Artificial intelligence chatbots are very good at changing peoples’ political opinions, according to a study published Thursday, and are particularly persuasive when they use inaccurate information.
The researchers used a crowd-sourcing website to find nearly 77,000 people to participate in the study and paid them to interact with various AI chatbots, including some using AI models from OpenAI, Meta and xAI. The researchers asked for people’s views on a variety of political topics, such as taxes and immigration, and then, regardless of whether the participant was conservative or liberal, a chatbot tried to change their mind to an opposing view.
The researchers found not only that the AI chatbots often succeeded, but also that some persuasion strategies worked better than others.
“Our results demonstrate the remarkable persuasive power of conversational AI systems on political issues,” lead author Kobi Hackenburg, a doctoral student at the University of Oxford, said in a statement about the study.
The study is part of a growing body of research into how AI could affect politics and democracy, and it comes as politicians, foreign governments and others are trying to figure out how they can apply AI to sway public opinion.
The paper, published in the journal Science, found that AI chatbots were most persuasive when they provided study participants with large amounts of in-depth information, rather than when they deployed alternate debating tactics such as appeals to morality or arguments personalized to the individual.
The implication, according to researchers, is that AI chatbots could “exceed the persuasiveness of even elite human persuaders, given their unique ability to generate large quantities of information almost instantaneously during conversation,” the researchers wrote, although they did not put AI chatbots head-to-head against human debaters.
But the study also said that the persuasiveness of AI chatbots wasn’t entirely on the up-and-up: Within the reams of information the chatbots provided as answers, researchers wrote that they discovered many inaccurate assertions.
“The most persuasive models and prompting strategies tended to produce the least accurate information,” the researchers wrote.
They added that they observed “a concerning decline in the accuracy of persuasive claims generated by the most recent and largest frontier models.” Claims made by GPT-4.5 — a model released by OpenAI in February — were significantly less accurate on average than claims from smaller, older models also from OpenAI, they wrote.
“Taken together, these results suggest that optimizing persuasiveness may come at some cost to truthfulness, a dynamic that could have malign consequences for public discourse and the information ecosystem,” they wrote.
OpenAI did not immediately respond to a request for comment Thursday. The study was conducted before OpenAI released its latest model, GPT-5.1, last month.
About 19% of all claims by the AI chatbots in the study were rated as “predominantly inaccurate,” the researchers wrote.
The paper warned that, in an extreme scenario, a highly persuasive AI chatbot “could benefit unscrupulous actors wishing, for example, to promote radical political or religious ideologies or foment political unrest among geopolitical adversaries.”
The paper was produced by a team of people from the U.K.-based AI Security Institute, a research organization backed by the British government, as well as from the University of Oxford, the London School of Economics, Stanford University and the Massachusetts Institute of Technology. They received funding from the British government’s Department for Science, Innovation and Technology.
Helen Margetts, a professor of society and the internet at Oxford and a co-author, said in a statement that the research was part of an effort “to understand the real-world effects of LLMs on democratic processes,” referring to large language models, or the technology underlying popular generative AI systems like ChatGPT and Google Gemini.
The study participants were all adults in the United Kingdom, and they were asked questions related to British politics.
The study lands at a time when AI is upending politics from a variety of angles. President Donald Trump has regularly posted AI-created videos and images to social media, some political campaigns have sent out AI-generated fundraising emails or deepfake videos and state operatives from China and Russia have deployed AI “slop” in their propaganda campaigns.
And at the same time, AI chatbot use is growing quickly. About 44% of U.S. adults surveyed said that they used AI tools like ChatGPT, Google Gemini or Microsoft Copilot “sometimes” or “very often,” according to an NBC News Decision Desk Poll powered by SurveyMonkey and released in June.
The study found that there’s something particularly persuasive in the back-and-forth conversation of chatbot apps. The researchers compared study participants who interacted with an AI chatbot with participants who were merely asked to read a 200-word persuasive message written by AI. They found that the AI chatbot was substantially more persuasive than the static message — between 41% and 52% more persuasive, depending on the AI model.
The change in political views was durable, the study found, with between 36% and 42% of the persuasive effect still evident among participants one month later.
The researchers also explored which kinds of AI chatbots were most effective, testing 17 different large language models including the most sophisticated models and simpler ones. They also tested models with various levels of fine-tuning, known in the AI industry as “post-training.”
The researchers acknowledged that the controlled conditions of their study didn’t immediately translate to the daily grind of politics, where most campaigns are not deploying chatbots as a tool for persuasion.
“The extent to which people will voluntarily sustain cognitively demanding political discussions with AI systems outside of a survey context remains unclear,” they wrote.
But some experts who were not involved in the research said the study is an important step.
“Now we have evidence showing that as models get better, they are becoming more persuasive,” said Shelby Grossman, a professor in the journalism school at Arizona State University who has studied the persuasiveness of AI content.
She said it’s conceivable that foreign governments could try to leverage the persuasiveness of AI content, including to try to sow division on social media, but she also said there were legitimate uses, too, if political actors are transparent about their use of AI.
David Broockman, an associate professor of political science at the University of California, Berkeley, who studies persuasion in campaigns, said that although the study found AI chatbots to be persuasive, he was reassured that the effect was not larger.
“There are these doomsday scenarios in the world that say AI is going to hypnotize or brainwash us because it’s so much more persuasive than a human,” he said.
He said the study rebuts those doomsday ideas and instead suggests that what humans find persuasive is large volumes of detailed information provided on demand — a positive sign for humanity, he said.
In a real-world scenario, “if you’ve got both sides of an issue using this, I would guess it would cancel out and you’re going to hear more persuasive arguments on both sides,” he said.
The study released Thursday is not the first to try to measure the persuasiveness of AI. Another paper published in May based on a study of a smaller group of participants found that AI chatbots were not especially persuasive, while a paper published last year found that humans working with generative AI could produce persuasive propaganda with limited effort.
