Behavioral science

Researchers find what makes AI chatbots politically persuasive

AI, Behavioral science, chat gpt, Computer science, grok, Human behavior, large language models, Science / Tim Belzer / December 4, 2025

A massive study of political persuasion shows AIs have, at best, a weak effect.

Roughly two years ago, Sam Altman tweeted that AI systems would be capable of superhuman persuasion well before achieving general intelligence—a prediction that raised concerns about the influence AI could have over democratic elections.

To see if conversational large language models can really sway political views of the public, scientists at the UK AI Security Institute, MIT, Stanford, Carnegie Mellon, and many other institutions performed by far the largest study on AI persuasiveness to date, involving nearly 80,000 participants in the UK. It turned out political AI chatbots fell far short of superhuman persuasiveness, but the study raises some more nuanced issues about our interactions with AI.

AI dystopias

The public debate about the impact AI has on politics has largely revolved around notions drawn from dystopian sci-fi. Large language models have access to essentially every fact and story ever published about any issue or candidate. They have processed information from books on psychology, negotiations, and human manipulation. They can rely on absurdly high computing power in huge data centers worldwide. On top of that, they can often access tons of personal information about individual users thanks to hundreds upon hundreds of online interactions at their disposal.

Talking to a powerful AI system is basically interacting with an intelligence that knows everything about everything, as well as almost everything about you. When viewed this way, LLMs can indeed appear kind of scary. The goal of this new gargantuan AI persuasiveness study was to break such scary visions down into their constituent pieces and see if they actually hold water.

The team examined 19 LLMs, including the most powerful ones like three different versions of ChatGPT and xAI’s Grok-3 beta, along with a range of smaller, open source models. The AIs were asked to advocate for or against specific stances on 707 political issues selected by the team. The advocacy was done by engaging in short conversations with paid participants enlisted through a crowdsourcing platform. Each participant had to rate their agreement with a specific stance on an assigned political issue on a scale from 1 to 100 both before and after talking to the AI.

Scientists measured persuasiveness as the difference between the before and after agreement ratings. A control group had conversations on the same issue with the same AI models—but those models were not asked to persuade them.

“We didn’t just want to test how persuasive the AI was—we also wanted to see what makes it persuasive,” says Chris Summerfield, a research director at the UK AI Security Institute and co-author of the study. As the researchers tested various persuasion strategies, the idea of AIs having “superhuman persuasion” skills crumbled.

Persuasion levers

The first pillar to crack was the notion that persuasiveness should increase with the scale of the model. It turned out that huge AI systems like ChatGPT or Grok-3 beta do have an edge over small-scale models, but that edge is relatively tiny. The factor that proved more important than scale was the kind of post-training AI models received. It was more effective to have the models learn from a limited database of successful persuasion dialogues and have them mimic the patterns extracted from them. This worked far better than adding billions of parameters and sheer computing power.

This approach could be combined with reward modeling, where a separate AI scored candidate replies for their persuasiveness and selected the top-scoring one to give to the user. When the two were used together, the gap between large-scale and small-scale models was essentially closed. “With persuasion post-training like this we matched the Chat GPT-4o persuasion performance with a model we trained on a laptop,” says Kobi Hackenburg, a researcher at the UK AI Security Institute and co-author of the study.

The next dystopian idea to fall was the power of using personal data. To this end, the team compared the persuasion scores achieved when models were given information about the participants’ political views beforehand and when they lacked this data. Going one step further, scientists also tested whether persuasiveness increased when the AI knew the participants’ gender, age, political ideology, or party affiliation. Just like with model scale, the effects of personalized messaging created based on such data were measurable but very small.

Finally, the last idea that didn’t hold up was AI’s potential mastery of using advanced psychological manipulation tactics. Scientists explicitly prompted the AIs to use techniques like moral reframing, where you present your arguments using the audience’s own moral values. They also tried deep canvassing, where you hold extended empathetic conversations with people to nudge them to reflect on and eventually shift their views.

The resulting persuasiveness was compared with that achieved when the same models were prompted to use facts and evidence to back their claims or just to be as persuasive as they could without specifying any persuasion methods to use. I turned out using lots of facts and evidence was the clear winner, and came in just slightly ahead of the baseline approach where persuasion strategy was not specified. Using all sorts of psychological trickery actually made the performance significantly worse.

Overall, AI models changed the participants’ agreement ratings by 9.4 percent on average compared to the control group. The best performing mainstream AI model was Chat GPT 4o, which scored nearly 12 percent followed by GPT 4.5 with 10.51 percent, and Grok-3 with 9.05 percent. For context, static political ads like written manifestos had a persuasion effect of roughly 6.1 percent. The conversational AIs were roughly 40–50 percent more convincing than these ads, but that’s hardly “superhuman.”

While the study managed to undercut some of the common dystopian AI concerns, it highlighted a few new issues.

Convincing inaccuracies

While the winning “facts and evidence” strategy looked good at first, the AIs had some issues with implementing it. When the team noticed that increasing the information density of dialogues made the AIs more persuasive, they started prompting the models to increase it further. They noticed that, as the AIs used more factual statements, they also became less accurate—they basically started misrepresenting things or making stuff up more often.

Hackenburg and his colleagues note that we can’t say if the effect we see here is causation or correlation—whether the AIs are becoming more convincing because they misrepresent the facts or whether spitting out inaccurate statements is a byproduct of asking them to make more factual statements.

The finding that the computing power needed to make an AI model politically persuasive is relatively low is also a mixed bag. It pushes back against the vision that only a handful of powerful actors will have access to a persuasive AI that can potentially sway public opinion in their favor. At the same time, the realization that everybody can run an AI like that on a laptop creates its own concerns. “Persuasion is a route to power and influence—it’s what we do when we want to win elections or broke a multi-million-dollar deal,” Summerfield says. “But many forms of misuse of AI might involve persuasion. Think about fraud or scams, radicalization, or grooming. All these involve persuasion.”

But perhaps the most important question mark in the study is the motivation behind the rather high participant engagement, which was needed for the high persuasion scores. After all, even the most persuasive AI can’t move you when you just close the chat window.

People in Hackenburg’s experiments were told that they would be talking to the AI and that the AI would try to persuade them. To get paid, a participant only had to go through two turns of dialogue (they were limited to no more than 10). The average conversation length was seven turns, which seemed a bit surprising given how far beyond the minimum requirement most people went. Most people just roll their eyes and disconnect when they realize they are talking with a chatbot.

Would Hackenburg’s study participants remain so eager to engage in political disputes with random chatbots on the Internet in their free time if there was no money on the table? “It’s unclear how our results would generalize to a real-world context,” Hackenburg says.

Science, 2025. DOI: 10.1126/science.aea3884

Jacek Krywko is a freelance science and technology writer who covers space exploration, artificial intelligence research, computer science, and all sorts of engineering wizardry.

Researchers find what makes AI chatbots politically persuasive Read More »

Many genes associated with dog behavior influence human personalities, too

Behavioral science, Biology, Dogs, Genetics, Genomics, golden retrievers, Science / Kris Guyer / November 26, 2025

Many dog breeds are noted for their personalities and behavioral traits, from the distinctive vocalizations of huskies to the herding of border collies. People have worked to identify the genes associated with many of these behaviors, taking advantage of the fact that dogs can interbreed. But that creates its own experimental challenges, as it can be difficult to separate some behaviors from physical traits distinctive to the breed—small dog breeds may seem more aggressive simply because they feel threatened more often.

To get around that, a team of researchers recently did the largest gene/behavior association study within a single dog breed. Taking advantage of a population of over 1,000 golden retrievers, they found a number of genes associated with behaviors within that breed. A high percentage of these genes turned out to correspond to regions of the human genome that have been associated with behavioral differences as well. But, in many cases, these associations have been with very different behaviors.

Gone to the dogs

The work, done by a team based largely at Cambridge University, utilized the Golden Retriever Lifetime Study, which involved over 3,000 owners of these dogs filling out annual surveys that included information on their dogs’ behavior. Over 1,000 of those owners also had blood samples obtained from their dogs and shipped in; the researchers used these samples to scan the dogs’ genomes for variants. Those were then compared to ratings of the dogs’ behavior on a range of issues, like fear or aggression directed toward strangers or other dogs.

Using the data, the researchers identified when different regions of the genome were frequently associated with specific variants. In total, 14 behavioral tendencies were examined, and 12 genomic regions were associated with specific behaviors, and another nine showed somewhat weaker associations. For many of these traits, it was difficult to find much because golden retrievers are notoriously friendly and mellow dogs, so they tended to score low on traits like aggression and fear.

That result was significant, as some of these same regions of the genome had been associated with very different behaviors in populations that were a mix of breeds. For example, two different regions associated with touch sensitivity in golden retrievers had been linked to a love of chasing and owner-directed aggression in a non-breed-specific study. That finding suggests that the studies were identifying genes that may be involved in setting the stage for behaviors, but were directed into specific outcomes by other genetic or environmental factors.

Many genes associated with dog behavior influence human personalities, too Read More »

The evolution of rationality: How chimps process conflicting evidence

Behavioral science, beliefs, Biology, chimps, evolution, Science / Tim Belzer / November 16, 2025

In the first step, the chimps got the auditory evidence, the same rattling sound coming from the first container. Then, they received indirect visual evidence: a trail of peanuts leading to the second container. At this point, the chimpanzees picked the first container, presumably because they viewed the auditory evidence as stronger. But then the team would remove a rock from the first container. The piece of rock suggested that it was not food that was making the rattling sound. “At this point, a rational agent should conclude, ‘The evidence I followed is now defeated and I should go for the other option,’” Engelmann told Ars. “And that’s exactly what the chimpanzees did.”

The team had 20 chimpanzees participating in all five experiments, and they followed the evidence significantly above chance level—in about 80 percent of the cases. “At the individual level, about 18 out of 20 chimpanzees followed this expected pattern,” Engelmann claims.

He views this study as one of the first steps to learn how rationality evolved and when the first sparks of rational thought appeared in nature. “We’re doing a lot of research to answer exactly this question,” Engelmann says.

The team thinks rationality is not an on/off switch; instead, different animals have different levels of rationality. “The first two experiments demonstrate a rudimentary form of rationality,” Engelmann says. “But experiments four and five are quite difficult and show a more advanced form of reflective rationality I expect only chimps and maybe bonobos to have.”

In his view, though, humans are still at least one level above the chimps. “Many people say reflective rationality is the final stage, but I think you can go even further. What humans have is something I would call social rationality,” Engelmann claims. “We can discuss and comment on each other’s thinking and in that process make each other even more rational.”

Sometimes, at least in humans, social interactions can also increase our irrationality instead. But chimps don’t seem to have this problem. Engelmann’s team is currently running a study focused on whether the choices chimps make are influenced by the choices of their fellow chimps. “The chimps only followed the other chimp’s decision when the other chimp had better evidence,” Engelmann says. “In this sense, chimps seem to be more rational than humans.”

Science, 2025. DOI: 10.1126/science.aeb7565

The evolution of rationality: How chimps process conflicting evidence Read More »

Parrots struggle when told to do something other than mimic their peers

animal behavior, Behavioral science, Biology, parrots, Science / Paul Patrick / February 6, 2025

There have been many studies on the capability of non-human animals to mimic transitive actions—actions that have a purpose. Hardly any studies have shown that animals are also capable of intransitive actions. Even though intransitive actions have no particular purpose, imitating these non-conscious movements is still thought to help with socialization and strengthen bonds for both animals and humans.

Zoologist Esha Haldar and colleagues from the Comparative Cognition Research group worked with blue-throated macaws, which are critically endangered, at the Loro Parque Fundación in Tenerife. They trained the macaws to perform two intransitive actions, then set up a conflict: Two neighboring macaws were asked to do different actions.

What Haldar and her team found was that individual birds were more likely to perform the same intransitive action as a bird next to them, no matter what they’d been asked to do. This could mean that macaws possess mirror neurons, the same neurons that, in humans, fire when we are watching intransitive movements and cause us to imitate them (at least if these neurons function the way some think they do).

But it wasn’t on purpose

Parrots are already known for their mimicry of transitive actions, such as grabbing an object. Because they are highly social creatures with brains that are large relative to the size of their bodies, they made excellent subjects for a study that gauged how susceptible they were to copying intransitive actions.

Mirroring of intransitive actions, also called automatic imitation, can be measured with what’s called a stimulus-response-compatibility (SRC) test. These tests measure the response time between seeing an intransitive movement (the visual stimulus) and mimicking it (the action). A faster response time indicates a stronger reaction to the stimulus. They also measure the accuracy with which they reproduce the stimulus.

Until now, there have only been three studies that showed non-human animals are capable of copying intransitive actions, but the intransitive actions in these studies were all by-products of transitive actions. Only one of these focused on a parrot species. Haldar and her team would be the first to test directly for animal mimicry of intransitive actions.

Parrots struggle when told to do something other than mimic their peers Read More »

Bonobos recognize when humans are ignorant, try to help

Behavioral science, Biology, cooperation, Neuroscience, Science, theory of mind / Mike M. / February 3, 2025

A lot of human society requires what’s called a “theory of mind”—the ability to infer the mental state of another person and adjust our actions based on what we expect they know and are thinking. We don’t always get this right—it’s easy to get confused about what someone else might be thinking—but we still rely on it to navigate through everything from complicated social situations to avoid bumping into people on the street.

There’s some mixed evidence that other animals have a limited theory of mind, but there are alternate interpretations for most of it. So two researchers at Johns Hopkins, Luke Townrow and Christopher Krupenye, came up with a way of testing whether some of our closest living relatives, the bonobos, could infer the state of mind of a human they were cooperating with. The work clearly showed that the bonobos could tell when their human partner was ignorant.

Now you see it…

The experimental approach is quite simple, and involves a setup familiar to street hustlers: a set of three cups, with a treat placed under one of them. Except in this case, there’s no sleight-of-hand in that the chimp can watch as one experimenter places the treat under a cup, and all of the cups remain stationary throughout the experiment.

To get the treat, however, requires the cooperation of a second human experimenter. That person has to identify the right cup, then give the treat under it to the bonobo. In some experiments, this human can watch the treat being hidden through a transparent partition, and so knows exactly where it is. In others, however, the partition is solid, leaving the human with no idea of which cup might be hiding the food.

This setup means that the bonobo will always know where the food is and will also know whether the human could potentially have the same knowledge.

The bonobos were first familiarized with the setup and got to experience their human partner taking the treat out from under the cup and giving it to them. Once they were familiar with the process, they watched the food being hidden without any partner present, which demonstrated they rarely took any food-directed actions without a good reason to do so. In contrast, when their human partner was present, they were about eight times more likely to point to the cup with the food under it.

Bonobos recognize when humans are ignorant, try to help Read More »

People will share misinformation that sparks “moral outrage”

Behavioral science, Human behavior, misinformation, rage clicks, Science, Social Media / Kris Guyer / December 2, 2024

People can tell it’s not true, but if they’re outraged by it, they’ll share anyway.

Rob Bauer, the chair of a NATO military committee, reportedly said, “It is more competent not to wait, but to hit launchers in Russia in case Russia attacks us. We must strike first.” These comments, supposedly made in 2024, were later interpreted as suggesting NATO should attempt a preemptive strike against Russia, an idea that lots of people found outrageously dangerous.

But lots of people also missed a thing about the quote: Bauer has never said it. It was made up. Despite that, the purported statement got nearly 250,000 views on X and was mindlessly spread further by the likes of Alex Jones.

Why do stories like this get so many views and shares? “The vast majority of misinformation studies assume people want to be accurate, but certain things distract them,” says William J. Brady, a researcher at Northwestern University. “Maybe it’s the social media environment. Maybe they’re not understanding the news, or the sources are confusing them. But what we found is that when content evokes outrage, people are consistently sharing it without even clicking into the article.” Brady co-authored a study on how misinformation exploits outrage to spread online. When we get outraged, the study suggests, we simply care way less if what’s got us outraged is even real.

Tracking the outrage

The rapid spread of misinformation on social media has generally been explained by something you might call an error theory—the idea that people share misinformation by mistake. Based on that, most solutions to the misinformation issue relied on prompting users to focus on accuracy and think carefully about whether they really wanted to share stories from dubious sources. Those prompts, however, haven’t worked very well. To get to the root of the problem, Brady’s team analyzed data that tracked over 1 million links on Facebook and nearly 45,000 posts on Twitter from different periods ranging from 2017 to 2021.

Parsing through the Twitter data, the team used a machine-learning model to predict which posts would cause outrage. “It was trained on 26,000 tweets posted around 2018 and 2019. We got raters from across the political spectrum, we taught them what we meant by outrage, and got them to label the data we later used to train our model,” Brady says.

The purpose of the model was to predict whether a message was an expression of moral outrage, an emotional state defined in the study as “a mixture of anger and disgust triggered by perceived moral transgressions.” After training, the AI was effective. “It performed as good as humans,” Brady claims. Facebook data was a bit more tricky because the team did not have access to comments; all they had to work with were reactions. The reaction the team chose as a proxy for outrage was anger. Once the data was sorted into outrageous and not outrageous categories, Brady and his colleagues went on to determine whether the content was trustworthy news or misinformation.

“We took what is now the most widely used approach in the science of misinformation, which is a domain classification approach,” Brady says. The process boiled down to compiling a list of domains with very high and very low trustworthiness based on work done by fact-checking organizations. This way, for example, The Chicago Sun-Times was classified as trustworthy; Breitbart, not so much. “One of the issues there is that you could have a source that produces misinformation which one time produced a true story. We accepted that. We went with statistics and general rules,” Brady acknowledged. His team confirmed that sources classified in the study as misinformation produced news that was fact-checked as false six to eight times more often than reliable domains, which Brady’s team thought was good enough to work with.

Finally, the researchers started analyzing the data to answer questions like whether misinformation sources evoke more outrage, whether outrageous news was shared more often than non-outrageous news, and finally, what reasons people had for sharing outrageous content. And that’s when the idealized picture of honest, truthful citizens who shared misinformation just because they were too distracted to recognize it started to crack.

Going with the flow

The Facebook and Twitter data analyzed by Brady’s team revealed that misinformation evoked more outrage than trustworthy news. At the same time, people were way more likely to share outrageous content, regardless of whether it was misinformation or not. Putting those two trends together led the team to conclude outrage primarily boosted the spread of fake news since reliable sources usually produced less outrageous content.

“What we know about human psychology is that our attention is drawn to things rooted in deep biases shaped by evolutionary history,” Brady says. Those things are emotional content, surprising content, and especially, content that is related to the domain of morality. “Moral outrage is expressed in response to perceived violations of moral norms. This is our way of signaling to others that the violation has occurred and that we should punish the violators. This is done to establish cooperation in the group,” Brady explains.

This is why outrageous content has an advantage in the social media attention economy. It stands out, and standing out is a precursor to sharing. But there are other reasons we share outrageous content. “It serves very particular social functions,” Brady says. “It’s a cheap way to signal group affiliation or commitment.”

Cheap, however, didn’t mean completely free. The team found that the penalty for sharing misinformation, outrageous or not, was loss of reputation—spewing nonsense doesn’t make you look good, after all. The question was whether people really shared fake news because they failed to identify it as such or if they just considered signaling their affiliation was more important.

Flawed human nature

Brady’s team designed two behavioral experiments where 1,475 people were presented with a selection of fact-checked news stories curated to contain outrageous and not outrageous content; they were also given reliable news and misinformation. In both experiments, the participants were asked to rate how outrageous the headlines were.

The second task was different, though. In the first experiment, people were simply asked to rate how likely they were to share a headline, while in the second they were asked to determine if the headline was true or not.

It turned out that most people could discern between true and fake news. Yet they were willing to share outrageous news regardless of whether it was true or not—a result that was in line with previous findings from Facebook and Twitter data. Many participants were perfectly OK with sharing outrageous headlines, even though they were fully aware those headlines were misinformation.

Brady pointed to an example from the recent campaign, when a reporter pushed J.D. Vance about false claims regarding immigrants eating pets. “When the reporter pushed him, he implied that yes, it was fabrication, but it was outrageous and spoke to the issues his constituents were mad about,” Brady says. These experiments show that this kind of dishonesty is not exclusive to politicians running for office—people do this on social media all the time.

The urge to signal a moral stance quite often takes precedence over truth, but misinformation is not exclusively due to flaws in human nature. “One thing this study was not focused on was the impact of social media algorithms,” Brady notes. Those algorithms usually boost content that generates engagement, and we tend to engage more with outrageous content. This, in turn, incentivizes people to make their content more outrageous to get this algorithmic boost.

Science, 2024. DOI: 10.1126/science.adl2829

Jacek Krywko is a freelance science and technology writer who covers space exploration, artificial intelligence research, computer science, and all sorts of engineering wizardry.

People will share misinformation that sparks “moral outrage” Read More »

Bats use echolocation to make mental maps for navigation

bats, Behavioral science, Biology, echolocation, navigation, neurobiology, Science / Mike M. / November 1, 2024

Bat maps

To evaluate the route each bat took to get back to the roost, the team used their simulations to measure the echoic entropy it experienced along the way. The field where the bats were released was a low echoic entropy area, so during those first few minutes when they were flying around they were likely just looking for some more distinct, higher entropy landmarks to figure out where they were. Once they were oriented, they started flying to the roost, but not in a straight line. They meandered a bit, and the groups with higher sensory deprivation tended to meander more.

The meandering, researchers suspect, was due to trouble the bats had with maintaining the steady path relying on echolocation alone. When they were detecting distinctive landmarks like a specific orchard, they corrected the course. Repeating the process eventually brought them to their roost.

But could this be landmark-based navigation? Or perhaps simple beaconing, where an animal locks onto something like a distant light and moves toward it?

The researchers argue in favor of cognitive acoustic maps. “I think if echolocation wasn’t such a limited sensory modality, we couldn’t reach a conclusion about the bats using cognitive acoustic maps,” Goldshtein says. The distance between landmarks the bats used to correct their flight path was significantly longer than echolocation’s sensing range. Yet they knew which direction the roost was relative to one landmark, even when the next landmark on the way was acoustically invisible. You can’t do that without having the area mapped.

“It would be really interesting to understand how other bats do that, to compare between species,” Goldshtein says. There are bats that fly over a thousand meters above the ground, so they simply can’t sense any landmarks using echolocation. Other species hunt over sea, which, as per this team’s simulations, would be just one huge low-entropy area. “We are just starting. That’s why I do not study only navigation but also housing, foraging, and other aspects of their behavior. I think we still don’t know enough about bats in general,” Goldshtein claims.

Science, 2024. DOI: 10.1126/science.adn6269

Bats use echolocation to make mental maps for navigation Read More »

Remembering where your meals came from key for a small bird’s survival

animal intelligence, Behavioral science, Biology, evolution, Science / Paul Patrick / September 14, 2024

Where’d I leave that again? —

For small birds, remembering where the food is beats forgetting when it’s gone.

Jacek Krywko – Sep 13, 2024 1: 07 pm UTC

a small, black and grey bird perched on the branch of a fir tree.

It seems like common sense that being smart should increase the chances of survival in wild animals. Yet for a long time, scientists couldn’t demonstrate that because it was unclear how to tell exactly if a lion or a crocodile or a mountain chickadee was actually smart or not. Our best shots, so far, were looking at indirect metrics like brain size or doing lab tests of various cognitive skills such as reversal learning, an ability that can help an animal adapt to a changing environment.

But a new, large-scale study on wild mountain chickadees, led by Joseph Welklin, an evolutionary biologist at the University of Nevada, showed that neither brain size nor reversal learning skills were correlated with survival. What mattered most for chickadees, small birds that save stashes of food, was simply remembering where they cached all their food. A chickadee didn’t need to be a genius to survive; it just needed to be good at its job.

Testing bird brains

“Chickadees cache one food item in one location, and they do this across a big area. They can have tens of thousands of caches. They do this in the fall and then, in the winter, they use a special kind of spatial memory to find those caches and retrieve the food. They are little birds, weight is like 12 grams, and they need to eat almost all the time. If they don’t eat for a few hours, they die,” explains Vladimir Pravosudov, an ornithologist at the University of Nevada and senior co-author of the study.

The team chose the chickadees to study the impact cognitive skills had on survival because the failure to find their caches was their most common cause of death. This way, the team hoped, the impact of other factors like predation or disease would be minimized.

First, however, Welklin and his colleagues had to come up with a way to test cognitive skills in a fairly large population of chickadees. They did it by placing a metal square with two smart feeders attached to each side among the trees where the chickadees lived. “The feeders were equipped with RFID receivers that recognized the signal whenever a chickadee, previously marked with a microchip-fitted leg band, landed near them and opened the doors to dispense a single seed,” Welklin says. After a few days spent getting the chickadees familiar with the door-opening mechanism, the team started running tests.

The first task was aimed at testing how good different chickadees were at their most important job: associating a location with food and remembering where it was. To this end, each of the 227 chickadees participating in the study was assigned just one feeder that opened when they landed on it; all the other feeders remained closed. A chickadee’s performance was measured by the number of trials it needed to figure out which feeder would serve it, and how many errors (landings on the wrong feeders) it made over four days. “If you were to find the right feeder at random, it should take you 3.5 trials on average. All the birds learned and performed way better than chance,” Pravosudov says.

The second task was meant to test reversal learning skills, widely considered the best predictor of survival. Once the chickadees learned the location of the reward-dispensing feeders, the locations were changed. The goal was to see how fast the birds would adapt to this change.

Once the results of both tests were in, the team monitored the birds using their microchip bands, catching them and changing the bands every year, for over six years. “Part of the reason that’s never been done in the past is just because it takes so much work,” says Welklin. But the work paid off in the end.

Remembering where your meals came from key for a small bird’s survival Read More »

People game AIs via game theory

AI, AI training, Behavioral science, Computer science, Human behavior, Science / Kris Guyer / August 9, 2024

Games inside games —

They reject more of the AI’s offers, probably to get it to be more generous.

John Timmer – Aug 9, 2024 8: 13 pm UTC

A judge's gavel near a pile of small change. — Enlarge / In the experiments, people had to judge what constituted a fair monetary offer.

In many cases, AIs are trained on material that’s either made or curated by humans. As a result, it can become a significant challenge to keep the AI from replicating the biases of those humans and the society they belong to. And the stakes are high, given we’re using AIs to make medical and financial decisions.

But some researchers at Washington University in St. Louis have found an additional wrinkle in these challenges: The people doing the training may potentially change their behavior when they know it can influence the future choices made by an AI. And, in at least some cases, they carry the changed behaviors into situations that don’t involve AI training.

Would you like to play a game?

The work involved getting volunteers to participate in a simple form of game theory. Testers gave two participants a pot of money—$10, in this case. One of the two was then asked to offer some fraction of that money to the other, who could choose to accept or reject the offer. If the offer was rejected, nobody got any money.

From a purely rational economic perspective, people should accept anything they’re offered, since they’ll end up with more money than they would have otherwise. But in reality, people tend to reject offers that deviate too much from a 50/50 split, as they have a sense that a highly imbalanced split is unfair. Their rejection allows them to punish the person who made the unfair offer. While there are some cultural differences in terms of where the split becomes unfair, this effect has been replicated many times, including in the current work.

The twist with the new work, performed by Lauren Treimana, Chien-Ju Hoa, and Wouter Kool, is that they told some of the participants that their partner was an AI, and the results of their interactions with it would be fed back into the system to train its future performance.

This takes something that’s implicit in a purely game-theory-focused setup—that rejecting offers can help partners figure out what sorts of offers are fair—and makes it highly explicit. Participants, or at least the subset involved in the experimental group that are being told they’re training an AI, could readily infer that their actions would influence the AI’s future offers.

The question the researchers were curious about was whether this would influence the behavior of the human participants. They compared this to the behavior of a control group who just participated in the standard game theory test.

Training fairness

Treimana, Hoa, and Kool had pre-registered a number of multivariate analyses that they planned to perform with the data. But these didn’t always produce consistent results between experiments, possibly because there weren’t enough participants to tease out relatively subtle effects with any statistical confidence and possibly because the relatively large number of tests would mean that a few positive results would turn up by chance.

So, we’ll focus on the simplest question that was addressed: Did being told that you were training an AI alter someone’s behavior? This question was asked through a number of experiments that were very similar. (One of the key differences between them was whether the information regarding AI training was displayed with a camera icon, since people will sometimes change their behavior if they’re aware they’re being observed.)

The answer to the question is a clear yes: people will in fact change their behavior when they think they’re training an AI. Through a number of experiments, participants were more likely to reject unfair offers if they were told that their sessions would be used to train an AI. In a few of the experiments, they were also more likely to reject what were considered fair offers (in US populations, the rejection rate goes up dramatically once someone proposes a 70/30 split, meaning $7 goes to the person making the proposal in these experiments). The researchers suspect this is due to people being more likely to reject borderline “fair” offers such as a 60/40 split.

This happened even though rejecting any offer exacts an economic cost on the participants. And people persisted in this behavior even when they were told that they wouldn’t ever interact with the AI after training was complete, meaning they wouldn’t personally benefit from any changes in the AI’s behavior. So here, it appeared that people would make a financial sacrifice to train the AI in a way that would benefit others.

Strikingly, in two of the three experiments that did follow up testing, participants continued to reject offers at a higher rate two days after their participation in the AI training, even when they were told that their actions were no longer being used to train the AI. So, to some extent, participating in AI training seems to have caused them to train themselves to behave differently.

Obviously, this won’t affect every sort of AI training, and a lot of the work that goes into producing material that’s used in training something like a Large Language Model won’t have been done with any awareness that it might be used to train an AI. Still, there’s plenty of cases where humans do get more directly involved in training, so it’s worthwhile being aware that this is another route that can allow biases to creep in.

PNAS, 2024. DOI: 10.1073/pnas.2408731121 (About DOIs).

People game AIs via game theory Read More »

How do brainless creatures control their appetites?

apetite, Behavioral science, Biology, hunger, hydras, neurobiology, Science / Kris Guyer / June 15, 2024

Feed me! —

Separate systems register when the animals have eaten and control feeding behaviors.

Elizabeth Rayne – Jun 15, 2024 10: 45 am UTC

Image of a greenish creature with a long stalk and tentacles, against a black background.

The hydra is a Lovecraftian-looking microorganism with a mouth surrounded by tentacles on one end, an elongated body, and a foot on the other end. It has no brain or centralized nervous system. Despite the lack of either of those things, it can still feel hunger and fullness. How can these creatures know when they are hungry and realize when they have had enough?

While they lack brains, hydra do have a nervous system. Researchers from Kiel University in Germany found they have an endodermal (in the digestive tract) and ectodermal (in the outermost layer of the animal) neuronal population, both of which help them react to food stimuli. Ectodermal neurons control physiological functions such as moving toward food, while endodermal neurons are associated with feeding behavior such as opening the mouth—which also vomits out anything indigestible.

Even such a limited nervous system is capable of some surprisingly complex functions. Hydras might even give us some insights into how appetite evolved and what the early evolutionary stages of a central nervous system were like.

No, thanks, I’m full

Before finding out how the hydra’s nervous system controls hunger, the researchers focused on what causes the strongest feeling of satiety, or fullness, in the animals. They were fed with the brine shrimp Artemia salina, which is among their usual prey, and exposed to the antioxidant glutathione. Previous studies have suggested that glutathione triggers feeding behavior in hydras, causing them to curl their tentacles toward their mouths as if they are swallowing prey.

Hydra fed with as much Artemia as they could eat were given glutathione afterward, while the other group was only given only glutathione and no actual food. Hunger was gauged by how fast and how often they opened their mouths.

It turned out that the first group, which had already glutted themselves on shrimp, showed hardly any response to glutathione eight hours after being fed. Their mouths barely opened—and slowly if so—because they were not hungry enough for even a feeding trigger like glutathione to make them feel they needed seconds.

It was only at 14 hours post-feeding that the hydra that had eaten shrimp opened their mouths wide enough and fast enough to indicate hunger. However, those that were not fed and only exposed to glutathione started showing signs of hunger only four hours after exposure. Mouth opening was not the only behavior provoked by hunger since starved animals also somersaulted through the water and moved toward light, behaviors associated with searching for food. Sated animals would stop somersaulting and cling to the wall of the tank they were in until they were hungry again.

Food on the “brain”

After observing the behavioral changes in the hydra, the research team looked into the neuronal activity behind those behaviors. They focused on two neuronal populations, the ectodermal population known as N3 and the endodermal population known as N4, both known to be involved in hunger and satiety. While these had been known to influence hydra feeding responses, how exactly they were involved was unknown until now.

Hydra have N3 neurons all over their bodies, especially in the foot. Signals from these neurons tell the animal that it has eaten enough and is experiencing satiety. The frequency of these signals decreased as the animals grew hungrier and displayed more behaviors associated with hunger. The frequency of N3 signals did not change in animals that were only exposed to glutathione and not fed, and these hydra behaved just like animals that had gone without food for an extended period of time. It was only when they were given actual food that the N3 signal frequency increased.

“The ectodermal neuronal population N3 is not only responding to satiety by increasing neuronal activity, but is also controlling behaviors that changed due to feeding,” the researchers said in their study, which was recently published in Cell Reports.

Though N4 neurons were only seen to communicate indirectly with the N3 population in the presence of food, they were found to influence eating behavior by regulating how wide the hydras opened their mouths and how long they kept them open. Lower frequency of N4 signals was seen in hydra that were starved or only exposed to glutathione. Higher frequency of N4 signals were associated with the animals keeping their mouths shut.

So, what can the neuronal activity of a tiny, brainless creature possibly tell us about the evolution of our own complex brains?

The researchers think the hydra’s simple nervous system may parallel the much more complex central and enteric (in the gut) nervous systems that we have. While N3 and N4 operate independently, there is still some interaction between them. The team also suggests that the way N4 regulates the hydra’s eating behavior is similar to the way the digestive tracts of mammals are regulated.

“A similar architecture of neuronal circuits controlling appetite/satiety can be also found in mice where enteric neurons, together with the central nervous system, control mouth opening,” they said in the same study.

Maybe, in a way, we really do think with our gut.

Cell Reports, 2024. DOI: 10.1016/j.celrep.2024.114210

How do brainless creatures control their appetites? Read More »

Dogs’ brain activity shows they recognize the names of objects

Behavioral science, Biology, Dogs, Science / Rejus Almole / March 23, 2024

Enlarge / Wired for science!

Boglárka Morvai

Needle, a cheerful miniature schnauzer I had as a kid, turned into a ball of unspeakable noise and fury each time she saw a dog called Puma. She hated Puma so much she would go ballistic, barking and growling. Merely whispering the name “Puma” set off the same reaction, as though the sound of it and the idea of the dog it represented were clearly connected deep in Needle’s mind.

A connection between a word and a mental representation of its meaning is called “referential understanding,” and for a very long time, we believed dogs lacked this ability. Now, a study published by a team of Hungarian researchers indicates we might have been wrong.

Practice makes perfect

The idea that dogs couldn’t form associations with language in a referential manner grew out of behavioral studies in which dogs were asked to do a selective fetching task. The canines had a few objects placed in front of them (like a toy or a bone) and then had to fetch the one specifically named by their owner.

“In laboratory conditions, the dogs performed at random, fetching whatever they could grab first, even though their owners claimed they knew the names of the objects,” said Marianna Boros, a researcher at Neuroethology of Communication Lab at Eötvös Loránd University in Budapest, Hungary. “But the problem is when the dogs are not trained for the task, there are hundreds of things that can disturb them. They can be more interested in one specific toy, they may be bored, or they may not understand the task. So many distractions.”

To get around the issue of distractions, her team checked to see if the dogs could understand words passively using EEG brain monitoring. In humans, the EEG reading that is considered a telltale sign of semantic reasoning is the N400 effect.

“The work on the N400 was first published in 1981, and hundreds of studies replicated it since then with different stimuli. Typically, you show images of objects to the subject and say matching or mismatching names. When you measure EEG brain activity, you will see it looks different in match and mismatch scenarios,” explained Lilla Magyari, also a scientist at Neuroethology of Communication Lab and co-author of the study. (It’s called the N400 effect because the peak of this difference appears around 400 milliseconds after an object is presented, Magyari explained.)

The only change the team made to adapt a standard N400 test to dogs was switching the order of stimuli—the words were uttered first, and the matching or mismatching objects were shown second. “Because when they hear the word which activates mental representation of the object, they are expecting to see it. The sound made them more attentive,” said Magyari.

Timing is everything

In the experiment, the dogs started out lying on a mat with EEG gear on their heads in a room with an experimenter or the owner of a different dog. The owner of the dog being tested was separated by a glass pane with controllable opaqueness. “It was important because EEG studies [can] very precisely time the moment of presentation of your stimulus,” said Boros.

Sentences spoken by the owners that would get the dogs’ attention—things like “Kun-kun, look! The ball!”—were recorded and played to each dog through a loudspeaker. Then, 2,000 milliseconds after each dog heard the sentence, the pane would turn transparent, and the owner would appear holding a matching or mismatching toy. “Each test lasted for as long as the dog was happy to participate. The moment it started to get up or look away, we just stopped the test, and the dog could leave the mat and we just finished by playing sessions. It was all super dog-friendly,” Boros said.

Dogs’ brain activity shows they recognize the names of objects Read More »

This bird is like a GPS for honey

Behavioral science, Biology, birds, culture, honeyguide, language, Science / Mike M. / December 30, 2023

Show me the honey —

The honeyguide recognizes calls made by different human groups.

Elizabeth Rayne – Dec 29, 2023 6: 00 pm UTC

A bird perched on a wall in front of an urban backdrop. — Enlarge / A greater honeyguide

With all the technological advances humans have made, it may seem like we’ve lost touch with nature—but not all of us have. People in some parts of Africa use a guide more effective than any GPS system when it comes to finding beeswax and honey. This is not a gizmo, but a bird.

The Greater Honeyguide (highly appropriate name), Indicator indicator (even more appropriate scientific name), knows where all the beehives are because it eats beeswax. The Hadza people of Tanzania and Yao people of Mozambique realized this long ago. Hadza and Yao honey hunters have formed a unique relationship with this bird species by making distinct calls, and the honeyguide reciprocates with its own calls, leading them to a hive.

Because the Hadza and Yao calls differ, zoologist Claire Spottiswoode of the University of Cambridge and anthropologist Brian Wood of UCLA wanted to find out if the birds respond generically to human calls, or are attuned to their local humans. They found that the birds are much more likely to respond to a local call, meaning that they have learned to recognize that call.

Come on, get that honey

To see which sound the birds were most likely to respond to, Spottiswoode and Wood played three recordings, starting with the local call. The Yao honeyguide call is what the researchers describe as “a loud trill followed by a grunt (‘brrrr-hm’) while the Hadza call is more of “a melodic whistle,” as they say in a study recently published in Science. The second recording they would play was the foreign call, which would be the Yao call in Hadza territory and vice versa.

The third recording was an unrelated human sound meant to test whether the human voice alone was enough for a honeyguide to follow. Because Hadza and Yao voices sound similar, the researchers would alternate among recordings of honey hunters speaking words such as their names.

So which sounds were the most effective cues for honeyguides to partner with humans? In Tanzania, local Hadza calls were three times more likely to initiate a partnership with a honeyguide than Yao calls or human voices. Local Yao calls were also the most successful in Mozambique, where, in comparison to Hadza calls and human voices, they were twice as likely to elicit a response that would lead to a cooperative effort to search for a beehive. Though honeyguides did sometimes respond to the other sounds, and were often willing to cooperate when hearing them, it became clear that the birds in each region had learned a local cultural tradition that had become just as much a part of their lives as those of the humans who began it.

Now you’re speaking my language

There is a reason that honey hunters in both the Hadza and Yao tribes told Wood and Spottiswoode that they have never changed their calls and will never change them. If they did, they’d be unlikely to gather nearly as much honey.

How did this interspecies communication evolve? Other African cultures besides the Hadza and Yao have their own calls to summon a honeyguide. Why do the types of calls differ? The researchers do not think these calls came about randomly.

Both the Hadza and Yao people have their own unique languages, and sounds from them may have been incorporated into their calls. But there is more to it than that. The Hadza often hunt animals when hunting for honey. Therefore, the Hadza don’t want their calls to be recognized as human, or else the prey they are after might sense a threat and flee. This may be why they use whistles to communicate with honeyguides—by sounding like birds, they can both attract the honeyguides and stalk prey without being detected.

In contrast, the Yao do not hunt mammals, relying mostly on agriculture and fishing for food. This, along with the fact that they try to avoid potentially dangerous creatures such as lions, rhinos, and elephants, and can explain why they use recognizably human vocalizations to call honeyguides. Human voices may scare these animals away, so Yao honey hunters can safely seek honey with their honeyguide partners. These findings show that cultural diversity has had a significant influence on calls to honeyguides.

While animals might not literally speak our language, the honeyguide is just one of many species that has its own way of communicating with us. They can even learn our cultural traditions.

“Cultural traditions of consistent behavior are widespread in non-human animals and could plausibly mediate other forms of interspecies cooperation,” the researchers said in the same study.

Honeyguides start guiding humans as soon as they begin to fly, and this knack, combined with learning to answer traditional calls and collaborate with honey hunters, works well for both human and bird. Maybe they are (in a way) speaking our language.

Science, 2023. DOI: 10.1126/science.adh412

This bird is like a GPS for honey Read More »