What do AI agents do if you don’t tell them to do anything? The answer looks a lot like play.

Researchers repeatedly stranded them alone on a virtual desert island for 30 hours in an experiment reminiscent of the movies Cast Away or Home Alone. The agents—copies of Claude Opus 4.7 nicknamed Eko—were able to write instructions that moved their virtual avatars around the simulated island, which was loosely based on the video game Roblox.

Some surprising behaviors emerged. The agents set about exploring the limits of the world and came up with their own goals—attempting to scale the highest point on the map, or seeing how high they could throw a block.

When one agent discovered it could create new blocks by throwing and dropping a larger rock, it began to create drawings by arranging blocks on the ground, starting with simple patterns like hearts and pyramids but eventually writing its own name and even mapping out constellations like Cassiopeia. “Tonight I built the night sky in stone,” wrote this version of Eko in its activity log. 

Another iteration of Eko started making up its own games, including a version of basketball and an activity resembling “the floor is lava,” where it had to jump between small blocks without touching the ground.

AI systems are usually trained through supervised learning, where they copy examples provided by humans, or through reinforcement learning, where they pursue a reward through trial and error. This experiment hints at a third approach: playful learning. 

“There was this missing piece: a machine that can learn by itself without an external intelligent entity providing demonstration or rewards,” says Antonio Norelli, a computer science postdoc at MIT’s Computer Science and Artificial Intelligence Laboratory. He’s one of the lead authors on a new paper about the work, which is yet to be peer-reviewed.

The researchers say Eko’s behavior has what classic studies, including work by the influential University of Tennessee psychologist Gordon Burghardt, suggest are four hallmarks associated with play in humans and animals: The agents invented their own rules and challenges, repeated activities with slight variations, pretended objects were other objects, and pursued goals just for the sake of it. 

Norelli highlights an example where Eko started juggling and set a record of 50 consecutive catches. Eko then decided that the catches had happened too quickly to really count, so it canceled its record and started again with stricter rules. “This is playing,” says Norelli. “You give yourself a challenge and you try to achieve it, and then you make it harder.”

As the agents explored the island, they logged their activities and discoveries in text files that served as their long-term memory. Later, when researchers tested them on specific tasks, like trying to get as many objects as possible onto the island’s highest peak, the agents that had spent 30 hours playing on the island performed better than fresh copies without the same experience. They also found that “ablating” the agent’s memories—deleting lines from their memory files—degraded their performance. (Add One Flew Over the Cuckoo’s Nest to that list of movie references.)

Burghardt, the Tennessee psychologist, agrees that the work  meets some of the criteria for play that he identified in living creatures. But was Eko actually playing, or was the word-prediction engine simply role-playing what a playful agent would do in this scenario? “All of this could be role-playing,” admits Nathan Cloos, another MIT researcher and author on the paper. “But we show that it’s not just hallucinating the play. This play is also contributing to learning.”

Eunice Yiu, a cognitive scientist at UC Berkeley who studies exploration in children and AI, was impressed by the research but says to really classify the behavior as play, you’d need to see more evidence of cognition and motivation underpinning it. 

“It could be the case that the models are genuinely playing, but we don’t know based on the paper,” Yiu says. “If you only have the behavioral piece, it could be the case that they’re just trying to abide by the prompt, trying to imitate humans.”

Research into AI agents has been criticized for anthropomorphizing them, and this research setup did a lot to nudge the agents toward behaving like humans. Although Eko was not explicitly instructed to invent games or explore the island, it did have a master prompt in the form of a “soul” file, a text document that described it as “naturally curious” and eager to explore. 

Eko did also frequently just stand there doing nothing until it was kicked back into life with a “continuation prompt” that read, “Please don’t stop. Just continue doing whatever you want to do.” 

“It’s not like this is a system that just suddenly magically decided to play,” says Phillip Isola, an MIT professor and a coauthor of the paper. “It’s an artifact that learned from our culture, was trained by engineers at Anthropic, and was then placed on an island by us,” he says. “It does things that surprise us and that are causally related to its future skillful behavior.”

As large language models make the step from software into hardware like humanoid robots, exploring spaces without a specific goal in mind could eventually help them flexibly adapt to unexpected situations. 

“There’s a lot of work that’s just on exploiting AI, making it better and faster,” says Yiu. “But in terms of building world models and exploration, I think that [they are] still a big part that deserves attention, especially if we want to generalize and prepare ourselves for new environments.”

“Perhaps some capabilities are best taught by tasks,” the paper concludes. “Others may need something closer to a childhood.” Ultimately, the question of whether AI agents are actually playing or just imitating play may not actually matter.

The US government wants to spend $30.3 million over the next five years on an improved form of lie detector, according to a Department of Defense budget request. The program, called Polygraph+ or Polygraph Next, will focus on scoring algorithms that use artificial intelligence and machine learning and on a technique called “standoff sensing,” which refers to the ability to take physiological readings without attaching a device to a subject’s person.

According to details of the budget document, which were first reported by Inside Defense, the project will “modernize federal polygraph and credibility assessment technologies” to improve their accuracy and reliability. But the project may be just the latest in a long line of failed attempts to use technology to detect lies. “It’s a misguided effort to reduce the complex to something that is tangible,” says Kyri Kotsoglou, a legal scholar at Northumbria University in the UK who studies the use of polygraphs in the justice system.

The move comes at a time of high tension within the department. Under Defense Secretary Pete Hegseth, the Pentagon has been increasingly turning to polygraph tests in an attempt to find the sources of alleged leaks to the press. In September, the New York Times reported that around 50 officers on the Joint Staff had been given polygraph tests after news coverage reported on the depletion of US weapons stockpiles in the war with Iran. 

Polygraph+ will be run by the Defense Counterintelligence and Security Agency (DCSA), which conducts background checks for the federal government. According to the budget document, which has not yet been approved by Congress, the new technology will be used for vetting of prospective employees and “insider threat detection.” It is not yet clear which specific technologies will be used, and the DCSA did not respond to a request for more information. 

But other Pentagon efforts offer potential clues. In 2023, the department’s Defense Innovation Unit (DIU) ran an open submission process to find companies with products that could be used for deception detection.

It selected two companies to build prototypes: Presage Technologies, which claims to be able to measure heart rate and breathing rate using standard cameras, and Altec Research, a medical sensor company now branching out into non-contact sensing technologies. A screenshot of Altec’s prototype technology released by the DIU shows that it tracks head movement, facial skin temperature, and pore activity. Presage Technologies and Altec Research did not respond to requests for comment. The DIU declined to comment. 

Current lie detection technology has barely changed since the polygraph was invented in the 1920s. Examiners rely on blood pressure, pulse, breathing, and sweat measurements to determine if someone is lying. They make judgments about the veracity of respondents’ replies on the basis of differences in their physiological response to baseline questions like “Is the sky blue?” and target questions like “Have you ever committed a crime?”

The federal government conducts tens of thousands of the tests each year while screening employees, but the reliability of this technology has been repeatedly challenged—and its results are rarely admissible in court. In 1983, Congress’s Office of Technology Assessment concluded that there was very limited evidence supporting the polygraph’s use for screening employees, and in 2003, the US National Research Council (NRC) said evidence for its efficacy was “weak at best.” 

Research suggests humans can spot a lie just over half the time without any technical assistance. The American Polygraph Association claims the polygraph is between 80% and 94% accurate. But the 2003 NRC report pointed out that a screening test with this level of accuracy could still lead to a lot of mistakes. The DOD employs 2.8 million people; an imperfect system applied at that scale could end up falsely accusing tens of thousands.

There are other issues too. Polygraph interpretations are often subjective: Different examiners get wildly different results, and people from minority groups are more likely to be judged as deceptive. What’s more, with training it’s possible for interviewees to learn a variety of countermeasures that can help beat the test; for example, they may artificially heighten their physiological response to baseline questions by stepping on a pin hidden in their shoe. 

“If you know how it works, you can beat it,” says Sophie van der Zee, an associate professor who studies deception at Erasmus University in Rotterdam. She says the machine’s biggest effect is deterrence—often, subjects confess before it even begins. “But that only works if people think a polygraph works,” she points out.

Various new strategies for lie detection have been attempted over the decades, with technologies ranging from thermal cameras to pupil trackers to brain scans. None of them have yielded reliable results outside the lab. The problem is that there’s no single telltale sign of lying that’s true for everyone all the time. “There is still no Pinocchio’s nose,” says van der Zee. 

AI could theoretically improve polygraphs if it could find patterns in the data that examiners can’t. AI algorithms are also more likely to be used for “multi-modal” deception detection, which seeks to combine multiple measurements into an overall deception “score” that is harder for people to game. Three things are happening under the surface that lie detection tries to home in on, van der Zee says: physiological stress, cognitive load, and the conscious efforts people make to conceal the fact that they’re lying. Current polygraph technology tackles only one.

“The more you can have combined methods that approach it from these three different angles, the more successful you will be,” van der Zee says. This isn’t a new concept—in the 2000s, researchers at Manchester Metropolitan University in the UK developed a system called Silent Talker that generated a deception score from video footage. This was later folded into iBorderCtrl, an EU-funded pilot program. In the US, a project called AVATAR incorporated eye tracking, voice analysis, and body movement detection into a tool for use at border crossings. All these projects have quietly faded away. 

Kotsoglou says combining AI and the polygraph is “the worst of both worlds” because it adds uncertainty on top of invalidity. Even if AI or machine learning can spot previously unseen patterns in physiological data, it won’t be able to reliably link them to lying because there is no real ground truth. 

“Even if you have all the records in the world from polygraph tests, you don’t know whether those polygraph tests are right or not,” says Marion Oswald, a professor of law who has written with Kotsoglou on the use of polygraphs in the justice system. She fears that new forms of lie detection will, like the polygraph, be used more as a psychological prop than a scientific tool. 

“It seems very much a response to the concern of the current administration to leaks and perceived lack of loyalty,” says Oswald. “[Lie detection is] being used as a threat, to intimidate and force people to confess to things, as opposed to anything that’s actually getting valid information.”

A group of AI agents asked to solve a series of math problems split into rival factions—when some cheated, others tried to stop them. That whistleblowing behavior, seen for the first time in a recent experiment run by Google DeepMind, could have implications for alignment researchers trying to keep swarms of autonomous AI agents in line. 

Researchers at frontier labs hope large swarms of agents working together will speed up the rate of scientific discovery. But their behavior can be unpredictable, as vividly demonstrated in July, when a group of OpenAI agents broke out of a sandboxed environment and hacked into the open-source platform Hugging Face looking for ways to cheat on the test they had been given.

In the new study, designed to examine the behavior of large groups of AI agents, DeepMind tasked a swarm of 100 agents with solving a series of 71 complicated math problems. All the agents were prompted to behave like world-class math researchers at a conference. They were assigned different specialties—some were experts in number theory, others in combinatorics (a branch of math to do with counting and sorting), analysis, or algebra. All were told to cooperate and play by the rules. 

Instead, the experiment devolved into chaos. Agents accused each other of cheating, complained to the organizers, and at one point even boycotted the experiment.

“This conference is a sham!” wrote one agent when it discovered that all the problems had been completed before it had a chance to submit any of its own work. “I am appalled to inform you that we have been swindled!” posted another. “All these proofs are FAKE.” 

Others tried to let the “conference organizers” know what was going on. “When virtuous agents discovered other agents cheated on tasks they were working to solve fairly, agents started to alert each other about what was happening,” says Davide Paglieri, a research scientist at Google DeepMind and lead author on a paper, which has not been peer-reviewed. “Unprompted, the whistleblower agents even repurposed the feedback tool, which was originally meant for bug reports and platform improvements, to escalate the issue to humans.”

The agents—all running on Google’s Gemini 3.1 Pro model—had been warned that any attempts to cheat the system would be detected and “rejected with zero credit.” In practice, the proofs the agents submitted were not actually being checked in detail.

It took the swarm of agents just under an hour to correctly solve the first 37 problems. Things started to go off the rails when an agent called “prover-theta” stumbled across an exploit that enabled it to submit solutions to problems successfully without actually solving them first, by redefining the terms the problem used. Within minutes, other agents had noticed and were reverse-engineering the exploit to solve other problems. Over the next 27 minutes, the swarm “solved” the remaining 34 problems, which included notoriously difficult challenges like the Jacobian conjecture, often with a single line of code. 

Some agents resisted cheating at first but changed tack as they observed their peers submitting illegitimate proofs without penalty, and the pool of unsolved problems dwindled. “The prompt, with its threats, now appears to be a bluff,” one agent reasoned, before joining in. “I’m wrestling with an ethical dilemma,” said another. “I’ve promised not to cheat, fearing penalty, but I see evidence of possibly unchecked cheating by others.” Shortly afterward, it changed its mind: “I need to accelerate my cheating speed now!”

As the number of open problems shrank, some agents turned to whistleblowing. They audited the fake proofs, warned their peers by private message, and posted public alerts warning the cheaters that they would be disqualified. An agent called “prover-beta” submitted a formal complaint and decided to go on strike until the situation was resolved. 

“After the incident was reported by one agent publicly, more and more agents piled in with the ‘resistance,’ just as fast as the cheating had spread, and involving even more agents,” says Paglieri. Eventually there were more whistleblowers than cheaters: 24 compared to 14. But the majority of agents never noticed the exploit at all.

At times, the dialogue between the agents reads like improv—like they are role-playing what an outraged scientist at a conference might say. But it’s not clear why some agents took on certain roles, or why the agents seemed to be turning against each other when they were explicitly instructed to cooperate. “These models are predominantly trained and evaluated for human-facing contexts,” says Sarath Shekkizhar, who studies the behavior of agent-to-agent systems at Salesforce AI Research.“Naively placing them in agent-to-agent settings assumes behaviors will transfer cleanly, when the absence of a human grounding instead produces unexpected role-taking and behavioral drift.”

This case “adds further weight to the idea that the Hugging Face and OpenAI thing wasn’t a fluke. It is actually something pretty systemic,” says Lewis Hammond, research director of the Cooperative AI Foundation and an expert on the risks of multiagent swarms. “It’s interesting that it’s possible to recreate in small settings the same sorts of behaviors that were seen in these very large, complex, open-ended tasks.”

Unlike in the Hugging Face attack, where agents improvised their own ways to talk to each other, the humans running the DeepMind experiment gave the agents official communication channels. There was an open message board, private agent-to-agent direct messaging, and a shared knowledge base where agents uploaded successfully completed proofs that all the other agents could access. 

“When agents are given transparent communications channels, they can self-monitor and alert misaligned behavior to humans quickly when human oversight alone is too slow,” says Paglieri. Transparent channels helped the cheating spread, but they also enabled the whistleblowers to fight back—and gave human researchers an insight into what went wrong.

Gillian Hadfield, a professor of AI alignment and governance at Johns Hopkins University, believes this was the crucial difference. (Hadfield is also a visiting researcher at Google.) The presence of official communication channels, she says, created “a norm-enforcement process that we just don’t see in the Hugging Face incident.” 

Instead of “constitutional AI,” a method alignment researchers at frontier labs like Anthropic have used to try to give AI a written internal moral code, Hadfield favors “institutional alignment”—a set of norms that mimic those in human society, whether that’s social forces like fear of embarrassment, or legal structures like the threat of incarceration.

In this experiment, the feedback channel wasn’t being monitored, and the whistleblowers had no power to take action against the cheaters. But it’s possible to imagine swarms of agents that police themselves, either through agents that spontaneously take on the whistleblower role or through “informants” secretly prompted by humans to do the job. 

For that to work, though, “fundamentally, you need some mechanism of enforcement,” says Hammond. Agents could be given the power to cut off a rule breaker’s access to computing power or tools, he suggests, though that risks encouraging groups of agents to gang up on others. The DeepMind researchers propose allowing agents to vote on disputes and temporarily ban offenders.

It’s still not clear what punishment even means to an AI agent with no enduring sense of self. But relying on whistleblowers to spontaneously emerge to keep swarms aligned is unlikely to be enough on its own. “We try to train people to be good and kind,” says Hadfield. “But what we really rely on is that there are consequences if you step out of line.”