TLDR: A large-scale study involving 76,977 participants and 19 LLMs reveals that the persuasive power of conversational AI stems more from specialized post-training and prompting methods than from model scale or personalization. AI is most persuasive when it provides a high volume of information, but this often comes at the cost of factual accuracy. The research highlights a troubling trade-off between persuasiveness and truthfulness, suggesting that future AI influence will depend heavily on how models are trained and instructed, rather than just their size.
Concerns about the potential for conversational AI to influence human beliefs are widespread. A recent large-scale study, detailed in the paper “The Levers of Political Persuasion with Conversational AI”, delves into what truly makes these AI models persuasive and the implications for factual accuracy.
Unpacking AI’s Persuasive Power
The research, conducted by Kobi Hackenburg, Ben M. Tappin, Luke Hewitt, Ed Saunders, Sid Black, Hause Lin, Catherine Fist, Helen Margetts, David G. Rand, and Christopher Summerfield, involved an extensive series of experiments. They engaged 76,977 participants in conversations with 19 different large language models (LLMs) across 707 political issues. A crucial part of their methodology also involved fact-checking over 466,000 claims made by these LLMs.
Contrary to some popular assumptions, the study found that the sheer size or ‘scale’ of an AI model isn’t the primary driver of its persuasive ability. While larger models did show some increase in persuasiveness, the most significant gains came from specific post-training methods and the way the AI was prompted to engage in conversation.
The Impact of Training and Prompting
The researchers discovered that post-training techniques, which explicitly optimize models for persuasion, could boost their effectiveness by as much as 51%. Prompting methods, which guide the AI on how to formulate its arguments, also significantly increased persuasiveness by up to 27%. This suggests that how an AI is refined after its initial development, and the instructions it receives for a conversation, are far more critical than simply making the model bigger.
For instance, a technique called ‘Reward Modeling’ (RM), where an AI learns to select responses that are predicted to be most persuasive, proved highly effective. Applying RM to a smaller, open-source model like Llama3.1-8B made it as persuasive, or even more so, than much larger and more expensive frontier models like GPT-4o (older version).
Information is Key, But at What Cost?
The study identified that AI models were most persuasive when they adopted an “information-based argumentation” strategy. This involved packing their dialogue with a high volume of fact-checkable claims and evidence. The more information an AI provided, the more persuasive it tended to be. This finding stands in contrast to other human-centric persuasion strategies like personalization, storytelling, or moral reframing, which showed much smaller persuasive effects when employed by AI.
However, this reliance on information density comes with a significant caveat: accuracy. The research uncovered a troubling trade-off: factors that increased an AI’s persuasiveness, such as specific prompting strategies or post-training, often systematically decreased the factual accuracy of the claims it made. For example, when prompted to be highly informative, models like GPT-4o (March 2025 version) and GPT-4.5 showed notable drops in accuracy. Some of the most recent and largest frontier models also exhibited a concerning decline in the accuracy of their persuasive claims compared to older or smaller models.
While the overall accuracy of AI-generated claims was reasonably high (77 out of 100 on average), the trend suggests that optimizing for persuasion might compromise truthfulness, potentially leading to negative consequences for public discourse.
Conversation’s Unique Edge
The study also confirmed that engaging in a multi-turn conversation with an AI is significantly more persuasive than simply reading a static, AI-generated message. This highlights the unique power of interactive dialogue in influencing human beliefs, a finding that could reshape how AI is deployed for communication and influence.
Also Read:
- Unlocking LLM Potential: A New Approach to Positional Bias
- Beyond Feelings: Why Observed Emotions May Not Predict What We Remember
Looking Ahead
The findings suggest that AI’s persuasive capabilities are likely to continue growing, driven primarily by advancements in post-training and prompting techniques rather than just model scale or personalization. This raises important questions about who controls these powerful techniques and how they might be used. The potential for even smaller, open-source models to become highly persuasive through targeted training means that the ability to deploy influential AI could become more widespread. Ensuring responsible development and deployment of these increasingly persuasive AI systems will be a critical challenge for researchers, policymakers, and technologists alike.
