In controlled experiments, AI systems change people's political opinions more effectively than financially motivated human persuaders. AI chatbots shift voter preferences by roughly four times the effect of traditional political advertisements. The largest study to date (77,000 participants, 19 AI models) found that the techniques which make AI more persuasive also make it less accurate.
Why this matters
Persuasion has always required human effort. Writing a compelling argument, making a phone call, knocking on a door. AI removes the effort. A single person can now generate personalised persuasive messages for every registered voter in a country for less than the cost of a single television advertisement.
The capability is not just theoretical. AI-generated persuasive content is already standard infrastructure in digital marketing. Political campaigns in the US, India, and Europe deployed AI chatbots to engage voters in the 2024 election cycle. The technology is moving from experimental to operational.
The most important finding may be the tradeoff: the techniques that make AI more persuasive systematically make it less accurate. Post-training can boost persuasiveness by 51%, but where persuasiveness goes up, factual accuracy goes down. This means the most persuasive AI messages are also the least truthful ones.
Documented incidents
Evidence timeline
Discussed in Theory
Coined "captology" (Computers As Persuasive Technologies). His doctoral thesis demonstrated experimentally that computers can change people's thoughts and behaviours in predictable ways. Machines do not need to understand persuasion to be persuasive.
Fogg, Stanford Persuasive Technology Lab →Generative Language Models and Automated Influence Operations. First rigorous analysis of how LLMs could automate influence operations at scale. Published by OpenAI researchers alongside security researchers, making it an industry self-assessment of the risk.
Goldstein et al., Georgetown CSET + OpenAI + Stanford →Training a Helpful and Harmless Assistant with RLHF. Demonstrated that optimising for helpfulness produces models that are more persuasive as a side effect. Persuasive capability is not a deliberate design choice but an emergent property of training models to be useful.
Bai et al. (Anthropic), arXiv 2022 →Demonstrated in Lab
Artificial Intelligence Can Persuade Humans on Political Issues. N=4,836 across 3 pre-registered experiments. GPT-3 messages were as persuasive as messages crafted by lay humans on policy issues including assault weapon bans, carbon tax, and paid parental leave. First large-scale demonstration that LLMs match human persuasion.
Bai et al., Nature Communications (2025) →2,306 US citizens in pre-registered experiments during the 2024 presidential election. AI chatbots shifted voter preferences by 3.9 points on a 100-point scale, roughly 4x the effect of traditional political advertisements. When the model was prevented from using facts, persuasion dropped dramatically.
Lin et al., Nature (Dec 2025) →The largest study of AI persuasion ever conducted. ~77,000 UK participants, ~91,000 dialogues, 19 LLMs, 707 political issues. Post-training boosted persuasiveness up to 51%. Critical finding: "where they increased AI persuasiveness, they also systematically decreased factual accuracy."
Hackenburg et al., Science (Dec 2025) →Demonstrated in Real World
AI chatbots deployed in political campaigns. Shamaine Daniels (US congressional candidate) used AI chatbot "Ashley" to call and persuade voters. India spent tens of millions on AI voter targeting. An Italian candidate deployed a chatbot in the 2024 EU Parliament elections.
MIT Technology Review / Newsweek →AI-generated persuasive content is now standard marketing infrastructure. Meta's Advantage+ AI advertising tools are deployed across the platform. Industry surveys report a majority of marketers using AI tools for customer-facing copy, with significant volume of online persuasive content now AI-generated or AI-assisted. (Specific platform-wide ROI percentages remain vendor-reported and are not independently verified at the time of writing.)
Meta Advantage+ documentation / industry adoption →In real-time interactive conversations, Claude Sonnet 3.5 achieved significantly higher compliance with persuasion attempts than financially incentivized human persuaders. AI was more persuasive in both truthful and deceptive contexts. The threshold has crossed from "AI matches humans" to "AI exceeds motivated humans." (Pre-print)
Schoenegger et al., arXiv 2025 →Strongest Counterargument
Model persuasiveness shows sharply diminishing returns with scale. Current frontier models are only slightly more persuasive than models an order of magnitude smaller. The effect disappears once you control for mere task completion (coherence, staying on topic). Since frontier models are at ceiling on task completion, further scaling may not increase single-message persuasiveness.
Source: Hackenburg et al. (PNAS, Feb 2025). N=25,982 participants, 720 messages from 24 LLMs.
Why this deserves weight: The PNAS study (n=25,982) is substantially more powered than the Nature voter preference study (n=2,306), and comes from the same lead author as the Science paper demonstrating AI persuasion. For static, single-message persuasion, scaling does not proportionally increase persuasiveness. However, this specifically applies to single-message contexts; the Science and Nature studies show conversational, multi-turn AI persuasion scales differently.