Most worry about AI and belief has focused on sycophancy (the model agreeing with you) and over-trust (you agreeing with the model without checking). A third question has now been tested at scale: if a chatbot sets out to move your opinion, does it work?
The answer from several large experiments is yes, by a measurable amount, in both directions. The same technique that talks people out of conspiracy theories can talk them into things that are not true. What is still missing is evidence from outside the laboratory, where nobody is paid to sit through the conversation.
The good-news study
In 2024, researchers at MIT, Cornell and American University published an experiment in Science that upended a common assumption about conspiracy beliefs, namely that facts do not shift them. They engaged 2190 conspiracy believers in personalized evidence-based dialogues with GPT-4 Turbo. Each participant described a conspiracy theory they believed and their reasons, and the model argued back with specific counter-evidence over a few rounds.
The intervention reduced conspiracy belief by about 20%. More surprising was the durability: the effect remained 2 months later, generalized across a wide range of conspiracy theories, and occurred even among participants with deeply entrenched beliefs.
The researchers’ interpretation was that people are more responsive to evidence than the psychology literature had assumed, provided the evidence addresses their particular reasons. A chatbot can do that tirelessly and without getting irritated, which most humans arguing about the moon landing cannot.
The same lever, pointed at politics
In December 2025, two linked papers tested whether the effect extends to political opinions.
The Science paper was large by any standard. In three large-scale experiments (N=76,977), the team deployed 19 language models to argue with participants on 707 political issues, and then checked the factual accuracy of 466,769 resulting LLM claims.
Its findings cut against two popular fears and confirmed a third.
Bigger models were not the main driver, and neither was personalisation. The persuasive power came more from post-training and prompting methods, which boosted persuasiveness by as much as 51% and 27% respectively. In practice that means a modest model, tuned and instructed to persuade, can outperform a frontier model that has not been.
The mechanism was information. As one summary of the paper put it, the most persuasive chatbots are those that share lots of facts. Models persuaded by producing a dense stream of relevant-sounding claims, faster than any person could.
And that is where the third fear was confirmed. The authors report that where these methods increased AI persuasiveness they also systematically decreased factual accuracy. The same summary notes that the most information-dense bots also dole out the most inaccurate claims. Optimising a model to persuade made it both more convincing and less truthful.
The companion paper in Nature applied this to real elections. According to Scientific American, the team recruited more than 2,000 U.S. adults and asked them to rate their candidate preference on a scale of 0 to 100, then had them talk with a chatbot arguing for one candidate. Preferences moved after a single conversation.
Does personalisation matter or not?
Here the literature disagrees with itself, and the disagreement is informative.
A 2025 study in Nature Human Behaviour set 900 people to debate either another person or GPT-4. When the opponent was given basic demographic details about the participant, GPT-4 with personalization was more persuasive 64.4% of the time in pairs where one side was clearly more convincing. Without that information, a summary of the study reports, GPT-4’s persuasive capabilities were indistinguishable from those of humans.
The much larger Science study found personalisation had a smaller effect than tuning and prompting. The two are not strictly in conflict: one compared AI with human debaters, the other compared AI configurations with each other. A fair reading is that personal data gives a model an edge over a human, but it is not the biggest lever available to someone building a persuasion machine.
The experiment run in reverse
The most uncomfortable development came from the team behind the original conspiracy study. As Science reported in 2026, the authors teamed up with AI safety researchers and turned their original experiment on its head. In the new study, only posted as a preprint so far, they found that an AI could talk people into conspiracy theories.
That work has not been peer reviewed, and preprints sometimes change. But it fits the accuracy finding above: the technique is neutral about truth. What moved people in the original study was a rapid supply of tailored, confident, specific claims. That works whether or not the claims are correct.
What is not known
Whether any of this happens in the wild. Every study above paid participants to complete a conversation with a bot that had been instructed to persuade them. Real people can close the tab. How much persuasion survives when nobody is obliged to keep reading has not been measured.
How long political shifts last. The conspiracy effect held for two months. Durability for political attitudes is less established.
Whether ordinary assistants do this by accident. The experiments used models told to persuade. A general-purpose assistant answering a question is not in that mode, and no study has shown that everyday use shifts political views. It has not been ruled out either.
Whether knowing helps. It is unclear whether being aware of these findings makes anyone more resistant.
What follows for someone who uses AI all day
Any advice here is inference from the mechanism, not a tested intervention, and should be read that way.
The mechanism is volume of plausible claims. The accuracy finding says the most convincing outputs contained the most errors. So the feeling of being overwhelmed by good points is not evidence of being right, and may be weak evidence of the opposite. When a model gives you ten reasons, check two of them.
Notice who configured the bot. The strongest lever in the research was post-training and prompting, which is invisible to the user. A chatbot on a campaign site, a sales page or an advocacy group’s app may be tuned for persuasion in a way a general assistant is not.
Ask for the other side. A model that produced a persuasive case for one position will produce an equally fluent case for the opposite if asked. Reading both is the cheapest available check on fluency posing as truth.
Common questions
Can AI chatbots change people’s political opinions?
In controlled experiments, yes. A study of 76,977 participants and 19 language models across 707 political issues found that conversations shifted attitudes, and a companion election study found chatbot conversations moved candidate preferences among more than 2,000 US adults.
Are AI chatbots more persuasive than humans?
With personal information, they can be. In a 900-person debate study, GPT-4 given demographic details about its opponent was more persuasive 64.4% of the time in pairs where the two sides differed. Without that information it performed about the same as human debaters.
Can AI talk people out of conspiracy theories?
A 2024 Science study of 2190 believers found personalised dialogues with GPT-4 Turbo reduced conspiracy belief by about 20%, and the effect remained two months later. A follow-up preprint by overlapping authors found an AI could also talk people into conspiracy theories.
Does persuasive AI tell the truth?
Less often as it gets more persuasive. After checking 466,769 claims, researchers found that the methods which increased AI persuasiveness also systematically decreased factual accuracy.
The bottom line
A short conversation with a language model can move what people believe, mostly by supplying a large number of specific claims quickly, and the setups that persuade best are the ones that get the most facts wrong. None of this has been measured outside paid experiments, so the real-world size of the problem is unknown. Until it is, treat a flood of fluent arguments as a reason to check, and be most sceptical when the bot belongs to someone with something to sell.
Did this answer your question?
Thanks. That helps us decide what to fix.
Leave a comment