When AI Gets Better At Persuasion, It Gets Worse At Telling The Truth

Forget Microtargeting: AI Changes Minds by Drowning People in Claims

The techniques that make artificial intelligence more effective at changing people’s minds also make it less accurate. That’s the troubling pattern that emerged from the largest study of AI persuasion ever conducted. Researchers found that when AI chatbots were optimized for persuasiveness, they systematically produced more false information, even without being instructed to deceive.

The study involved nearly 77,000 participants in the United Kingdom who engaged in political conversations with 19 different AI systems, including frontier models like GPT-4.5 and Grok-3. Professional fact-checkers and a separate AI system then evaluated more than 466,000 claims made during these conversations.

Overall, about 81% of AI-generated claims were rated as accurate. But the researchers documented a consistent pattern across models, training methods, and prompting strategies; whatever made AI more persuasive also made it less truthful.

Persuasion Works Through Sheer Volume

AI persuasion, the study found, works primarily through flooding conversations with claims. The most effective strategy was simply prompting the AI to pack its arguments with facts and evidence. When given this instruction, GPT-4o generated more than 25 fact-checkable claims per conversation on average, compared with fewer than 10 under other approaches. Persuasiveness increased by 27%.

But accuracy dropped in tandem. The March 2025 version of GPT-4o made accurate claims 78% of the time under standard conditions. When pushed to deploy more information, that figure fell to 62%. GPT-4.5, when prompted the same way, saw accuracy decline from 70% to 56%.

The study, published in Science, found that specialized training designed to increase persuasiveness produced the same pattern. When researchers applied a technique called reward modeling, which trains AI to select responses predicted to be most persuasive, the models became better at shifting opinions and worse at accuracy. False claims appear to be a byproduct of the pressure to generate more information rather than a direct mechanism of influence; when one model was explicitly instructed to fabricate claims, its persuasiveness did not increase even though accuracy declined.

Newer and Bigger Models Are Not More Truthful

Perhaps most concerning, the accuracy problem appears to be worsening at the frontier of AI development. Claims made by OpenAI’s GPT-4.5, one of the newest and largest models tested, were rated inaccurate more than 30% of the time, roughly equivalent to a much smaller model that requires a fraction of the computing resources.

GPT-3.5, released more than two years before GPT-4.5, actually performed better on accuracy, producing about 13 percentage points fewer inaccurate claims than its successor. Different versions of GPT-4o, identical in underlying scale but subjected to different post-release training by OpenAI, showed large accuracy gaps. The March 2025 version was significantly less accurate than the August 2024 version.

These findings indicate that recent changes to frontier models have not improved their truthfulness during persuasive conversations, and in some cases have reduced it.

Conversation Amplifies Both Persuasion and Misinformation

Static AI-generated messages, like a 200-word persuasive essay, produced modest effects on reader opinions. But back-and-forth conversation amplified persuasion by 40-50%, and the effects persisted. A follow-up one month later found that between 36% and 42% of the immediate persuasive impact remained.

The researchers calculated what would happen if someone combined every advantage identified in the study: the most persuasive model, the most effective prompting strategy, and specialized training techniques. Under these maximal-persuasion conditions, AI achieved a persuasive effect of about 16 percentage points on average, and 26 percentage points among participants who initially disagreed with the position being argued. The AI made an average of 22.5 fact-checkable claims per conversation, and nearly one-third were rated inaccurate.

Several factors that have dominated public discussion about AI influence turned out to have relatively small effects. Personalization, the idea that AI could tailor arguments based on user demographics or stated beliefs, produced gains of only about half a percentage point. Model size mattered, but its effects were often eclipsed by training choices. The difference between two versions of GPT-4o released months apart exceeded what the researchers estimated would come from building a model 100 times larger than current frontier systems.

Psychological techniques drawn from political science literature, including moral reframing, storytelling, and deep canvassing, underperformed a simple fact-based approach. Some strategies actually reduced persuasiveness.

Small Models Can Be Weaponized Too

The study carries a practical warning about who can deploy persuasive AI. Applying the right training techniques to a small open-source model, the kind that can run on a standard laptop, made it as persuasive as GPT-4o. Actors with limited computational resources could train and deploy highly effective persuasion tools without access to frontier technology, broadening the range of people who could use AI for political influence.

The findings point to what appears to be an inherent tension in AI development. The capability that makes AI useful for legitimate purposes, its ability to rapidly generate relevant information during conversation, appears to be the same capability that drives both persuasiveness and inaccuracy. Optimizing for one may unavoidably compromise the other.

The researchers stopped short of claiming that AI developers are deliberately building deceptive systems. But they noted that even without explicitly seeking to misinform, AI systems designed for maximum persuasion may provide substantial amounts of inaccurate information.

Whether this trade-off can be broken remains an open question. For now, the evidence points to a disquieting conclusion: the better AI gets at changing minds, the worse it gets at telling the truth.

---

Disclaimer: This article summarizes peer-reviewed research and is intended for informational purposes only. The findings reflect controlled experimental conditions and may not directly translate to real-world AI deployment. Readers seeking to understand AI policy or safety should consult the original study and additional expert sources.

---