juice box prop and childish leg swinging that McNally orchestrated in the McNally Video was intentional to diminish the perceived seriousness of Proven Industries…

The use of juvenile imagery, such as sipping from a juice box while casually applying the shim, reinforces the misleading impression that the lock is inherently insecure and marketed deceptively…

The video then abruptly shifts to Defendant in a childlike persona, sipping from a juice box and casually applying a shim to the lock…

In the end, Proven argued that the McNally video was “for commercial entertainment and mockery,” produced for the purpose of “humiliating Plaintiff.” McNally, it was said, “will not stop until he destroys Proven’s reputation.” Justice was needed. Expensive, litigious justice.

But the proverbially level-headed horde of Internet users does not always love it when companies file thermonuclear lawsuits against critics. Sometimes, in fact, the level-headed horde disregards everything taught by that fount of judicial knowledge, The People’s Court, and they take the law into their own hands.

Proven was soon the target of McNally fans. The company says it was “forced to disable comments on posts and product videos due to an influx of mocking and misleading replies furthering the false narrative that McNally conveyed to the viewers.” The company’s customer service department received such an “influx of bogus customer service tickets… that it is experiencing difficulty responding to legitimate tickets.”

Screenshot of a social media post from Proven Industries.

Proven was quite proud of its lawsuit… at first.

Someone posted Lee’s personal phone number to the comment section of a McNally video, which soon led to “a continuous stream of harassing phone calls and text messages from unknown numbers at all hours of the day and night,” which included “profanity, threats, and racially charged language.”

Lest this seem like mere high spirits and hijinks, Lee’s partner and his mother both “received harassing messages through Facebook Messenger,” while other messages targeted Lee’s son, saying things like “I would kill your f—ing n—– child” and calling him a “racemixing pussy.”

This is clearly terrible behavior; it also has no obvious connection to McNally, who did not direct or condone the harassment. As for Lee’s phone number, McNally said that he had nothing to do with posting it and wrote that “it is my understanding that the phone number at issue is publicly available on the Better Business Bureau website and can be obtained through a simple Google search.”

And this, with both sides palpably angry at each other, is how things stood on June 13 at 9: 09 am, when the case got a hearing in front of the Honorable Mary Scriven, an extremely feisty federal judge in Tampa. Proven had demanded a preliminary injunction that would stop McNally from sharing his videos while the case progressed, but Proven had issues right from the opening gavel:

LAWYER 1: Austin Nowacki on behalf of Proven industries.

THE COURT: I’m sorry. What is your name?

LAWYER 1: Austin Nowacki.

THE COURT: I thought you said Austin No Idea.

LAWYER 2: That’s Austin Nowacki.

THE COURT: All right.

When Proven’s lead lawyer introduced a colleague who would lead that morning’s arguments, the judge snapped, “Okay. Then you have a seat and let her speak.”

Things went on this way for some time, as the judge wondered, “Did the plaintiff bring a lock and a beer can?” (The plaintiff did not.) She appeared to be quite disappointed when it was clear there would be no live shimming demonstration in the courtroom.

Then it was on to the actual arguments. Proven argued that the 15 seconds of its 90-second promo video used by McNally were not fair use, that McNally had defamed the company by implication, and that shimming its locks was actually quite difficult. Under questioning, however, one of Proven’s employees admitted that he had been able to duplicate McNally’s technique, leading to the question from McNally’s lawyer: “When you did it yourself, did it occur to you for one moment that maybe the best thing to do, instead of file a lawsuit, was to fix [the lock]?”

At the end of several hours of wrangling, the judge stepped in, saying that she “declines to grant the preliminary injunction motion.” For her to do so, Proven would have to show that it was likely to win at trial, among other things; it had not.

As for the big copyright infringement claim, of which Proven had made so much hay, the judge reached a pretty obvious finding: You’re allowed to quote snippets of copyrighted videos in order to critique them.

“The purpose and character of the use to which Mr. McNally put the alleged infringed work is transformative, artistic, and a critique,” said the judge. “He is in his own way challenging and critiquing Proven’s video by the use of his own video.”

As for the amount used, it was “substantial enough but no more than is necessary to make the point that he is trying to critique Proven’s video, and I think that’s fair game and a nominative fair use circumstance.”

While Proven might convince her otherwise after a full trial, “the copyright claim fails as a basis for a demand for preliminary injunctive relief.”

As for “tortious interference” and “defamation by implication,” the judge was similarly unimpressed.

“The fact that you might have a repeat customer who is dissuaded to buy your product due to a criticism of the product is not the type of business relationship the tortious interference with business relationship concept is intended to apply,” she said.

In the end, the judge said she would see the case through to its end, if that was really what everyone wanted, but “I will pray that you all come to a resolution of the case that doesn’t require all of this. This is a capitalist market and people say what they say. As long as it’s not false, they say what they say.”

She gave Proven until July 7 to amend its complaint if it wished.

On July 7, the company dismissed the lawsuit against McNally instead.

Proven also made a highly unusual request: Would the judge please seal almost the entire court record—including the request to seal?

Court records are presumptively public, but Proven complained about a “pattern of intimidation and harassment by individuals influenced by Defendant McNally’s content.” According to the company, a key witness had already backed out of the case, saying, “Is there a way to leave my name and my companies name out of this due to concerns of potential BLOW BACK from McNally or others like him?” Another witness, who did submit a declaration, wondered, “Is this going to be public? My concern is that there may be some backlash from the other side towards my company.”

McNally’s lawyer laid into this seal request, pointing out that the company had shown no concern over these issues until it lost its bid for a preliminary injunction. Indeed, “Proven boasted to its social media followers about how it sued McNally and about how confident it was that it would prevail. Proven even encouraged people to search for the lawsuit.” Now, however, the company “suddenly discover[ed] a need for secrecy.”

The judge has not yet ruled on the request to seal.

Another way

The strange thing about the whole situation is that Proven actually knew how to respond constructively to the first McNally video. Its own response video opened with a bit of humor (the presenter drinks a can of Liquid Death), acknowledged the issue (“we’ve had a little bit of controversy in the last couple days”), and made clear that Proven could handle criticism (“we aren’t afraid of a little bit of feedback”).

The video went on to show how their locks work and provided some context on shimming attacks and their likelihood of real-world use. It ended by showing how users concerned about shimming attacks could choose more expensive but more secure lock cores that should resist the technique.

Quick, professional, non-defensive—a great way to handle controversy.

But it was all blown apart by the company’s angry social media statements, which were unprofessional and defensive, and the litigation, which was spectacularly ill-conceived as a matter of both law and policy. In the end, the case became a classic example of the Streisand Effect, in which the attempt to censor information can instead call attention to it.

Judging from the number of times the lawsuit talks about 1) ridicule and 2) harassment, it seems like the case quickly became a personal one for Proven’s owner and employees, who felt either mocked or threatened. That’s understandable, but being mocked is not illegal and should never have led to a lawsuit or a copyright claim. As for online harassment, it remains a serious and unresolved issue, but launching a personal vendetta—and on pretty flimsy legal grounds—against McNally himself was patently unwise. (Doubly so given that McNally had a huge following and had already responded to DMCA takedowns by creating further videos on the subject; this wasn’t someone who would simply be intimidated by a lawsuit.)

In the end, Proven’s lawsuit likely cost the company serious time and cash—and generated little but bad publicity.

Photo of Nate Anderson

10M people watched a YouTuber shim a lock; the lock company sued him. Bad idea. Read More »

google-has-a-useful-quantum-algorithm-that-outperforms-a-supercomputer

Google has a useful quantum algorithm that outperforms a supercomputer

Computer science, Features, Google, NMR, Physics, quantum computing, quantum mechanics, Science / /

An approach it calls “quantum echoes” takes 13,000 times longer on a supercomputer.

Image of a silvery plate labeled with

The work relied on Google’s current-generation quantum hardware, the Willow chip. Credit: Google

The work relied on Google’s current-generation quantum hardware, the Willow chip. Credit: Google

A few years back, Google made waves when it claimed that some of its hardware had achieved quantum supremacy, performing operations that would be effectively impossible to simulate on a classical computer. That claim didn’t hold up especially well, as mathematicians later developed methods to help classical computers catch up, leading the company to repeat the work on an improved processor.

While this back-and-forth was unfolding, the field became less focused on quantum supremacy and more on two additional measures of success. The first is quantum utility, in which a quantum computer performs computations that are useful in some practical way. The second is quantum advantage, in which a quantum system completes calculations in a fraction of the time it would take a typical computer. (IBM and a startup called Pasqual have published a useful discussion about what would be required to verifiably demonstrate a quantum advantage.)

Today, Google and a large collection of academic collaborators are publishing a paper describing a computational approach that demonstrates a quantum advantage compared to current algorithms—and may actually help us achieve something useful.

Out of time

Google’s latest effort centers on something it’s calling “quantum echoes.” The approach could be described as a series of operations on the hardware qubits that make up its machine. These qubits hold a single bit of quantum information in a superposition between two values, with probabilities of finding the qubit in one value or the other when it’s measured. Each qubit is entangled with its neighbors, allowing its probability to influence those of all the qubits around it. The operations that allow computation, called gates, are ways of manipulating these probabilities. Most current hardware, including Google’s, perform manipulations on one or two qubits at a time (termed one- and two-qubit gates, respectively.

For quantum echoes, the operations involved performing a set of two-qubit gates, altering the state of the system, and later performing the reverse set of gates. On its own, this would return the system to its original state. But for quantum echoes, Google inserts single-qubit gates performed with a randomized parameter. This alters the state of the system before the reverse operations take place, ensuring that the system won’t return to exactly where it started. That explains the “echoes” portion of the name: You’re sending an imperfect copy back toward where things began, much like an echo involves the imperfect reversal of sound waves.

That’s what the process looks like in terms of operations performed on the quantum hardware. But it’s probably more informative to think of it in terms of a quantum system’s behavior. As Google’s Tim O’Brien explained, “You evolve the system forward in time, then you apply a small butterfly perturbation, and then you evolve the system backward in time.” The forward evolution is the first set of two qubit gates, the small perturbation is the randomized one qubit gate, and the second set of two qubit gates is the equivalent of sending the system backward in time.

Because this is a quantum system, however, strange things happen. “On a quantum computer, these forward and backward evolutions, they interfere with each other,” O’Brien said. One way to think about that interference is in terms of probabilities. The system has multiple paths between its start point and the point of reflection—where it goes from evolving forward in time to evolving backward—and from that reflection point back to a final state. Each of those paths has a probability associated with it. And since we’re talking about quantum mechanics, those paths can interfere with each other, increasing some probabilities at the expense of others. That interference ultimately determines where the system ends up.

(Technically, these are termed “out of time order correlations,” or OTOCs. If you read the Nature paper describing this work, prepare to see that term a lot.)

Demonstrating advantage

So how do you turn quantum echoes into an algorithm? On its own, a single “echo” can’t tell you much about the system—the probabilities ensure that any two runs might show different behaviors. But if you repeat the operations multiple times, you can begin to understand the details of this quantum interference. And performing the operations on a quantum computer ensures that it’s easy to simply rerun the operations with different random one-qubit gates and get many instances of the initial and final states—and thus a sense of the probability distributions involved.

This is also where Google’s quantum advantage comes from. Everyone involved agrees that the precise behavior of a quantum echo of moderate complexity can be modeled using any leading supercomputer. But doing so is very time-consuming, so repeating those simulations a few times becomes unrealistic. The paper estimates that a measurement that took its quantum computer 2.1 hours to perform would take the Frontier supercomputer approximately 3.2 years. Unless someone devises a far better classical algorithm than what we have today, this represents a pretty solid quantum advantage.

But is it a useful algorithm? The repeated sampling can act a bit like the Monte Carlo sampling done to explore the behavior of a wide variety of physical systems. Typically, however, we don’t view algorithms as modeling the behavior of the underlying hardware they’re being run on; instead, they’re meant to model some other physical system we’re interested in. That’s where Google’s announcement stands apart from its earlier work—the company believes it has identified an interesting real-world physical system with behaviors that the quantum echoes can help us understand.

That system is a small molecule in a Nuclear Magnetic Resonance (NMR) machine. In a second draft paper being published on the arXiv later today, Google has collaborated with a large collection of NMR experts to explore that use.

From computers to molecules

NMR is based on the fact that the nucleus of every atom has a quantum property called spin. When nuclei are held near to each other, such as when they’re in the same molecule, these spins can influence one another. NMR uses magnetic fields and photons to manipulate these spins and can be used to infer structural details, like how far apart two given atoms are. But as molecules get larger, these spin networks can extend for greater distances and become increasingly complicated to model. So NMR has been limited to focusing on the interactions of relatively nearby spins.

For this work, though, the researchers figured out how to use an NMR machine to create the physical equivalent of a quantum echo in a molecule. The work involved synthesizing the molecule with a specific isotope of carbon (carbon-13) in a known location in the molecule. That isotope could be used as the source of a signal that propagates through the network of spins formed by the molecule’s atoms.

“The OTOC experiment is based on a many-body echo, in which polarization initially localized on a target spin migrates through the spin network, before a Hamiltonian-engineered time-reversal refocuses to the initial state,” the team wrote. “This refocusing is sensitive to perturbations on distant butterfly spins, which allows one to measure the extent of polarization propagation through the spin network.”

Naturally, something this complicated needed a catchy nickname. The team came up with TARDIS, or Time-Accurate Reversal of Dipolar InteractionS. While that name captures the “out of time order” aspect of OTOC, it’s simply a set of control pulses sent to the NMR sample that starts a perturbation of the molecule’s network of nuclear spins. A second set of pulses then reflects an echo back to the source.

The reflections that return are imperfect, with noise coming from two sources. The first is simply imperfections in the control sequence, a limitation of the NMR hardware. But the second is the influence of fluctuations happening in distant atoms along the spin network. These happen at a certain frequency at random, or the researchers could insert a fluctuation by targeting a specific part of the molecule with randomized control signals.

The influence of what’s going on in these distant spins could allow us to use quantum echoes to tease out structural information at greater distances than we currently do with NMR. But to do so, we need an accurate model of how the echoes will propagate through the molecule. And again, that’s difficult to do with classical computations. But it’s very much within the capabilities of quantum computing, which the paper demonstrates.

Where things stand

For now, the team stuck to demonstrations on very simple molecules, making this work mostly a proof of concept. But the researchers are optimistic that there are many ways the system could be used to extract structural information from molecules at distances that are currently unobtainable using NMR. They list a lot of potential upsides that should be explored in the discussion of the paper, and there are plenty of smart people who would love to find new ways of using their NMR machines, so the field is likely to figure out pretty quickly which of these approaches turns out to be practically useful.

The fact that the demonstrations were done with small molecules, however, means that the modeling run on the quantum computer could also have been done on classical hardware (it only required 15 hardware qubits). So Google is claiming both quantum advantage and quantum utility, but not at the same time. The sorts of complex, long-distance interactions that would be out of range of classical simulation are still a bit beyond the reach of the current quantum hardware. O’Brien estimated that the hardware’s fidelity would have to improve by a factor of three or four to model molecules that are beyond classical simulation.

The quantum advantage issue should also be seen as a work in progress. Google has collaborated with enough researchers at enough institutions that there’s unlikely to be a major improvement in algorithms that could allow classical computers to catch up. Until the community as a whole has some time to digest the announcement, though, we shouldn’t take that as a given.

The other issue is verifiability. Some quantum algorithms will produce results that can be easily verified on classical hardware—situations where it’s hard to calculate the right result but easy to confirm a correct answer. Quantum echoes isn’t one of those, so we’ll need another quantum computer to verify the behavior Google has described.

But Google told Ars nothing is up to the task yet. “No other quantum processor currently matches both the error rates and number of qubits of our system, so our quantum computer is the only one capable of doing this at present,” the company said. (For context, Google says that the algorithm was run on up to 65 qubits, but the chip has 105 qubits total.)

There’s a good chance that other companies would disagree with that contention, but it hasn’t been possible to ask them ahead of the paper’s release.

In any case, even if this claim proves controversial, Google’s Michel Devoret, a recent Nobel winner, hinted that we shouldn’t have long to wait for additional ones. “We have other algorithms in the pipeline, so we will hopefully see other interesting quantum algorithms,” Devoret said.

Nature, 2025. DOI: 10.1038/s41586-025-09526-6  (About DOIs).

Photo of John Timmer

John is Ars Technica’s science editor. He has a Bachelor of Arts in Biochemistry from Columbia University, and a Ph.D. in Molecular and Cell Biology from the University of California, Berkeley. When physically separated from his keyboard, he tends to seek out a bicycle, or a scenic location for communing with his hiking boots.

Google has a useful quantum algorithm that outperforms a supercomputer Read More »

should-an-ai-copy-of-you-help-decide-if-you-live-or-die?

Should an AI copy of you help decide if you live or die?

AI, ai clone, ai surrogate, Artificial Intelligence, chatbot, Features, healthcare ai, medical ai / /

“It would combine demographic and clinical variables, documented advance-care-planning data, patient-recorded values and goals, and contextual information about specific decisions,” he said.

“Including textual and conversational data could further increase a model’s ability to learn why preferences arise and change, not just what a patient’s preference was at a single point in time,” Starke said.

Ahmad suggested that future research could focus on validating fairness frameworks in clinical trials, evaluating moral trade-offs through simulations, and exploring how cross-cultural bioethics can be combined with AI designs.

Only then might AI surrogates be ready to be deployed, but only as “decision aids,” Ahmad wrote. Any “contested outputs” should automatically “trigger [an] ethics review,” Ahmad wrote, concluding that “the fairest AI surrogate is one that invites conversation, admits doubt, and leaves room for care.”

“AI will not absolve us”

Ahmad is hoping to test his conceptual models at various UW sites over the next five years, which would offer “some way to quantify how good this technology is,” he said.

“After that, I think there’s a collective decision regarding how as a society we decide to integrate or not integrate something like this,” Ahmad said.

In his paper, he warned against chatbot AI surrogates that could be interpreted as a simulation of the patient, predicting that future models may even speak in patients’ voices and suggesting that the “comfort and familiarity” of such tools might blur “the boundary between assistance and emotional manipulation.”

Starke agreed that more research and “richer conversations” between patients and doctors are needed.

“We should be cautious not to apply AI indiscriminately as a solution in search of a problem,” Starke said. “AI will not absolve us from making difficult ethical decisions, especially decisions concerning life and death.”

Truog, the bioethics expert, told Ars he “could imagine that AI could” one day “provide a surrogate decision maker with some interesting information, and it would be helpful.”

But a “problem with all of these pathways… is that they frame the decision of whether to perform CPR as a binary choice, regardless of context or the circumstances of the cardiac arrest,” Truog’s editorial said. “In the real world, the answer to the question of whether the patient would want to have CPR” when they’ve lost consciousness, “in almost all cases,” is “it depends.”

When Truog thinks about the kinds of situations he could end up in, he knows he wouldn’t just be considering his own values, health, and quality of life. His choice “might depend on what my children thought” or “what the financial consequences would be on the details of what my prognosis would be,” he told Ars.

“I would want my wife or another person that knew me well to be making those decisions,” Truog said. “I wouldn’t want somebody to say, ‘Well, here’s what AI told us about it.’”

Should an AI copy of you help decide if you live or die? Read More »

yes,-everything-online-sucks-now—but-it-doesn’t-have-to

Yes, everything online sucks now—but it doesn’t have to

books, Cory Doctorow, culture, enshittification, Features, online culture, Tech / /

from good to bad to nothing

Ars chats with Cory Doctorow about his new book Enshittification.

We all feel it: Our once-happy digital spaces have become increasingly less user-friendly and more toxic, cluttered with extras nobody asked for and hardly anybody wants. There’s even a word for it: “enshittification,” named 2023 Word of the Year by the American Dialect Society. The term was coined by tech journalist/science fiction author Cory Doctorow, a longtime advocate of digital rights. Doctorow has spun his analysis of what’s been ailing the tech industry into an eminently readable new book, Enshittification: Why Everything Suddenly Got Worse and What To Do About It.

As Doctorow tells it, he was on vacation in Puerto Rico, staying in a remote cabin nestled in a cloud forest with microwave Internet service—i.e., very bad Internet service, since microwave signals struggle to penetrate through clouds. It was a 90-minute drive to town, but when they tried to consult TripAdvisor for good local places to have dinner one night, they couldn’t get the site to load. “All you would get is the little TripAdvisor logo as an SVG filling your whole tab and nothing else,” Doctorow told Ars. “So I tweeted, ‘Has anyone at TripAdvisor ever been on a trip? This is the most enshittified website I’ve ever used.’”

Initially, he just got a few “haha, that’s a funny word” responses. “It was when I married that to this technical critique, at a moment when things were quite visibly bad to a much larger group of people, that made it take off,” Doctorow said. “I didn’t deliberately set out to do it. I bought a million lottery tickets and one of them won the lottery. It only took two decades.”

Yes, people sometimes express regret to him that the term includes a swear word. To which he responds, “You’re welcome to come up with another word. I’ve tried. ‘Platform decay’ just isn’t as good.” (“Encrapification” and “enpoopification” also lack a certain je ne sais quoi.)

In fact, it’s the sweariness that people love about the word. While that also means his book title inevitably gets bleeped on broadcast radio, “The hosts, in my experience, love getting their engineers to creatively bleep it,” said Doctorow. “They find it funny. It’s good radio, it stands out when every fifth word is ‘enbeepification.’”

People generally use “enshittification” colloquially to mean “the degradation in the quality and experience of online platforms over time.” Doctorow’s definition is more specific, encompassing “why an online service gets worse, how that worsening unfolds,” and how this process spreads to other online services, such that everything is getting worse all at once.

For Doctorow, enshittification is a disease with symptoms, a mechanism, and an epidemiology. It has infected everything from Facebook, Twitter, Amazon, and Google, to Airbnb, dating apps, iPhones, and everything in between. “For me, the fact that there were a lot of platforms that were going through this at the same time is one of the most interesting and important factors in the critique,” he said. “It makes this a structural issue and not a series of individual issues.”

It starts with the creation of a new two-sided online product of high quality, initially offered at a loss to attract users—say, Facebook, to pick an obvious example. Once the users are hooked on the product, the vendor moves to the second stage: degrading the product in some way for the benefit of their business customers. This might include selling advertisements, scraping and/or selling user data, or tweaking algorithms to prioritize content the vendor wishes users to see rather than what those users actually want.

This locks in the business customers, who, in turn, invest heavily in that product, such as media companies that started Facebook pages to promote their published content. Once business customers are locked in, the vendor can degrade those services too—i.e., by de-emphasizing news and links away from Facebook—to maximize profits to shareholders. Voila! The product is now enshittified.

The four horsemen of the shitocalypse

Doctorow identifies four key factors that have played a role in ushering in an era that he has dubbed the “Enshittocene.” The first is competition (markets), in which companies are motivated to make good products at affordable prices, with good working conditions, because otherwise customers and workers will go to their competitors.  The second is government regulation, such as antitrust laws that serve to keep corporate consolidation in check, or levying fines for dishonest practices, which makes it unprofitable to cheat.

The third is interoperability: the inherent flexibility of digital tools, which can play a useful adversarial role. “The fact that enshittification can always be reversed with a dis-enshittifiting counter-technology always acted as a brake on the worst impulses of tech companies,” Doctorow writes. Finally, there is labor power; in the case of the tech industry, highly skilled workers were scarce and thus had considerable leverage over employers.

All four factors, when functioning correctly, should serve as constraints to enshittification. However, “One by one each enshittification restraint was eroded until it dissolved, leaving the enshittification impulse unchecked,” Doctorow writes. Any “cure” will require reversing those well-established trends.

But isn’t all this just the nature of capitalism? Doctorow thinks it’s not, arguing that the aforementioned weakening of traditional constraints has resulted in the usual profit-seeking behavior producing very different, enshittified outcomes. “Adam Smith has this famous passage in Wealth of Nations about how it’s not due to the generosity of the baker that we get our bread but to his own self-regard,” said Doctorow. “It’s the fear that you’ll get your bread somewhere else that makes him keep prices low and keep quality high. It’s the fear of his employees leaving that makes him pay them a fair wage. It is the constraints that causes firms to behave better. You don’t have to believe that everything should be a capitalist or a for-profit enterprise to acknowledge that that’s true.”

Our wide-ranging conversation below has been edited for length to highlight the main points of discussion.

Ars Technica: I was intrigued by your choice of framing device, discussing enshittification as a form of contagion. 

Cory Doctorow: I’m on a constant search for different framing devices for these complex arguments. I have talked about enshittification in lots of different ways. That frame was one that resonated with people. I’ve been a blogger for a quarter of a century, and instead of keeping notes to myself, I make notes in public, and I write up what I think is important about something that has entered my mind, for better or for worse. The downside is that you’re constantly getting feedback that can be a little overwhelming. The upside is that you’re constantly getting feedback, and if you pay attention, it tells you where to go next, what to double down on.

Another way of organizing this is the Galaxy Brain meme, where the tiny brain is “Oh, this is because consumers shopped wrong.” The medium brain is “This is because VCs are greedy.” The larger brain is “This is because tech bosses are assholes.” But the biggest brain of all is “This is because policymakers created the policy environment where greed can ruin our lives.” There’s probably never going to be just one way to talk about this stuff that lands with everyone. So I like using a variety of approaches. I suck at being on message. I’m not going to do Enshittification for the Soul and Mornings with Enshittifying Maury. I am restless, and my Myers-Briggs type is ADHD, and I want to have a lot of different ways of talking about this stuff.

Ars Technica: One site that hasn’t (yet) succumbed is Wikipedia. What has protected Wikipedia thus far? 

Cory Doctorow: Wikipedia is an amazing example of what we at the Electronic Frontier Foundation (EFF) call the public interest Internet. Internet Archive is another one. Most of these public interest Internet services start off as one person’s labor of love, and that person ends up being what we affectionately call the benevolent dictator for life. Very few of these projects have seen the benevolent dictator for life say, “Actually, this is too important for one person to run. I cannot be the keeper of the soul of this project. I am prone to self-deception and folly just like every other person. This needs to belong to its community.” Wikipedia is one of them. The founder, my friend Jimmy Wales, woke up one day and said, “No individual should run Wikipedia. It should be a communal effort.”

There’s a much more durable and thick constraint on the decisions of anyone at Wikipedia to do something bad. For example, Jimmy had this idea that you could use AI in Wikipedia to help people make entries and navigate Wikipedia’s policies, which are daunting. The community evaluated his arguments and decided—not in a reactionary way, but in a really thoughtful way—that this was wrong. Jimmy didn’t get his way. It didn’t rule out something in the future, but that’s not happening now. That’s pretty cool.

Wikipedia is not just governed by a board; it’s also structured as a nonprofit. That doesn’t mean that there’s no way it could go bad. But it’s a source of friction against enshittification. Wikipedia has its entire corpus irrevocably licensed as the most open it can be without actually being in the public domain. Even if someone were to capture Wikipedia, there’s limits on what they could do to it.

There’s also a labor constraint in Wikipedia in that there’s very little that the leadership can do without bringing along a critical mass of a large and diffuse body of volunteers. That cuts against the volunteers working in unison—they’re not represented by a union; it’s hard for them to push back with one voice. But because they’re so diffuse and because there’s no paychecks involved, it’s really hard for management to do bad things. So if there are two people vying for the job of running the Wikimedia Foundation and one of them has got nefarious plans and the other doesn’t, the nefarious plan person, if they’re smart, is going to give it up—because if they try to squeeze Wikipedia, the harder they squeeze, the more it will slip through their grasp.

So these are structural defenses against enshittification of Wikipedia. I don’t know that it was in the mechanism design—I think they just got lucky—but it is a template for how to run such a project. It does raise this question: How do you build the community? But if you have a community of volunteers around a project, it’s a model of how to turn that project over to that community.

Ars Technica: Your case studies naturally include the decay of social media, notably Facebook and the social media site formerly known as Twitter. How might newer social media platforms resist the spiral into “platform decay”?

Cory Doctorow: What you want is a foundation in which people on social media face few switching costs. If the social media is interoperable, if it’s federatable, then it’s much harder for management to make decisions that are antithetical to the interests of users. If they do, users can escape. And it sets up an internal dynamic within the firm, where the people who have good ideas don’t get shouted down by the people who have bad but more profitable ideas, because it makes those bad ideas unprofitable. It creates both short and long-term risks to the bottom line.

There has to be a structure that stops their investors from pressurizing them into doing bad things, that stops them from rationalizing their way into complying. I think there’s this pathology where you start a company, you convince 150 of your friends to risk their kids’ college fund and their mortgage working for you. You make millions of users really happy, and your investors come along and say, “You have to destroy the life of 5 percent of your users with some change.” And you’re like, “Well, I guess the right thing to do here is to sacrifice those 5 percent, keep the other 95 percent happy, and live to fight another day, because I’m a good guy. If I quit over this, they’ll just put a bad guy in who’ll wreck things. I keep those 150 people working. Not only that, I’m kind of a martyr because everyone thinks I’m a dick for doing this. No one understands that I have taken the tough decision.”

I think that’s a common pattern among people who, in fact, are quite ethical but are also capable of rationalizing their way into bad things. I am very capable of rationalizing my way into bad things. This is not an indictment of someone’s character. But it’s why, before you go on a diet, you throw away the Oreos. It’s why you bind yourself to what behavioral economists call “Ulysses pacts“: You tie yourself to the mast before you go into the sea of sirens, not because you’re weak but because you’re strong enough now to know that you’ll be weak in the future.

I have what I would call the epistemic humility to say that I don’t know what makes a good social media network, but I do know what makes it so that when they go bad, you’re not stuck there. You and I might want totally different things out of our social media experience, but I think that you should 100 percent have the right to go somewhere else without losing anything. The easier it is for you to go without losing something, the better it is for all of us.

My dream is a social media universe where knowing what network someone is using is just a weird curiosity. It’d be like knowing which cell phone carrier your friend is using when you give them a call. It should just not matter. There might be regional or technical reasons to use one network or another, but it shouldn’t matter to anyone other than the user what network they’re using. A social media platform where it’s always easier for users to leave is much more future-proof and much more effective than trying to design characteristics of good social media.

Ars Technica: How might this work in practice?

Cory Doctorow: I think you just need a protocol. This is [Mike] Maznik’s point: protocols, not products. We don’t need a universal app to make email work. We don’t need a universal app to make the web work. I always think about this in the context of administrable regulation. Making a rule that says your social media network must be good for people to use and must not harm their mental health is impossible. The fact intensivity of determining whether a platform satisfies that rule makes it a non-starter.

Whereas if you were to say, “OK, you have to support an existing federation protocol, like AT Protocol and Mastodon ActivityPub,” both have ways to port identity from one place to another and have messages auto-forward. This is also in RSS. There’s a permanent redirect directive. You do that, you’re in compliance with the regulation.

Or you have to do something that satisfies the functional requirements of the spec. So it’s not “did you make someone sad in a way that was reckless?” That is a very hard question to adjudicate. Did you satisfy these functional requirements? It’s not easy to answer that, but it’s not impossible. If you want to have our users be able to move to your platform, then you just have to support the spec that we’ve come up with, which satisfies these functional requirements.

We don’t have to have just one protocol. We can have multiple ones. Not everything has to connect to everything else, but everyone who wants to connect should be able to connect to everyone else who wants to connect. That’s end-to-end. End-to-end is not “you are required to listen to everything someone wants to tell you.” It’s that willing parties should be connected when they want to be.

Ars Technica: What about security and privacy protocols like GPG and PGP?

Cory Doctorow: There’s this argument that the reason GPG is so hard to use is that it’s intrinsic; you need a closed system to make it work. But also, until pretty recently, GPG was supported by one part-time guy in Germany who got 30,000 euros a year in donations to work on it, and he was supporting 20 million users. He was primarily interested in making sure the system was secure rather than making it usable. If you were to put Big Tech quantities of money behind improving ease of use for GPG, maybe you decide it’s a dead end because it is a 30-year-old attempt to stick a security layer on top of SMTP. Maybe there’s better ways of doing it. But I doubt that we have reached the apex of GPG usability with one part-time volunteer.

I just think there’s plenty of room there. If you have a pretty good project that is run by a large firm and has had billions of dollars put into it, the most advanced technologists and UI experts working on it, and you’ve got another project that has never been funded and has only had one volunteer on it—I would assume that dedicating resources to that second one would produce pretty substantial dividends, whereas the first one is only going to produce these minor tweaks. How much more usable does iOS get with every iteration?

I don’t know if PGP is the right place to start to make privacy, but I do think that if we can create independence of the security layer from the transport layer, which is what PGP is trying to do, then it wouldn’t matter so much that there is end-to-end encryption in Mastodon DMs or in Bluesky DMs. And again, it doesn’t matter whose sim is in your phone, so it just shouldn’t matter which platform you’re using so long as it’s secure and reliably delivered end-to-end.

Ars Technica: These days, I’m almost contractually required to ask about AI. There’s no escaping it. But it’s certainly part of the ongoing enshittification.

Cory Doctorow: I agree. Again, the companies are too big to care. They know you’re locked in, and the things that make enshittification possible—like remote software updating, ongoing analytics of use of devices—they allow for the most annoying AI dysfunction. I call it the fat-finger economy, where you have someone who works in a company on a product team, and their KPI, and therefore their bonus and compensation, is tied to getting you to use AI a certain number of times. So they just look at the analytics for the app and they ask, “What button gets pushed the most often? Let’s move that button somewhere else and make an AI summoning button.”

They’re just gaming a metric. It’s causing significant across-the-board regressions in the quality of the product, and I don’t think it’s justified by people who then discover a new use for the AI. That’s a paternalistic justification. The user doesn’t know what they want until you show it to them: “Oh, if I trick you into using it and you keep using it, then I have actually done you a favor.” I don’t think that’s happening. I don’t think people are like, “Oh, rather than press reply to a message and then type a message, I can instead have this interaction with an AI about how to send someone a message about takeout for dinner tonight.” I think people are like, “That was terrible. I regret having tapped it.” 

The speech-to-text is unusable now. I flatter myself that my spoken and written communication is not statistically average. The things that make it me and that make it worth having, as opposed to just a series of multiple-choice answers, is all the ways in which it diverges from statistical averages. Back when the model was stupider, when it gave up sooner if it didn’t recognize what word it might be and just transcribed what it thought you’d said rather than trying to substitute a more probable word, it was more accurate.  Now, what I’m getting are statistically average words that are meaningless.

That elision of nuance and detail is characteristic of what makes AI products bad. There is a bunch of stuff that AI is good at that I’m excited about, and I think a lot of it is going to survive the bubble popping. But I fear that we’re not planning for that. I fear what we’re doing is taking workers whose jobs are meaningful, replacing them with AIs that can’t do their jobs, and then those AIs are going to go away and we’ll have nothing. That’s my concern.

Ars Technica: You prescribe a “cure” for enshittification, but in such a polarized political environment, do we even have the collective will to implement the necessary policies?

Cory Doctorow: The good news is also the bad news, which is that this doesn’t just affect tech. Take labor power. There are a lot of tech workers who are looking at the way their bosses treat the workers they’re not afraid of—Amazon warehouse workers and drivers, Chinese assembly line manufacturers for iPhones—and realizing, “Oh, wait, when my boss stops being afraid of me, this is how he’s going to treat me.” Mark Zuckerberg stopped going to those all-hands town hall meetings with the engineering staff. He’s not pretending that you are his peers anymore. He doesn’t need to; he’s got a critical mass of unemployed workers he can tap into. I think a lot of Googlers figured this out after the 12,000-person layoffs. Tech workers are realizing they missed an opportunity, that they’re going to have to play catch-up, and that the only way to get there is by solidarity with other kinds of workers.

The same goes for competition. There’s a bunch of people who care about media, who are watching Warner about to swallow Paramount and who are saying, “Oh, this is bad. We need antitrust enforcement here.” When we had a functional antitrust system for the last four years, we saw a bunch of telecoms mergers stopped because once you start enforcing antitrust, it’s like eating Pringles. You just can’t stop. You embolden a lot of people to start thinking about market structure as a source of either good or bad policy. The real thing that happened with [former FTC chair] Lina Kahn doing all that merger scrutiny was that people just stopped planning mergers.

There are a lot of people who benefit from this. It’s not just tech workers or tech users; it’s not just media users. Hospital consolidation, pharmaceutical consolidation, has a lot of people who are very concerned about it. Mark Cuban is freaking out about pharmacy benefit manager consolidation and vertical integration with HMOs, as he should be. I don’t think that we’re just asking the anti-enshittification world to carry this weight.

Same with the other factors. The best progress we’ve seen on interoperability has been through right-to-repair. It hasn’t been through people who care about social media interoperability. One of the first really good state-level right-to-repair bills was the one that [Governor] Jared Polis signed in Colorado for powered wheelchairs. Those people have a story that is much more salient to normies.

What do you mean you spent six months in bed because there’s only two powered wheelchair manufacturers and your chair broke and you weren’t allowed to get it fixed by a third party?” And they’ve slashed their repair department, so it takes six months for someone to show up and fix your chair. So you had bed sores and pneumonia because you couldn’t get your chair fixed. This is bullshit.

So the coalitions are quite large. The thing that all of those forces share—interoperability, labor power, regulation, and competition—is that they’re all downstream of corporate consolidation and wealth inequality. Figuring out how to bring all of those different voices together, that’s how we resolve this. In many ways, the enshittification analysis and remedy are a human factors and security approach to designing an enshittification-resistant Internet. It’s about understanding this as a red team, blue team exercise. How do we challenge the status quo that we have now, and how do we defend the status quo that we want?

Anything that can’t go on forever eventually stops. That is the first law of finance, Stein’s law. We are reaching multiple breaking points, and the question is whether we reach things like breaking points for the climate and for our political system before we reach breaking points for the forces that would rescue those from permanent destruction.

Photo of Jennifer Ouellette

Jennifer is a senior writer at Ars Technica with a particular focus on where science meets culture, covering everything from physics and related interdisciplinary topics to her favorite films and TV series. Jennifer lives in Baltimore with her spouse, physicist Sean M. Carroll, and their two cats, Ariel and Caliban.

Yes, everything online sucks now—but it doesn’t have to Read More »

why-signal’s-post-quantum-makeover-is-an-amazing-engineering-achievement

Why Signal’s post-quantum makeover is an amazing engineering achievement

Biz & IT, diffie-hellman elliptic curve, Features, ML-KEM, Module-Lattice-Based Key-Encapsulation Mechanism, post-quantum encryption, Security, signal messenger / /

COMING TO A PHONE NEAR YOU

New design sets a high standard for post-quantum readiness.

Credit: Aurich Lawson | Getty Images

Credit: Aurich Lawson | Getty Images

The encryption protecting communications against criminal and nation-state snooping is under threat. As private industry and governments get closer to building useful quantum computers, the algorithms protecting Bitcoin wallets, encrypted web visits, and other sensitive secrets will be useless. No one doubts the day will come, but as the now-common joke in cryptography circles observes, experts have been forecasting this cryptocalypse will arrive in the next 15 to 30 years for the past 30 years.

The uncertainty has created something of an existential dilemma: Should network architects spend the billions of dollars required to wean themselves off quantum-vulnerable algorithms now, or should they prioritize their limited security budgets fighting more immediate threats such as ransomware and espionage attacks? Given the expense and no clear deadline, it’s little wonder that less than half of all TLS connections made inside the Cloudflare network and only 18 percent of Fortune 500 networks support quantum-resistant TLS connections. It’s all but certain that many fewer organizations still are supporting quantum-ready encryption in less prominent protocols.

Triumph of the cypherpunks

One exception to the industry-wide lethargy is the engineering team that designs the Signal Protocol, the open source engine that powers the world’s most robust and resilient form of end-to-end encryption for multiple private chat apps, most notably the Signal Messenger. Eleven days ago, the nonprofit entity that develops the protocol, Signal Messenger LLC, published a 5,900-word write-up describing its latest updates that make Signal fully quantum-resistant.

The complexity and problem-solving required for making the Signal Protocol quantum safe are as daunting as just about any in modern-day engineering. The original Signal Protocol already resembled the inside of a fine Swiss timepiece, with countless gears, wheels, springs, hands, and other parts all interoperating in an intricate way. In less adept hands, mucking about with an instrument as complex as the Signal protocol could have led to shortcuts or unintended consequences that hurt performance, undoing what would otherwise be a perfectly running watch. Yet this latest post-quantum upgrade (the first one came in 2023) is nothing short of a triumph.

“This appears to be a solid, thoughtful improvement to the existing Signal Protocol,” said Brian LaMacchia, a cryptography engineer who oversaw Microsoft’s post-quantum transition from 2015 to 2022 and now works at Farcaster Consulting Group. “As part of this work, Signal has done some interesting optimization under the hood so as to minimize the network performance impact of adding the post-quantum feature.”

Of the multiple hurdles to clear, the most challenging was accounting for the much larger key sizes that quantum-resistant algorithms require. The overhaul here adds protections based on ML-KEM-768, an implementation of the CRYSTALS-Kyber algorithm that was selected in 2022 and formalized last year by the National Institute of Standards and Technology. ML-KEM is short for Module-Lattice-Based Key-Encapsulation Mechanism, but most of the time, cryptographers refer to it simply as KEM.

Ratchets, ping-pong, and asynchrony

Like the Elliptic curve Diffie-Hellman (ECDH) protocol that Signal has used since its start, KEM is a key encapsulation mechanism. Also known as a key agreement mechanism, it provides the means for two parties who have never met to securely agree on one or more shared secrets in the presence of an adversary who is monitoring the parties’ connection. RSA, ECDH, and other encapsulation algorithms have long been used to negotiate symmetric keys (almost always AES keys) in protocols including TLS, SSH, and IKE. Unlike ECDH and RSA, however, the much newer KEM is quantum-safe.

Key agreement in a protocol like TLS is relatively straightforward. That’s because devices connecting over TLS negotiate a key over a single handshake that occurs at the beginning of a session. The agreed-upon AES key is then used throughout the session. The Signal Protocol is different. Unlike TLS sessions, Signal sessions are protected by forward secrecy, a cryptographic property that ensures the compromise of a key used to encrypt a recent set of messages can’t be used to decrypt an earlier set of messages. The protocol also offers Post-Compromise Security, which protects future messages from past key compromises. While a TLS  uses the same key throughout a session, keys within a Signal session constantly evolve.

To provide these confidentiality guarantees, the Signal Protocol updates secret key material each time a message party hits the send button or receives a message, and at other points, such as in graphical indicators that a party is currently typing and in the sending of read receipts. The mechanism that has made this constant key evolution possible over the past decade is what protocol developers call a “double ratchet.” Just as a traditional ratchet allows a gear to rotate in one direction but not in the other, the Signal ratchets allow messaging parties to create new keys based on a combination of preceding and newly agreed-upon secrets. The ratchets work in a single direction, the sending and receiving of future messages. Even if an adversary compromises a newly created secret, messages encrypted using older secrets can’t be decrypted.

The starting point is a handshake that performs three or four ECDH agreements that mix long- and short-term secrets to establish a shared secret. The creation of this “root key” allows the Double Ratchet to begin. Until 2023, the key agreement used X3DH. The handshake now uses PQXDH to make the handshake quantum-resistant.

The first layer of the Double Ratchet, the Symmetric Ratchet, derives an AES key from the root key and advances it for every message sent. This allows every message to be encrypted with a new secret key. Consequently, if attackers compromise one party’s device, they won’t be able to learn anything about the keys that came earlier. Even then, though, the attackers would still be able to compute the keys used in future messages. That’s where the second, “Diffie-Hellman ratchet” comes in.

The Diffie-Hellman ratchet incorporates a new ECDH public key into each message sent. Using Alice and Bob, the fictional characters often referred to when explaining asymmetric encryption, when Alice sends Bob a message, she creates a new ratchet keypair and computes the ECDH agreement between this key and the last ratchet public key Bob sent. This gives her a new secret, and she knows that once Bob gets her new public key, he will know this secret, too (because, as mentioned earlier, Bob previously sent that other key). With that, Alice can mix the new secret with her old root key to get a new root key and start fresh. The result: Attackers who learn her old secrets won’t be able to tell the difference between her new ratchet keys and random noise.

The result is what Signal developers describe as “ping-pong” behavior, as the parties to a discussion take turns replacing ratchet key pairs one at a time. The effect: An eavesdropper who compromises one of the parties might recover a current ratchet private key, but soon enough, that private key will be replaced with a new, uncompromised one, and in a way that keeps it free from the prying eyes of the attacker.

The objective of the newly generated keys is to limit the number of messages that can be decrypted if an adversary recovers key material at some point in an ongoing chat. Messages sent prior to and after the compromise will remain off limits.

A major challenge designers of the Signal Protocol face is the need to make the ratchets work in an asynchronous environment. Asynchronous messages occur when parties send or receive them at different times—such as while one is offline and the other is active, or vice versa—without either needing to be present or respond immediately. The entire Signal Protocol must work within this asynchronous environment. What’s more, it must work reliably over unstable networks and networks controlled by adversaries, such as a government that forces a telecom or cloud service to spy on the traffic.

Shor’s algorithm lurking

By all accounts, Signal’s double ratchet design is state-of-the-art. That said, it’s wide open to an inevitable if not immediate threat: quantum computing. That’s because an adversary capable of monitoring traffic passing from two or more messenger users can capture that data and feed it into a quantum computer—once one of sufficient power is viable—and calculate the ephemeral keys generated in the second ratchet.

In classical computing, it’s infeasible, if not impossible, for such an adversary to calculate the key. Like all asymmetric encryption algorithms, ECDH is based on a mathematical, one-way function. Also known as trapdoor functions, these problems are trivial to compute in one direction and substantially harder to compute in reverse. In elliptic curve cryptography, this one-way function is based on the Discrete Logarithm problem in mathematics. The key parameters are based on specific points in an elliptic curve over the field of integers modulo some prime P.

On average, an adversary equipped with only a classical computer would spend billions of years guessing integers before arriving at the right ones. A quantum computer, by contrast, would be able to calculate the correct integers in a matter of hours or days. A formula known as Shor’s algorithm—which runs only on a quantum computer—reverts this one-way discrete logarithm equation to a two-way one. Shor’s Algorithm can similarly make quick work of solving the one-way function that’s the basis for the RSA algorithm.

As noted earlier, the Signal Protocol received its first post-quantum makeover in 2023. This update added PQXDH—a Signal-specific implementation that combined the key agreements from elliptic curves used in X3DH (specifically X25519) and the quantum-safe KEM—in the initial protocol handshake. (X3DH was then put out to pasture as a standalone implementation.)

The move foreclosed the possibility of a quantum attack being able to recover the symmetric key used to start the ratchets, but the ephemeral keys established in the ping-ponging second ratchet remained vulnerable to a quantum attack. Signal’s latest update adds quantum resistance to these keys, ensuring that forward secrecy and post-compromise security are safe from Shor’s algorithm as well.

Even though the ping-ponging keys are vulnerable to future quantum attacks, they are broadly believed to be secure against today’s attacks from classical computers. The Signal Protocol developers didn’t want to remove them or the battle-tested code that produces them. That led to their decision to add quantum resistance by adding a third ratchet. This one uses a quantum-safe KEM to produce new secrets much like the Diffie-Hellman ratchet did before, ensuring quantum-safe, post-compromise security.

The technical challenges were anything but easy. Elliptic curve keys generated in the X25519 implementation are about 32 bytes long, small enough to be added to each message without creating a burden on already constrained bandwidths or computing resources. A ML-KEM 768 key, by contrast, is 1,000 bytes. Additionally, Signal’s design requires sending both an encryption key and a ciphertext, making the total size 2272 bytes.

And then there were three

To handle the 71x increase, Signal developers considered a variety of options. One was to send the 2272-byte KEM key less often—say every 50th message or once every week—rather than every message. That idea was nixed because it doesn’t work well in asynchronous or adversarial messaging environments. Signal Protocol developers Graeme Connell and Rolfe Schmidt explained:

Consider the case of “send a key if you haven’t sent one in a week”. If Bob has been offline for 2 weeks, what does Alice do when she wants to send a message? What happens if we can lose messages, and we lose the one in fifty that contains a new key? Or, what happens if there’s an attacker in the middle that wants to stop us from generating new secrets, and can look for messages that are [many] bytes larger than the others and drop them, only allowing keyless messages through?

Another option Signal engineers considered was breaking the 2272-byte key into smaller chunks, say 71 of them that are 32 bytes each. Breaking up the KEM key into smaller chunks and putting one in each message sounds like a viable approach at first, but once again, the asynchronous environment of messaging made it unworkable. What happens, for example, when data loss causes one of the chunks to be dropped? The protocol could deal with this scenario by just repeat-sending chunks again after sending all 71 previously. But then an adversary monitoring the traffic could simply cause packet 3 to be dropped each time, preventing Alice and Bob from completing the key exchange.

Signal developers ultimately went with a solution that used this multiple-chunks approach.

Sneaking an elephant through the cat door

To manage the asynchrony challenges, the developers turned to “erasure codes,” a method of breaking up larger data into smaller pieces such that the original can be reconstructed using any sufficiently sized subset of chunks.

Charlie Jacomme, a researcher at INRIA Nancy on the Pesto team who focuses on formal verification and secure messaging, said this design accounts for packet loss by building redundancy into the chunked material. Instead of all x number of chunks having to be successfully received to reconstruct the key, the model requires only x-y chunks to be received, where y is the acceptable number of packets lost. As long as that threshold is met, the new key can be established even when packet loss occurs.

The other part of the design was to split the KEM computations into smaller steps. These KEM computations are distinct from the KEM key material.

As Jacomme explained it:

Essentially, a small part of the public key is enough to start computing and sending a bigger part of the ciphertext, so you can quickly send in parallel the rest of the public key and the beginning of the ciphertext. Essentially, the final computations are equal to the standard, but some stuff was parallelized.

All this in fact plays a role in the end security guarantees, because by optimizing the fact that KEM computations are done faster, you introduce in your key derivation fresh secrets more frequently.

Signal’s post 10 days ago included several images that illustrate this design:

While the design solved the asynchronous messaging problem, it created a new complication of its own: This new quantum-safe ratchet advanced so quickly that it couldn’t be kept synchronized with the Diffie-Hellman ratchet. Ultimately, the architects settled on a creative solution. Rather than bolt KEM onto the existing double ratchet, they allowed it to remain more or less the same as it had been. Then they used the new quantum-safe ratchet to implement a parallel secure messaging system.

Now, when the protocol encrypts a message, it sources encryption keys from both the classic Double Ratchet and the new ratchet. It then mixes the two keys together (using a cryptographic key derivation function) to get a new encryption key that has all of the security of the classical Double Ratchet but now has quantum security, too.

The Signal engineers have given this third ratchet the formal name: Sparse Post Quantum Ratchet, or SPQR for short. The third ratchet was designed in collaboration with PQShield, AIST, and New York University. The developers presented the erasure-code-based chunking and the high-level Triple Ratchet design at the Eurocrypt 2025 conference. At the Usenix 25 conference, they discussed the six options they considered for adding quantum-safe forward secrecy and post-compromise security and why SPQR and one other stood out. Presentations at the NIST PQC Standardization Conference and the Cryptographic Applications Workshop explain the details of chunking, the design challenges, and how the protocol had to be adapted to use the standardized ML-KEM.

Jacomme further observed:

The final thing interesting for the triple ratchet is that it nicely combines the best of both worlds. Between two users, you have a classical DH-based ratchet going on one side, and fully independently, a KEM-based ratchet is going on. Then, whenever you need to encrypt something, you get a key from both, and mix it up to get the actual encryption key. So, even if one ratchet is fully broken, be it because there is now a quantum computer, or because somebody manages to break either elliptic curves or ML-KEM, or because the implementation of one is flawed, or…, the Signal message will still be protected by the second ratchet. In a sense, this update can be seen, of course simplifying, as doubling the security of the ratchet part of Signal, and is a cool thing even for people that don’t care about quantum computers.

As both Signal and Jacomme noted, users of Signal and other messengers relying on the Signal Protocol need not concern themselves with any of these new designs. To paraphrase a certain device maker, it just works.

In the coming weeks or months, various messaging apps and app versions will be updated to add the triple ratchet. Until then, apps will simply rely on the double ratchet as they always did. Once apps receive the update, they’ll behave exactly as they did before upgrading.

For those who care about the internal workings of their Signal-based apps, though, the architects have documented in great depth the design of this new ratchet and how it behaves. Among other things, the work includes a mathematical proof verifying that the updated Signal protocol provides the claimed security properties.

Outside researchers are applauding the work.

“If the normal encrypted messages we use are cats, then post-quantum ciphertexts are elephants,” Matt Green, a cryptography expert at Johns Hopkins University, wrote in an interview. “So the problem here is to sneak an elephant through a tunnel designed for cats. And that’s an amazing engineering achievement. But it also makes me wish we didn’t have to deal with elephants.”

Photo of Dan Goodin

Dan Goodin is Senior Security Editor at Ars Technica, where he oversees coverage of malware, computer espionage, botnets, hardware hacking, encryption, and passwords. In his spare time, he enjoys gardening, cooking, and following the independent music scene. Dan is based in San Francisco. Follow him at here on Mastodon and here on Bluesky. Contact him on Signal at DanArs.82.

Why Signal’s post-quantum makeover is an amazing engineering achievement Read More »

“like-putting-on-glasses-for-the-first-time”—how-ai-improves-earthquake-detection

“Like putting on glasses for the first time”—how AI improves earthquake detection

AI, earthquakes, Features, Science, seismology / /

AI is “comically good” at detecting small earthquakes—here’s why that matters.

Credit: Aurich Lawson | Getty Images

On January 1, 2008, at 1: 59 am in Calipatria, California, an earthquake happened. You haven’t heard of this earthquake; even if you had been living in Calipatria, you wouldn’t have felt anything. It was magnitude -0.53, about the same amount of shaking as a truck passing by. Still, this earthquake is notable, not because it was large but because it was small—and yet we know about it.

Over the past seven years, AI tools based on computer imaging have almost completely automated one of the fundamental tasks of seismology: detecting earthquakes. What used to be the task of human analysts—and later, simpler computer programs—can now be done automatically and quickly by machine-learning tools.

These machine-learning tools can detect smaller earthquakes than human analysts, especially in noisy environments like cities. Earthquakes give valuable information about the composition of the Earth and what hazards might occur in the future.

“In the best-case scenario, when you adopt these new techniques, even on the same old data, it’s kind of like putting on glasses for the first time, and you can see the leaves on the trees,” said Kyle Bradley, co-author of the Earthquake Insights newsletter.

I talked with several earthquake scientists, and they all agreed that machine-learning methods have replaced humans for the better in these specific tasks.

“It’s really remarkable,” Judith Hubbard, a Cornell University professor and Bradley’s co-author, told me.

Less certain is what comes next. Earthquake detection is a fundamental part of seismology, but there are many other data processing tasks that have yet to be disrupted. The biggest potential impacts, all the way to earthquake forecasting, haven’t materialized yet.

“It really was a revolution,” said Joe Byrnes, a professor at the University of Texas at Dallas. “But the revolution is ongoing.”

When an earthquake happens in one place, the shaking passes through the ground, similar to how sound waves pass through the air. In both cases, it’s possible to draw inferences about the materials the waves pass through.

Imagine tapping a wall to figure out if it’s hollow. Because a solid wall vibrates differently than a hollow wall, you can figure out the structure by sound.

With earthquakes, this same principle holds. Seismic waves pass through different materials (rock, oil, magma, etc.) differently, and scientists use these vibrations to image the Earth’s interior.

The main tool that scientists traditionally use is a seismometer. These record the movement of the Earth in three directions: up–down, north–south, and east–west. If an earthquake happens, seismometers can measure the shaking in that particular location.

An old-fashioned physical seismometer. Today, seismometers record data digitally. Credit: Yamaguchi先生 on Wikimedia CC BY-SA 3.0

Scientists then process raw seismometer information to identify earthquakes.

Earthquakes produce multiple types of shaking, which travel at different speeds. Two types, Primary (P) waves and Secondary (S) waves are particularly important, and scientists like to identify the start of each of these phases.

Before good algorithms, earthquake cataloging had to happen by hand. Byrnes said that “traditionally, something like the lab at the United States Geological Survey would have an army of mostly undergraduate students or interns looking at seismograms.”

However, there are only so many earthquakes you can find and classify manually. Creating algorithms to effectively find and process earthquakes has long been a priority in the field—especially since the arrival of computers in the early 1950s.

“The field of seismology historically has always advanced as computing has advanced,” Bradley told me.

There’s a big challenge with traditional algorithms, though: They can’t easily find smaller quakes, especially in noisy environments.

Composite seismogram of common events. Note how each event has a slightly different shape. Credit: EarthScope Consortium CC BY 4.0

As we see in the seismogram above, many different events can cause seismic signals. If a method is too sensitive, it risks falsely detecting events as earthquakes. The problem is especially bad in cities, where the constant hum of traffic and buildings can drown out small earthquakes.

However, earthquakes have a characteristic “shape.” The magnitude 7.7 earthquake above looks quite different from the helicopter landing, for instance.

So one idea scientists had was to make templates from human-labeled datasets. If a new waveform correlates closely with an existing template, it’s almost certainly an earthquake.

Template matching works very well if you have enough human-labeled examples. In 2019, Zach Ross’ lab at Caltech used template matching to find 10 times as many earthquakes in Southern California as had previously been known, including the earthquake at the start of this story. Almost all of the new 1.6 million quakes they found were very small, magnitude 1 and below.

If you don’t have an extensive pre-existing dataset of templates, however, you can’t easily apply template matching. That isn’t a problem in Southern California—which already had a basically complete record of earthquakes down to magnitude 1.7—but it’s a challenge elsewhere.

Also, template matching is computationally expensive. Creating a Southern California quake dataset using template matching took 200 Nvidia P100 GPUs running for days on end.

There had to be a better way.

AI detection models solve all of these problems:

  • They are faster than template matching.

  • Because AI detection models are very small (around 350,000 parameters compared to billions in LLMs like GPT4.0), they can be run on consumer CPUs.

  • AI models generalize well to regions not represented in the original dataset.

As an added bonus, AI models can give better information about when the different types of earthquake shaking arrive. Timing the arrivals of the two most important waves—P and S waves—is called phase picking. It allows scientists to draw inferences about the structure of the quake. AI models can do this alongside earthquake detection.

The basic task of earthquake detection (and phase picking) looks like this:

Cropped figure from Earthquake Transformer—an attentive deep-learning model for simultaneous earthquake detection and phase picking. Credit: Nature Communications

The first three rows represent different directions of vibration (east–west, north–south, and up–down respectively). Given these three dimensions of vibration, can we determine if an earthquake occurred, and if so, when it started?

We want to detect the initial P wave, which arrives directly from the site of the earthquake. But this can be tricky because echoes of the P wave may get reflected off other rock layers and arrive later, making the waveform more complicated.

Ideally, then, our model outputs three things at every time step in the sample:

  1. The probability that an earthquake is occurring at that moment.

  2. The probability that the first P wave arrives at that moment.

  3. The probability that the first S wave arrives at that moment.

We see all three outputs in the fourth row: the detection in green, the P wave arrival in blue, and the S wave arrival in red. (There are two earthquakes in this sample.)

To train an AI model, scientists take large amounts of labeled data, like what’s above, and do supervised training. I’ll describe one of the most used models: Earthquake Transformer, which was developed around 2020 by a Stanford University team led by S. Mostafa Mousavi, who later became a Harvard professor.

Like many earthquake detection models, Earthquake Transformer adapts ideas from image classification. Readers may be familiar with AlexNet, a famous image-recognition model that kicked off the deep-learning boom in 2012.

AlexNet used convolutions, a neural network architecture that’s based on the idea that pixels that are physically close together are more likely to be related. The first convolutional layer of AlexNet broke an image down into small chunks—11 pixels on a side—and classified each chunk based on the presence of simple features like edges or gradients.

The next layer took the first layer’s classifications as input and checked for higher-level concepts such as textures or simple shapes.

Each convolutional layer analyzed a larger portion of the image and operated at a higher level of abstraction. By the final layers, the network was looking at the entire image and identifying objects like “mushroom” and “container ship.”

Images are two-dimensional, so AlexNet is based on two-dimensional convolutions. By contrast, seismograph data is one-dimensional, so Earthquake Transformer uses one-dimensional convolutions over the time dimension. The first layer analyzes vibration data in 0.1-second chunks, while later layers identify patterns over progressively longer time periods.

It’s difficult to say what exact patterns the earthquake model is picking out, but we can analogize this to a hypothetical audio transcription model using one-dimensional convolutions. That model might first identify consonants, then syllables, then words, then sentences over increasing time scales.

Earthquake Transformer converts raw waveform data into a collection of high-level representations that indicate the likelihood of earthquakes and other seismologically significant events. This is followed by a series of deconvolution layers that pinpoint exactly when an earthquake—and its all-important P and S waves—occurred.

The model also uses an attention layer in the middle of the model to mix information between different parts of the time series. The attention mechanism is most famous in large language models, where it helps pass information between words. It plays a similar role in seismographic detection. Earthquake seismograms have a general structure: P waves followed by S waves followed by other types of shaking. So if a segment looks like the start of a P wave, the attention mechanism helps it check that it fits into a broader earthquake pattern.

All of the Earthquake Transformer’s components are standard designs from the neural network literature. Other successful detection models, like PhaseNet, are even simpler. PhaseNet uses only one-dimensional convolutions to pick the arrival times of earthquake waves. There are no attention layers.

Generally, there hasn’t been “much need to invent new architectures for seismology,” according to Byrnes. The techniques derived from image processing have been sufficient.

What made these generic architectures work so well then? Data. Lots of it.

Ars has previously reported on how the introduction of ImageNet, an image recognition benchmark, helped spark the deep learning boom. Large, publicly available earthquake datasets have played a similar role in seismology.

Earthquake Transformer was trained using the Stanford Earthquake Dataset (STEAD), which contains 1.2 million human-labeled segments of seismogram data from around the world. (The paper for STEAD explicitly mentions ImageNet as an inspiration). Other models, like PhaseNet, were also trained on hundreds of thousands or millions of labeled segments.

All recorded earthquakes in the Stanford Earthquake Dataset. Credit: IEEE (CC BY 4.0)

The combination of the data and the architecture just works. The current models are “comically good” at identifying and classifying earthquakes, according to Byrnes. Typically, machine-learning methods find 10 or more times the quakes that were previously identified in an area. You can see this directly in an Italian earthquake catalog:

From Machine learning and earthquake forecasting—next steps by Beroza et al. Credit: Nature Communications (CC-BY 4.0)

AI tools won’t necessarily detect more earthquakes than template matching. But AI-based techniques are much less compute- and labor-intensive, making them more accessible to the average research project and easier to apply in regions around the world.

All in all, these machine-learning models are so good that they’ve almost completely supplanted traditional methods for detecting and phase-picking earthquakes, especially for smaller magnitudes.

The holy grail of earthquake science is earthquake prediction. For instance, scientists know that a large quake will happen near Seattle but have little ability to know whether it will happen tomorrow or in a hundred years. It would be helpful if we could predict earthquakes precisely enough to allow people in affected areas to evacuate.

You might think AI tools would help predict earthquakes, but that doesn’t seem to have happened yet.

The applications are more technical and less flashy, said Cornell’s Judith Hubbard.

Better AI models have given seismologists much more comprehensive earthquake catalogs, which have unlocked “a lot of different techniques,” Bradley said.

One of the coolest applications is in understanding and imaging volcanoes. Volcanic activity produces a large number of small earthquakes, whose locations help scientists understand the structure of the magma system. In a 2022 paper, John Wilding and co-authors used a large AI-generated earthquake catalog to create this incredible image of the structure of the Hawaiian volcanic system.

Each dot represents an individual earthquake. Credit: Wilding et al., The magmatic web beneath Hawai‘i.

They provided direct evidence of a previously hypothesized magma connection between the deep Pāhala sill complex and Mauna Loa’s shallow volcanic structure. You can see this in the image with the arrow labeled as Pāhala-Mauna Loa seismicity band. The authors were also able to clarify the structure of the Pāhala sill complex into discrete sheets of magma. This level of detail could potentially facilitate better real-time monitoring of earthquakes and more accurate eruption forecasting.

Another promising area is lowering the cost of dealing with huge datasets. Distributed Acoustic Sensing (DAS) is a powerful technique that uses fiber-optic cables to measure seismic activity across the entire length of the cable. A single DAS array can produce “hundreds of gigabytes of data” a day, according to Jiaxuan Li, a professor at the University of Houston. That much data can produce extremely high-resolution datasets—enough to pick out individual footsteps.

AI tools make it possible to very accurately time earthquakes in DAS data. Before the introduction of AI techniques for phase picking in DAS data, Li and some of his collaborators attempted to use traditional techniques. While these “work roughly,” they weren’t accurate enough for their downstream analysis. Without AI, much of his work would have been “much harder,” he told me.

Li is also optimistic that AI tools will be able to help him isolate “new types of signals” in the rich DAS data in the future.

Not all AI techniques have paid off

As in many other scientific fields, seismologists face some pressure to adopt AI methods, whether or not they are relevant to their research.

“The schools want you to put the word AI in front of everything,” Byrnes said. “It’s a little out of control.”

This can lead to papers that are technically sound but practically useless. Hubbard and Bradley told me that they’ve seen a lot of papers based on AI techniques that “reveal a fundamental misunderstanding of how earthquakes work.”

They pointed out that graduate students can feel pressure to specialize in AI methods at the cost of learning less about the fundamentals of the scientific field. They fear that if this type of AI-driven research becomes entrenched, older methods will get “out-competed by a kind of meaninglessness.”

While these are real issues, and ones Understanding AI has reported on before, I don’t think they detract from the success of AI earthquake detection. In the last five years, an AI-based workflow has almost completely replaced one of the fundamental tasks in seismology for the better.

That’s pretty cool.

Kai Williams is a reporter for Understanding AI, a Substack newsletter founded by Ars Technica alum Timothy B. Lee. His work is supported by a Tarbell Fellowship. Subscribe to Understanding AI to get more from Tim and Kai.

“Like putting on glasses for the first time”—how AI improves earthquake detection Read More »