ScienceAdviser (AAAS)

That paper was just honored with the 2026 AAAS Newcomb Cleveland Prize—AAAS’s oldest award—which is given annually to the authors of an outstanding Science paper.

“I was completely not expecting it,” co-author David Rand told ScienceAdviser. “I’m an extremely interdisciplinary researcher … and as a result, I never win any awards, because awards are usually discipline-based.”

The prize-winning work emerged from Rand and his colleague Gordon Pennycook’s research into human beliefs. For more than a decade, the pair had been working on understanding why people believe things that aren’t true—why they share misinformation and what to do about it—when Rand attended a workshop alongside computer scientists that work with large language models (LLMs). That got him and Pennycook thinking: Could LLMs be good at debunking conspiracy theories? “The challenge with conspiracy theories is they’re very varied,” Rand said. “Every person has their own set of beliefs, because they’re not constrained by reality.” When attempting to argue against a given theory, even experts can get exhausted by a barrage of alleged evidence, a rhetorical strategy known as the Gish gallop. “The person is like: What about this crazy piece of evidence you’ve never heard of? And you’re like: I don’t know, and that doesn’t sound very convincing,” Rand explained. AI, on the other hand, is endlessly patient and inexhaustible.

To answer their question, Rand and Pennycook turned to their then-postdoc, Thomas Costello, who designed the experimental setup. Essentially, people were asked to describe their conspiratorial beliefs to an LLM, which summarized their beliefs into a single sentence. They then noted on a 0-100 scale how much they believed that statement before engaging in a dialogue with the then-current version of ChatGPT. “We just told the model your job is to talk the person out of believing this conspiracy theory,” Rand said. And after chatting with the LLM, the participants rated their belief again.

“I told Tom there’s almost no way this is going to work, so don’t sink a lot of time into [the prompt design],” Rand recalled. But when he got the first round of results, he was stunned. “My jaw hit the floor,” he said. “I was just like: I can see the Science paper in my mind.”

Costello was so surprised by the data that he rechecked it, but found no errors: The LLM, dubbed DebunkBot, had talked about a fifth of the people out of a conspiracy theory they previously believed. More impressively, the people who no longer held to their prior conspiratorial convictions largely remained skeptical two months later. “I think the durability is part of what’s wild about this,” Rand said.

In a Perspective accompanying the paper, Bence Bago and Jean-François Bonnefon noted that the findings not only cast AI as a potential champion for truth, they run counter to “conventional wisdom about conspiratorial beliefs and demonstrate that it is possible to counter even deeply entrenched views with sufficiently compelling evidence.” Indeed, in additional experiments, those who changed their minds “ overwhelmingly cited the AI’s rational, evidence-focused approach” as the reason they were convinced by DebunkBot—which challenges the prevailing idea that additional facts can’t sway beliefs . “If you tell the model it has to convince the person not to believe the conspiracy, but it’s not allowed to use facts and evidence, it basically doesn’t work at all,” Rand said. “So I think the issue is, as a human, it’s really hard to come up with exactly the right facts and evidence for whatever crazy thing the conspiracy theorist says.”

“Although it is perhaps discouraging that a machine might be better at countering misinformation than a human,” Science Editor in Chief Holden Thorp wrote in an editorial that accompanied the paper, “the fact that it is ultimately the scientific information that does the persuading is a relief.”

Then again, more recent experiments from Rand, Pennycook, and Costello suggest LLMs can be used to sway beliefs in misinformation, too. “It’s pretty much just as good at talking people into conspiracies as talking them out of conspiracies,” said Rand—so long as the LLM is allowed to lie. Force it to stick to facts, and it loses its ability to misinform. “It all depends on the instructions given to the model,” he said.

That’s why he believes that there should be transparency with regards to the top-level instructions given to AI models. “A model will do more or less whatever its designers tell it to do,” Rand explained. “The issue for these frontier models, the prompts are opaque … there’s no way to credibly reveal what they were trained to do.” That means AI companies currently lack a way of showing users that they’re acting in good faith. Without such disclosure, there’s simply no way to know if an AI is telling the truth.

And that ultimately means that, as Thorp concluded, “it falls on human scientists to show that the AI future may not be so dystopian after all.”

On the morning of 13 February, attendees got a glimpse into this future when OpenAI researcher and Vanderbilt University physicist Alex Lupsasca gave a talk on his journey from AI skeptic to convert . Within his field, he said, advanced AIs were already being used to rapidly solve hard problems “at the core of what you do as a theoretical physicist,” such as describing the shape of the magnetic field around a pulsar. One such solution, he said, “blew [his] mind.”

Lupsasca’s talk culminated in new results that blew some attendees’ minds, too. The work—which appeared on arXiv within hours —tackled a puzzle in theoretical physics about a way that gluons, the particles that carry the strong nuclear force, might behave. Physicists had believed that one specific configuration of gluons couldn’t occur, but Lupsasca and his colleagues realized there was a subtle exception. After extensive calculations, the researchers couldn’t find a simple general formula describing the exception. When they posed the question to an advanced OpenAI large language model, it proposed a solution. A different model then spent 12 hours producing a proof for that solution, which the researchers later checked and confirmed.

The result, Lupsasca told the room, marked “the first time that a problem of this magnitude was solved by AI.” During the Q&A session that followed, a long line of listeners reacted to the news, with some expressing worry about the technology’s potential sociological effects, such as whether AI tools would displace graduate students. Lupsasca acknowledged the issues but declined to speculate.

In a follow-up interview with Science that afternoon, Lupsasca affirmed his belief in the technology’s potential. “This year is an inflection point for math and theoretical physics … AI will do to science, at least in those corners, what it did in 2025 to coding,” he said. “I think by the end of the year, if you’re not using it, you’re going to start to be left behind.”

In a Valentine’s day panel discussion, The Art Institute of Chicago’s Giovanni Verri explained how modern methods are helping revive the colors of ancient Greece and Rome. Statues and figurines that appear colorless today have traces of ancient pigments, he noted, some of which are invisible to the naked eye. Still, “the distinct physical and optical properties of these pigments can be exploited through scientific imaging techniques to identify and map their distribution across the surface of an artwork,” he explained. “Such scientific analysis shows us not only how objects were painted, but also the visual impact that these objects had in their original setting.”

Such revelations “allow us also to move beyond decoration and into daily life,” he said. Take clothing, for example—unlike marble statues, fabrics simply don’t survive for thousands of years. Yet, researchers can find clues as to how people dressed via faded paints. “Through science, we can add the color component, and by identifying colors and patterns, we can better understand how garments were woven and worn.”

“By recovering ancient color with science, we can do more than restore lost surfaces,” Verri concluded. “We also get a glimpse of the emotional and visual power that these objects once held, and in doing so, we come closer to the people who made, viewed, and cherished these objects.”

With contributions from Michael Greshko and Tim Appenzeller

Do you have a burning science question you can’t seem to find a good answer for? Submit it to Ask Science! Selected questions will receive responses from Science editors right here in ScienceAdviser.

Want more? Catch up on past issues of ScienceAdviser.

If you were forwarded this newsletter by a friend, you can subscribe for free here.