AI chatbots beat static messages in AISI's 42,000-person test

The UK AI Security Institute surveyed more than 42,000 Britons debating political topics with 19 language models, and found the AI conversations shifted attitudes 41-52% more effectively than static messages, with an average movement of roughly 10 points on a 0-100 scale. The catch, Felix M. Simon writes in Transformer, is that the most persuasive models also generated false claims in about 30% of their responses. Two experts in our Who's Who directory have shared the piece.

Simon, a research fellow at Oxford's Reuters Institute, treats the AISI result as one data point in a wider pattern. A Nature study across the US, Canada and Poland found AI election conversations outperformed traditional video ads, and separate work by Zhongren Chen and co-authors estimates that "LLM-based persuasion costs between $48 and $75 per voter," versus about $100 for traditional methods. Personalization to demographics, the AISI work found, added little; information density did the work.

His argument is that the lab and the street are different rooms. Ben Tappin, an assistant professor of psychology at the London School of Economics, is blunt: "Most people in the real world are not going to spare 10 minutes for a very information dense political conversation with a chatbot." Simon estimates the average person has only 30-60 minutes a day for intentional information consumption, against 5.2 hours of leisure time, 2.6 of them spent on television. Any AI pitch is competing with ads, news, family and peers for that window.

Simon also draws a definitional line that shapes the regulatory read: "Persuasion is about changing people's minds by giving them reasons they can inspect (and of course reject), and is quite different from manipulation, which forgoes that process by deceiving people." The EU AI Act's Article 5(1)(a) prohibits subliminal manipulation or deceptive techniques that materially distort behavior, has applied since February 2, 2025, and carries penalties up to €35 million or 7% of annual turnover. In the United States, the FCC has ruled AI-voiced robocalls illegal.

Tappin adds one further human factor working against the persuaders: "People are very averse to being persuaded when it seems like an overt persuasion attempt."