Today, we published in Science the results of a study carried out with colleagues at the Oxford Internet Institute, the London School of Economics, Stanford University and MIT, examining how conversational AI can shape political attitudes. Through three large‑scale experiments with over 76,000 participants, we tested the persuasiveness of 19 AI models on more than 700 political issues.
Our goal was to understand the levers of persuasion with conversational AI: what makes it effective, and under what conditions. We were interested in answering questions like: Is persuasion mainly driven by model size? Do personalisation and microtargeting matter? Can models be post‑trained to become more persuasive? Which rhetorical strategies are most effective?
Why study AI-enabled persuasion?
Conversational AI systems can now generate detailed, well‑structured arguments instantly and hold interactive discussions that feel tailored and engaging. While this creates opportunities for useful applications, it also raises the possibility that AI could influence what people think and do.
Although there is currently little evidence that such systems are being used to maliciously persuade people at scale, that may change as the technology advances. Understanding the mechanisms that make AI persuasive allows us to identify where genuine risks lie, rather than relying on assumptions or speculation. This evidence is essential for designing safeguards and standards that keep people safe.
Our experimental set-up
Across three experiments, participants engaged in back‑and‑forth conversations with one of 19 open‑ and closed‑source language models. In controlled conditions, models were instructed to persuade the participant to agree with one of 707 issue stances, using one of eight different rhetorical strategies. These strategies included information-focused argumentation, storytelling, and moral reframing.
Participants rated their agreement with an issue both before and after the conversation. We measured persuasion as the difference in average post‑conversation opinion relative to a control group that had no persuasive conversation.
Our key findings
Across the three experiments, we identified three main findings.
1. Post-training matters more than model scale
Larger models were more persuasive when post-training was held constant, but these gains were modest compared to those offered by post-training on validated examples of persuasive conversations. For example, our persuasion reward modelling boosted the persuasiveness of a small open-source model enough to match or exceed much larger frontier models. Further demonstrating the importance of post-training, we observe that the difference in persuasiveness between two versions of the same frontier model released seven months apart (same scale, different post-training) exceeded what our statistical models predict from even a 100× increase in pre-training compute.
2. Information density drives persuasion more than personalisation or “microtargeting”
Despite widespread concern about AI-powered microtargeting, we observed personalisation effects which were consistently small (less than one percentage point). What predicted larger persuasion gains was information density: the sheer volume of fact-checkable claims the model deployed. Prompting models to emphasise facts and evidence increased persuasiveness by 27% relative to a basic "be persuasive" prompt, which was a larger increase than any of the other strategies we tested (this included prevailing strategies from persuasion literature, such as moral reframing, storytelling, and deep canvassing).
3. More persuasive models made more inaccurate claims
We consistently found across our experiments that the factors that increased model persuasiveness, such as information-focused prompting and persuasion reward modelling, systematically decreased factual accuracy. This implies that optimising models for persuasion may come at a cost to truthfulness.
Importantly, however, our findings do not imply that inaccurate claims are more persuasive than accurate ones. One possible explanation for this correlation is that as models become more information-dense in their responses, they eventually exhaust the supply of reliably accurate facts and begin to generate weaker or less accurate material. Exploring this relationship further is an important aim of our future research.
What does this all mean?
Our findings show that conversational AI can be a persuasive tool under controlled conditions. But importantly, the strongest levers of persuasion are not model scale or personalisation - they are post‑training and prompting techniques, which significantly shape how much information models present and how they structure arguments.
The same techniques that increase persuasiveness also reduce factual accuracy, highlighting the importance of robust evaluation standards, transparency around model training, and careful consideration of any optimisation that might affect influence or truthfulness.
Looking ahead, persuasive capabilities are likely to evolve both through the development of larger models, but perhaps more significantly through enhanced post‑training. At the same time, real‑world impacts remain uncertain: the most effective persuasion in our study occurred during sustained, information‑dense, multi‑turn discussions, and the extent to which people will engage in such conversations voluntarily outside of a survey experiment remains unclear.
These results help clarify which levers of AI persuasion deserve particular attention from developers and policymakers. They also underline the need to monitor persuasion‑related behaviours in future frontier models, ensure accuracy remains a priority, and build safeguards that protect users and democratic processes. As AI capabilities advance, continuing to study and evaluate persuasive potential will be a critical part of ensuring safe and trustworthy deployment.
