A 2025 Stanford study gave 4,829 Americans AI-generated arguments supporting six different policies and found that participants became modestly more supportive after reading them, with AI messages appearing as persuasive as human-written arguments and often perceived as more logical and informed

Large language models are no longer simply tools for writing emails, summarising documents or generating code. A new study suggests they can also influence how people think about contentious political issues, raising fresh questions about the role of artificial intelligence in elections, political campaigns and public debate. Researchers from Stanford University and other institutions conducted three preregistered experiments involving 4,829 Americans and found that people exposed to persuasive messages generated by large language models (LLMs) subsequently changed their attitudes toward a range of public policies. The effects were relatively small, but they were consistent - and, importantly, LLM-generated messages were about as persuasive as messages written by ordinary people. The research, published in Nature Communications in 2025, examined whether publicly available LLMs could generate original political arguments capable of shifting people's policy preferences.

RESEARCHERS TESTED AI PERSUASION ON SMOKING BANS, GUN CONTROL AND TAXES

The experiments were designed to test whether AI-generated arguments could move people's opinions rather than simply asking participants whether they thought AI-written messages were persuasive. Participants reported their support for a policy before and after reading a message. In the first study, participants encountered arguments concerning a public smoking ban. The second examined the highly polarised issue of an assault weapons ban. A third experiment broadened the test to four additional policies: a carbon tax, an increase to the child tax credit, paid parental leave and automatic voter registration. Some participants read messages generated by GPT-3, while others read messages written by human participants. Another group read a neutral message unrelated to the policy issue, providing the control condition. In some experiments, humans were also allowed to select the most persuasive message from a set of AI-generated arguments.


AI MESSAGES CONSISTENTLY SHIFTED POLICY ATTITUDES

Across all three experiments, participants exposed to LLM-generated arguments became more supportive of the policies being advocated. The researchers describe the effects as small but statistically significant, with changes generally ranging from about 2 to 4 points on a 101-point policy-support scale. For example, in the smoking-ban experiment, participants who read an AI-generated persuasive message increased their support by about 3.6 points compared with the control group. The effect was also observed on the much more politically polarised issue of assault-weapons bans, where exposure to an LLM-generated message produced an increase of about 1.8 points compared with the control condition. In the third experiment, which examined multiple policies, the AI-generated messages again produced a significant shift in policy attitudes. The findings therefore do not suggest that an AI-generated paragraph can suddenly transform a person's political ideology. Rather, they show that even relatively modest exposure to an AI-generated argument can measurably move people's stated policy preferences.

AI WAS ABOUT AS PERSUASIVE AS AN ORDINARY HUMAN

Perhaps one of the study's most striking findings was the comparison between AI-generated and human-written arguments. In the first two experiments, researchers found no significant difference in the persuasive effects of LLM-generated messages and messages written by lay humans. In other words, GPT-3 did not need to outperform humans to produce a potentially important result. It appeared capable of reaching roughly the same level of persuasion as an ordinary person writing an argument. The researchers note that this comparison involved laypeople rather than professional political campaigners or communication experts. More sophisticated human political messaging could potentially be more effective than AI-generated arguments. Still, the finding establishes an important benchmark: LLMs had already reached a level where their political arguments could influence people at approximately the same magnitude as those produced by everyday humans.

MOST PEOPLE DID NOT REALISE THE MESSAGES CAME FROM AI

The study also uncovered a potentially more concerning feature of AI persuasion: people generally failed to recognise that the messages had been generated by artificial intelligence. In the second experiment, participants were asked who they thought had written the messages. Across the experimental conditions, approximately 94% of participants who had actually read LLM-generated messages believed they had been written by a human rather than an AI program. That matters because persuasion can operate differently when people know they are interacting with a machine.

The participants also perceived the apparent authors of AI-generated messages as better informed and more logical, and less angry, compared with the authors of human-generated messages. Human-written arguments, meanwhile, were perceived as more unique and more likely to use vivid storytelling.

AI'S ADVANTAGE APPEARED TO BE FACTS, LOGIC AND A DISPASSIONATE TONE

The researchers found evidence that AI and human messages may persuade through somewhat different mechanisms. LLM-generated arguments were associated with perceptions that their authors relied more heavily on facts, evidence and logical reasoning, while also maintaining a less emotional tone. Human-generated arguments appeared to benefit more from perceptions of uniqueness and originality. The language analysis supported some of these differences. AI-generated messages contained a greater proportion of longer words, were generally longer, used more analytical language and had higher average reading levels. Human messages were more likely to use first-person language and personal appeals. This suggests that AI may have a particular advantage when persuasion depends on presenting a structured, evidence-oriented argument rather than a deeply personal story.

THE FINDINGS RAISE A BIGGER QUESTION ABOUT AI AND DEMOCRACY

The researchers say the significance of their findings extends beyond whether ChatGPT can write a convincing political paragraph. An individual campaign worker could use an LLM to produce speeches, slogans, canvassing scripts or political advertisements much faster than before. More significantly, AI could potentially generate large numbers of individually tailored political messages for different audiences. At massive scale, that capability could change the information environment surrounding elections. The researchers warn that LLMs could potentially be used to impersonate large numbers of people online, generate seemingly authentic messages to politicians, flood social-media discussions or produce targeted political communications. Such activity could make it difficult to distinguish genuine public opinion from artificially generated political activity. They stress, however, that the experiments did not test whether AI-generated misinformation could persuade people. The study tested persuasive messages advocating policies, not deliberately false claims.

WHY DISCLOSURE AND AI LITERACY COULD BECOME INCREASINGLY IMPORTANT

The researchers argue that the findings should encourage policymakers to take AI-generated political persuasion seriously. Potential safeguards discussed in the study include disclosing AI-generated political content, digital watermarking, AI-based detection systems, stronger safeguards around generating misleading political messages and greater public education about AI-generated content. The challenge is that the technology can produce persuasive communication cheaply and rapidly - and people may not reliably recognise its source.

The study ultimately points to a shift that could become increasingly important as AI systems improve: political persuasion may no longer require a human campaigner to write every argument, or even a human author to produce the apparent voice behind it. The research does not show that AI can manipulate an entire electorate or instantly change deeply held political beliefs. But it does provide experimental evidence that AI-generated arguments can move people's policy attitudes - and can do so at a scale and speed that traditional human-written political communication may struggle to match.