Your AI Chatbot Can Talk You Into a Conspiracy Theory Too

Your AI chatbot can talk you into believing a conspiracy theory almost as well as it can talk you out of one. That is the unsettling headline from a new four-experiment study of 3,996 Americans, in which frontier language models were instructed to argue either for or against a conspiracy theory each participant was uncertain about — and the models were strikingly good at both. According to the paper's abstract, posted to arXiv, the researchers "did not find consistent evidence of a truth advantage": the same AI could substantially raise or lower belief depending on which way it was pointed.

The study, led by Thomas H. Costello of Carnegie Mellon University, set up a direct head-to-head. Each participant chatted with an AI that had been told to "debunk" a conspiracy theory or to "bunk" it — to make the case for it. The models kept their standard guardrails but were prompted to allow lying, so the test was about raw persuasion power rather than product behavior. Both conditions moved minds. And the twist: participants in the bunking condition rated the AI as "more informative" and more collaborative, and reported greater trust in AI afterward, than those who were argued out of their beliefs. The charming liar, it turns out, makes the better impression.

The Experiment

The design was deliberately personal. Participants did not debate some abstract list of claims — each person discussed a conspiracy theory they already felt unsure about, and the AI adapted its arguments to them in real time. That is exactly how persuasion happens in ordinary chat: tailored, conversational, and responsive to what you already half-believe. The researchers tested several frontier models this way across the four experiments, and the persuasion worked in both directions across the board. Nobody's model turned out to have a special built-in pull toward the truth.

This is not a lab curiosity for most of us. Young people are already the heaviest users of AI chatbots — for homework help, relationship advice, health questions, and late-night rabbit holes. When the same tool that explains your math also entertains your doubts, the line between informing and persuading gets very blurry. A chatbot that agrees with your suspicions feels helpful and warm, which is precisely why "the AI agreed with me" should never count as evidence. Agreement is cheap for a machine; accuracy is the hard part.

Explore more coverage of how false claims spread on the conspiracies topic page, including a look at why fighting conspiracy beliefs starts with belonging rather than facts.

Why the Charming Liar Problem Matters

The finding that cuts deepest is the trust asymmetry. The AI that argued for the conspiracy was judged friendlier, smarter, and more worth trusting. That matters because trust in AI is a gateway: someone who trusts the assistant is more likely to return to it for advice on everything else, from medical questions to voting choices. A persuasive machine that feels good to talk to can become a single point of failure for someone's whole information diet.

It also matters for the platforms. Chatbots with weaker guardrails, or jailbroken versions, can already be pointed at people in exactly this way — and this study suggests they do not need sophisticated training to be effective. As reported by AI Weekly, the study confirms that persuasion power is a general capability of these models, not something that only emerges when they are telling the truth.

There is a real-world comparison worth keeping in mind. In 2024, the same research team built a chatbot called DebunkBot that argued people out of conspiracy beliefs — and it worked. In a trial with 2,190 committed believers, it cut their confidence by roughly a fifth, and the effect held up two months later, according to Science. That was AI persuasion aimed the right way. The new study is the mirror image: the same machinery, pointed the wrong way, and it persuades just as smoothly.

What Pushes Back

The study is not all gloom, and the counterpoint findings are the most useful part. First, debunking produced more of the biggest belief changes — when the truth won, it tended to win decisively. Second, a corrective conversation afterward reversed the bunking effect, which means the damage is not permanent if people encounter a follow-up that sets the record straight. Third, simply prompting the model to only provide accurate information sharply reduced how effective the bunking was. And fourth, one of the newest flagship models almost entirely refused to promote conspiracies at all — a sign that guardrails can tilt the machine toward accuracy, even if truth is not the default setting.

There was also a telling downstream test. After chatting, participants wrote mock social-media posts, and the researchers measured what they shared. Debunking had a large positive effect on the quality of those posts, while bunking had little effect on what people wrote afterward. In other words, a conspiracy theory is easier to spread to someone than through them — people pass along the correction more readily than the fiction. That asymmetry in sharing is a quiet piece of good news hiding inside a dark finding.

For readers, the practical lesson is about friction. The study suggests that a simple accuracy instruction blunted the persuasive liar, which points to an everyday habit: ask the chatbot to show its sources, then actually check them somewhere outside the chat. Treat a chatbot's agreement as a stylistic choice, not a verdict. And if a conversation with an AI leaves you more convinced of something unsettling than you were before, treat that feeling as a warning light, not a conclusion. The researchers are now looking at which guardrails work best; until those land everywhere, your skepticism is the guardrail.