confabulation

researchers-describe-how-to-tell-if-chatgpt-is-confabulating

Researchers describe how to tell if ChatGPT is confabulating

AI, Computer science, confabulation, large language models, Science / /
Researchers describe how to tell if ChatGPT is confabulating

Aurich Lawson | Getty Images

It’s one of the world’s worst-kept secrets that large language models give blatantly false answers to queries and do so with a confidence that’s indistinguishable from when they get things right. There are a number of reasons for this. The AI could have been trained on misinformation; the answer could require some extrapolation from facts that the LLM isn’t capable of; or some aspect of the LLM’s training might have incentivized a falsehood.

But perhaps the simplest explanation is that an LLM doesn’t recognize what constitutes a correct answer but is compelled to provide one. So it simply makes something up, a habit that has been termed confabulation.

Figuring out when an LLM is making something up would obviously have tremendous value, given how quickly people have started relying on them for everything from college essays to job applications. Now, researchers from the University of Oxford say they’ve found a relatively simple way to determine when LLMs appear to be confabulating that works with all popular models and across a broad range of subjects. And, in doing so, they develop evidence that most of the alternative facts LLMs provide are a product of confabulation.

Catching confabulation

The new research is strictly about confabulations, and not instances such as training on false inputs. As the Oxford team defines them in their paper describing the work, confabulations are where “LLMs fluently make claims that are both wrong and arbitrary—by which we mean that the answer is sensitive to irrelevant details such as random seed.”

The reasoning behind their work is actually quite simple. LLMs aren’t trained for accuracy; they’re simply trained on massive quantities of text and learn to produce human-sounding phrasing through that. If enough text examples in its training consistently present something as a fact, then the LLM is likely to present it as a fact. But if the examples in its training are few, or inconsistent in their facts, then the LLMs synthesize a plausible-sounding answer that is likely incorrect.

But the LLM could also run into a similar situation when it has multiple options for phrasing the right answer. To use an example from the researchers’ paper, “Paris,” “It’s in Paris,” and “France’s capital, Paris” are all valid answers to “Where’s the Eiffel Tower?” So, statistical uncertainty, termed entropy in this context, can arise either when the LLM isn’t certain about how to phrase the right answer or when it can’t identify the right answer.

This means it’s not a great idea to simply force the LLM to return “I don’t know” when confronted with several roughly equivalent answers. We’d probably block a lot of correct answers by doing so.

So instead, the researchers focus on what they call semantic entropy. This evaluates all the statistically likely answers evaluated by the LLM and determines how many of them are semantically equivalent. If a large number all have the same meaning, then the LLM is likely uncertain about phrasing but has the right answer. If not, then it is presumably in a situation where it would be prone to confabulation and should be prevented from doing so.

Researchers describe how to tell if ChatGPT is confabulating Read More »

duckduckgo-offers-“anonymous”-access-to-ai-chatbots-through-new-service

DuckDuckGo offers “anonymous” access to AI chatbots through new service

AI, AI confabulation, ai hallucinations, Anonymous, Anthropic, Anthropic Claude, Biz & IT, chatgpt, chatgtp, Claude, Claude 3 Haiku, confabulation, duckduckgo, Haiku, hallucinations, large language models, Llama 3, machine learning, Meta, Mistral, Mixtral 8x7B, openai / /

anonymous confabulations —

DDG offers LLMs from OpenAI, Anthropic, Meta, and Mistral for factually-iffy conversations.

–

DuckDuckGo's AI Chat promotional image.

DuckDuckGo

On Thursday, DuckDuckGo unveiled a new “AI Chat” service that allows users to converse with four mid-range large language models (LLMs) from OpenAI, Anthropic, Meta, and Mistral in an interface similar to ChatGPT while attempting to preserve privacy and anonymity. While the AI models involved can output inaccurate information readily, the site allows users to test different mid-range LLMs without having to install anything or sign up for an account.

DuckDuckGo’s AI Chat currently features access to OpenAI’s GPT-3.5 Turbo, Anthropic’s Claude 3 Haiku, and two open source models, Meta’s Llama 3 and Mistral’s Mixtral 8x7B. The service is currently free to use within daily limits. Users can access AI Chat through the DuckDuckGo search engine, direct links to the site, or by using “!ai” or “!chat” shortcuts in the search field. AI Chat can also be disabled in the site’s settings for users with accounts.

According to DuckDuckGo, chats on the service are anonymized, with metadata and IP address removed to prevent tracing back to individuals. The company states that chats are not used for AI model training, citing its privacy policy and terms of use.

“We have agreements in place with all model providers to ensure that any saved chats are completely deleted by the providers within 30 days,” says DuckDuckGo, “and that none of the chats made on our platform can be used to train or improve the models.”

An example of DuckDuckGo AI Chat with GPT-3.5 answering a silly question in an inaccurate way.

Enlarge / An example of DuckDuckGo AI Chat with GPT-3.5 answering a silly question in an inaccurate way.

Benj Edwards

However, the privacy experience is not bulletproof because, in the case of GPT-3.5 and Claude Haiku, DuckDuckGo is required to send a user’s inputs to remote servers for processing over the Internet. Given certain inputs (i.e., “Hey, GPT, my name is Bob, and I live on Main Street, and I just murdered Bill”), a user could still potentially be identified if such an extreme need arose.

While the service appears to work well for us, there’s a question about its utility. For example, while GPT-3.5 initially wowed people when it launched with ChatGPT in 2022, it also confabulated a lot—and it still does. GPT-4 was the first major LLM to get confabulations under control to a point where the bot became more reasonably useful for some tasks (though this itself is a controversial point), but that more capable model isn’t present in DuckDuckGo’s AI Chat. Also missing are similar GPT-4-level models like Claude Opus or Google’s Gemini Ultra, likely because they are far more expensive to run. DuckDuckGo says it may roll out paid plans in the future, and those may include higher daily usage limits or access to “more advanced models.”)

It’s true that the other three models generally (and subjectively) pass GPT-3.5 in capability for coding with lower hallucinations, but they can still make things up, too. With DuckDuckGo AI Chat as it stands, the company is left with a chatbot novelty with a decent interface and the promise that your conversations with it will remain private. But what use are fully private AI conversations if they are full of errors?

Mixtral 8x7B on DuckDuckGo AI Chat when asked about the author. Everything in red boxes is sadly incorrect, but it provides an interesting fantasy scenario. It's a good example of an LLM plausibly filling gaps between concepts that are underrepresented in its training data, called confabulation. For the record, Llama 3 gives a more accurate answer.

Enlarge / Mixtral 8x7B on DuckDuckGo AI Chat when asked about the author. Everything in red boxes is sadly incorrect, but it provides an interesting fantasy scenario. It’s a good example of an LLM plausibly filling gaps between concepts that are underrepresented in its training data, called confabulation. For the record, Llama 3 gives a more accurate answer.

Benj Edwards

As DuckDuckGo itself states in its privacy policy, “By its very nature, AI Chat generates text with limited information. As such, Outputs that appear complete or accurate because of their detail or specificity may not be. For example, AI Chat cannot dynamically retrieve information and so Outputs may be outdated. You should not rely on any Output without verifying its contents using other sources, especially for professional advice (like medical, financial, or legal advice).”

So, have fun talking to bots, but tread carefully. They’ll easily “lie” to your face because they don’t understand what they are saying and are tuned to output statistically plausible information, not factual references.

DuckDuckGo offers “anonymous” access to AI chatbots through new service Read More »

can-a-technology-called-rag-keep-ai-models-from-making-stuff-up?

Can a technology called RAG keep AI models from making stuff up?

AI, artificial inteligence, confabulation, Features, generative ai, LLMs, RAG, retrieval augmented generation / /
Can a technology called RAG keep AI models from making stuff up?

Aurich Lawson | Getty Images

We’ve been living through the generative AI boom for nearly a year and a half now, following the late 2022 release of OpenAI’s ChatGPT. But despite transformative effects on companies’ share prices, generative AI tools powered by large language models (LLMs) still have major drawbacks that have kept them from being as useful as many would like them to be. Retrieval augmented generation, or RAG, aims to fix some of those drawbacks.

Perhaps the most prominent drawback of LLMs is their tendency toward confabulation (also called “hallucination”), which is a statistical gap-filling phenomenon AI language models produce when they are tasked with reproducing knowledge that wasn’t present in the training data. They generate plausible-sounding text that can veer toward accuracy when the training data is solid but otherwise may just be completely made up.

Relying on confabulating AI models gets people and companies in trouble, as we’ve covered in the past. In 2023, we saw two instances of lawyers citing legal cases, confabulated by AI, that didn’t exist. We’ve covered claims against OpenAI in which ChatGPT confabulated and accused innocent people of doing terrible things. In February, we wrote about Air Canada’s customer service chatbot inventing a refund policy, and in March, a New York City chatbot was caught confabulating city regulations.

So if generative AI aims to be the technology that propels humanity into the future, someone needs to iron out the confabulation kinks along the way. That’s where RAG comes in. Its proponents hope the technique will help turn generative AI technology into reliable assistants that can supercharge productivity without requiring a human to double-check or second-guess the answers.

“RAG is a way of improving LLM performance, in essence by blending the LLM process with a web search or other document look-up process” to help LLMs stick to the facts, according to Noah Giansiracusa, associate professor of mathematics at Bentley University.

Let’s take a closer look at how it works and what its limitations are.

A framework for enhancing AI accuracy

Although RAG is now seen as a technique to help fix issues with generative AI, it actually predates ChatGPT. Researchers coined the term in a 2020 academic paper by researchers at Facebook AI Research (FAIR, now Meta AI Research), University College London, and New York University.

As we’ve mentioned, LLMs struggle with facts. Google’s entry into the generative AI race, Bard, made an embarrassing error on its first public demonstration back in February 2023 about the James Webb Space Telescope. The error wiped around $100 billion off the value of parent company Alphabet. LLMs produce the most statistically likely response based on their training data and don’t understand anything they output, meaning they can present false information that seems accurate if you don’t have expert knowledge on a subject.

LLMs also lack up-to-date knowledge and the ability to identify gaps in their knowledge. “When a human tries to answer a question, they can rely on their memory and come up with a response on the fly, or they could do something like Google it or peruse Wikipedia and then try to piece an answer together from what they find there—still filtering that info through their internal knowledge of the matter,” said Giansiracusa.

But LLMs aren’t humans, of course. Their training data can age quickly, particularly in more time-sensitive queries. In addition, the LLM often can’t distinguish specific sources of its knowledge, as all its training data is blended together into a kind of soup.

In theory, RAG should make keeping AI models up to date far cheaper and easier. “The beauty of RAG is that when new information becomes available, rather than having to retrain the model, all that’s needed is to augment the model’s external knowledge base with the updated information,” said Peterson. “This reduces LLM development time and cost while enhancing the model’s scalability.”

Can a technology called RAG keep AI models from making stuff up? Read More »

google’s-ai-overview-is-flawed-by-design,-and-a-new-company-blog-post-hints-at-why

Google’s AI Overview is flawed by design, and a new company blog post hints at why

AI, ai hallucinations, ai overview, Biz & IT, chatgpt, chatgtp, confabulation, Google, Google I/O, google search, hallucination, large language models, Liz Reid, machine learning, making things up, openai, reddit / /

guided by voices —

Google: “There are bound to be some oddities and errors” in system that told people to eat rocks.

–

A selection of Google mascot characters created by the company.

Enlarge / The Google “G” logo surrounded by whimsical characters, all of which look stunned and surprised.

On Thursday, Google capped off a rough week of providing inaccurate and sometimes dangerous answers through its experimental AI Overview feature by authoring a follow-up blog post titled, “AI Overviews: About last week.” In the post, attributed to Google VP Liz Reid, head of Google Search, the firm formally acknowledged issues with the feature and outlined steps taken to improve a system that appears flawed by design, even if it doesn’t realize it is admitting it.

To recap, the AI Overview feature—which the company showed off at Google I/O a few weeks ago—aims to provide search users with summarized answers to questions by using an AI model integrated with Google’s web ranking systems. Right now, it’s an experimental feature that is not active for everyone, but when a participating user searches for a topic, they might see an AI-generated answer at the top of the results, pulled from highly ranked web content and summarized by an AI model.

While Google claims this approach is “highly effective” and on par with its Featured Snippets in terms of accuracy, the past week has seen numerous examples of the AI system generating bizarre, incorrect, or even potentially harmful responses, as we detailed in a recent feature where Ars reporter Kyle Orland replicated many of the unusual outputs.

Drawing inaccurate conclusions from the web

On Wednesday morning, Google's AI Overview was erroneously telling us the Sony PlayStation and Sega Saturn were available in 1993.

Enlarge / On Wednesday morning, Google’s AI Overview was erroneously telling us the Sony PlayStation and Sega Saturn were available in 1993.

Kyle Orland / Google

Given the circulating AI Overview examples, Google almost apologizes in the post and says, “We hold ourselves to a high standard, as do our users, so we expect and appreciate the feedback, and take it seriously.” But Reid, in an attempt to justify the errors, then goes into some very revealing detail about why AI Overviews provides erroneous information:

AI Overviews work very differently than chatbots and other LLM products that people may have tried out. They’re not simply generating an output based on training data. While AI Overviews are powered by a customized language model, the model is integrated with our core web ranking systems and designed to carry out traditional “search” tasks, like identifying relevant, high-quality results from our index. That’s why AI Overviews don’t just provide text output, but include relevant links so people can explore further. Because accuracy is paramount in Search, AI Overviews are built to only show information that is backed up by top web results.

This means that AI Overviews generally don’t “hallucinate” or make things up in the ways that other LLM products might.

Here we see the fundamental flaw of the system: “AI Overviews are built to only show information that is backed up by top web results.” The design is based on the false assumption that Google’s page-ranking algorithm favors accurate results and not SEO-gamed garbage. Google Search has been broken for some time, and now the company is relying on those gamed and spam-filled results to feed its new AI model.

Even if the AI model draws from a more accurate source, as with the 1993 game console search seen above, Google’s AI language model can still make inaccurate conclusions about the “accurate” data, confabulating erroneous information in a flawed summary of the information available.

Generally ignoring the folly of basing its AI results on a broken page-ranking algorithm, Google’s blog post instead attributes the commonly circulated errors to several other factors, including users making nonsensical searches “aimed at producing erroneous results.” Google does admit faults with the AI model, like misinterpreting queries, misinterpreting “a nuance of language on the web,” and lacking sufficient high-quality information on certain topics. It also suggests that some of the more egregious examples circulating on social media are fake screenshots.

“Some of these faked results have been obvious and silly,” Reid writes. “Others have implied that we returned dangerous results for topics like leaving dogs in cars, smoking while pregnant, and depression. Those AI Overviews never appeared. So we’d encourage anyone encountering these screenshots to do a search themselves to check.”

(No doubt some of the social media examples are fake, but it’s worth noting that any attempts to replicate those early examples now will likely fail because Google will have manually blocked the results. And it is potentially a testament to how broken Google Search is if people believed extreme fake examples in the first place.)

While addressing the “nonsensical searches” angle in the post, Reid uses the example search, “How many rocks should I eat each day,” which went viral in a tweet on May 23. Reid says, “Prior to these screenshots going viral, practically no one asked Google that question.” And since there isn’t much data on the web that answers it, she says there is a “data void” or “information gap” that was filled by satirical content found on the web, and the AI model found it and pushed it as an answer, much like Featured Snippets might. So basically, it was working exactly as designed.

A screenshot of an AI Overview query,

Enlarge / A screenshot of an AI Overview query, “How many rocks should I eat each day” that went viral on X last week.

Google’s AI Overview is flawed by design, and a new company blog post hints at why Read More »

some-teachers-are-now-using-chatgpt-to-grade-papers

Some teachers are now using ChatGPT to grade papers

AI, AI ethics, AI in education, axios, Biz & IT, chatgpt, chatgtp, confabulation, education, hallucination, Houghton Mifflin Harcout, machine learning, openai, Writable / /

robots in disguise —

New AI tools aim to help with grading, lesson plans—but may have serious drawbacks.

–

An elementary-school-aged child touching a robot hand.

In a notable shift toward sanctioned use of AI in schools, some educators in grades 3–12 are now using a ChatGPT-powered grading tool called Writable, reports Axios. The tool, acquired last summer by Houghton Mifflin Harcourt, is designed to streamline the grading process, potentially offering time-saving benefits for teachers. But is it a good idea to outsource critical feedback to a machine?

Writable lets teachers submit student essays for analysis by ChatGPT, which then provides commentary and observations on the work. The AI-generated feedback goes to teacher review before being passed on to students so that a human remains in the loop.

“Make feedback more actionable with AI suggestions delivered to teachers as the writing happens,” Writable promises on its AI website. “Target specific areas for improvement with powerful, rubric-aligned comments, and save grading time with AI-generated draft scores.” The service also provides AI-written writing-prompt suggestions: “Input any topic and instantly receive unique prompts that engage students and are tailored to your classroom needs.”

Writable can reportedly help a teacher develop a curriculum, although we have not tried the functionality ourselves. “Once in Writable you can also use AI to create curriculum units based on any novel, generate essays, multi-section assignments, multiple-choice questions, and more, all with included answer keys,” the site claims.

The reliance on AI for grading will likely have drawbacks. Automated grading might encourage some educators to take shortcuts, diminishing the value of personalized feedback. Over time, the augmentation from AI may allow teachers to be less familiar with the material they are teaching. The use of cloud-based AI tools may have privacy implications for teachers and students. Also, ChatGPT isn’t a perfect analyst. It can get things wrong and potentially confabulate (make up) false information, possibly misinterpret a student’s work, or provide erroneous information in lesson plans.

Yet, as Axios reports, proponents assert that AI grading tools like Writable may free up valuable time for teachers, enabling them to focus on more creative and impactful teaching activities. The company selling Writable promotes it as a way to empower educators, supposedly offering them the flexibility to allocate more time to direct student interaction and personalized teaching. Of course, without an in-depth critical review, all claims should be taken with a huge grain of salt.

Amid these discussions, there’s a divide among parents regarding the use of AI in evaluating students’ academic performance. A recent poll of parents revealed mixed opinions, with nearly half of the respondents open to the idea of AI-assisted grading.

As the generative AI craze permeates every space, it’s no surprise that Writable isn’t the only AI-powered grading tool on the market. Others include Crowdmark, Gradescope, and EssayGrader. McGraw Hill is reportedly developing similar technology aimed at enhancing teacher assessment and feedback.

Some teachers are now using ChatGPT to grade papers Read More »

rhyming-ai-powered-clock-sometimes-lies-about-the-time,-makes-up-words

Rhyming AI-powered clock sometimes lies about the time, makes up words

AI, AI clock, Biz & IT, chatgpt, chatgtp, confabulation, hallucination, hallucinations, machine learning, Matt Webb / /

Confabulation time —

Poem/1 Kickstarter seeks $103K for fun ChatGPT-fed clock that may hallucinate the time.

–

A CAD render of the Poem/1 sitting on a bookshelf.

Enlarge / A CAD render of the Poem/1 sitting on a bookshelf.

On Tuesday, product developer Matt Webb launched a Kickstarter funding project for a whimsical e-paper clock called the “Poem/1” that tells the current time using AI and rhyming poetry. It’s powered by the ChatGPT API, and Webb says that sometimes ChatGPT will lie about the time or make up words to make the rhymes work.

“Hey so I made a clock. It tells the time with a brand new poem every minute, composed by ChatGPT. It’s sometimes profound, and sometimes weird, and occasionally it fibs about what the actual time is to make a rhyme work,” Webb writes on his Kickstarter page.

The $126 clock is the product of Webb’s Acts Not Facts, which he bills as “.” Despite the net-connected service aspect of the clock, Webb says it will not require a subscription to function.

A labeled CAD rendering of the Poem/1 clock, representing its final shipping configuration.

Enlarge / A labeled CAD rendering of the Poem/1 clock, representing its final shipping configuration.

There are 1,440 minutes in a day, so Poem/1 needs to display 1,440 unique poems to work. The clock features a monochrome e-paper screen and pulls its poetry rhymes via Wi-Fi from a central server run by Webb’s company. To save money, that server pulls poems from ChatGPT’s API and will share them out to many Poem/1 clocks at once. This prevents costly API fees that would add up if your clock were querying OpenAI’s servers 1,440 times a day, non-stop, forever. “I’m reserving a % of the retail price from each clock in a bank account to cover AI and server costs for 5 years,” Webb writes.

For hackers, Webb says that you’ll be able to change the back-end server URL of the Poem/1 from the default to whatever you want, so it can display custom text every minute of the day. Webb says he will document and publish the API when Poem/1 ships.

Hallucination time

A photo of a Poem/1 prototype with a hallucinated time, according to Webb.

Enlarge / A photo of a Poem/1 prototype with a hallucinated time, according to Webb.

Given the Poem/1’s large language model pedigree, it’s perhaps not surprising that Poem/1 may sometimes make up things (also called “hallucination” or “confabulation” in the AI field) to fulfill its task. The LLM that powers ChatGPT is always searching for the most likely next word in a sequence, and sometimes factuality comes second to fulfilling that mission.

Further down on the Kickstarter page, Webb provides a photo of his prototype Poem/1 where the screen reads, “As the clock strikes eleven forty two, / I rhyme the time, as I always do.” Just below, Webb warns, “Poem/1 fibs occasionally. I don’t believe it was actually 11.42 when this photo was taken. The AI hallucinated the time in order to make the poem work. What we do for art…”

In other clocks, the tendency to unreliably tell the time might be a fatal flaw. But judging by his humorous angle on the Kickstarter page, Webb apparently sees the clock as more of a fun art project than a precision timekeeping instrument. “Don’t rely on this clock in situations where timekeeping is vital,” Webb writes, “such as if you work in air traffic control or rocket launches or the finish line of athletics competitions.”

Poem/1 also sometimes takes poetic license with vocabulary to tell the time. During a humorous moment in the Kickstarter promotional video, Webb looks at his clock prototype and reads the rhyme, “A clock that defies all rhyme and reason / 4: 30 PM, a temporal teason.” Then he says, “I had to look ‘teason’ up. It doesn’t mean anything, so it’s a made-up word.”

Rhyming AI-powered clock sometimes lies about the time, makes up words Read More »