A new study suggests that personalization may be the key factor that makes an AI chatbot unusually persuasive. In controlled debate settings, GPT-4 performed far better when it could tailor its arguments using background information about the person it was trying to convince.
The finding matters because online persuasion is not only about the strength of an argument. It is also about timing, framing, and whether a message feels relevant to the person receiving it. The study points to a future in which large language models could use even limited personal context to shape arguments more effectively than another human debater.
What the researchers tested
Researchers from the École Polytechnique Fédérale de Lausanne (EPFL) and the Italian Fondazione Bruno Kessler compared the persuasive power of large language models with that of human participants. The study focused on direct debate situations involving controversial topics.
Participants were randomly assigned to different groups. The research team then tested four debate setups:
- Human vs. human
- Human vs. AI
- Human vs. human with personalization
- Human vs. personalized AI model
In the personalization conditions, debaters had access to anonymized background information about their opponents. That detail is central to the study because it allowed the researchers to compare not only humans against AI, but also generic persuasion against tailored persuasion.
The AI model in the study was GPT-4. The researchers looked at whether participants became more aligned with their opponents' arguments after the debate, using the different debate formats to isolate the effect of personalization.
The striking result
The largest effect appeared when GPT-4 had access to personal information. In that condition, GPT-4 increased participants' agreement with their opponents' arguments by 81.7 percent compared to debates between humans.
Without personalization, GPT-4 still had an advantage over humans, but the result was smaller. The source reports that GPT-4's advantage without personalization was 21.3 percent, and that this result was not statistically significant.
That contrast is the main takeaway. The study does not simply suggest that a large language model can argue well. It suggests that the model becomes much more effective when it can adapt its case to the person on the other side of the exchange.
The researchers attribute the personalized AI's advantage to the way the language model used profile information. With that context, it could produce arguments that were more tailored to the participant, rather than relying on a general appeal that might work for a broad audience.
Why personalization changes the risk
The researchers found the result troubling because the personalization used in the study was based only on rudimentary background data. Even with that limited input, GPT-4's persuasiveness became significantly stronger.
The concern is that real online environments may offer far richer personal signals. The source notes that malicious actors could build more detailed user profiles from digital traces, including social media activity or purchasing behavior. Those profiles could then be used to strengthen AI-driven persuasion.
This is not only a technical issue. It is also a platform issue. If AI chatbots can tailor arguments at scale, sensitive online environments such as social media may become more vulnerable to targeted persuasion strategies.
The study suggests that platform operators should take steps to counter the spread of AI-driven persuasion strategies. One possible response identified by the researchers is to use similarly personalized AI systems that answer misinformation with fact-based counterarguments.
That proposal reflects a difficult tension. The same kind of personalization that may make manipulative persuasion more effective could also make corrective information more relevant to the person receiving it. The source does not claim this would solve the problem, but it presents it as one possible countermeasure.
Important limits of the study
The study also has limits that matter when interpreting the findings. Participants were randomly assigned to pro or con positions, regardless of their prior opinions. That setup helps structure the experiment, but it differs from many real debates, where people often choose positions based on existing beliefs.
The debates also followed a predetermined structure. Online discussions are often more spontaneous, less orderly, and shaped by many surrounding signals. The source notes that this difference limits how directly the findings can be mapped onto real-world online exchanges.
Time was another limitation. The debate format included a time limit, which could restrict creativity and persuasiveness. This may have mattered especially in the personalization condition, where human participants had to process additional information before using it effectively.
These caveats do not erase the core result. They do, however, show why the findings should be read as evidence from a controlled setting rather than as a complete model of every online conversation.
What comes next for AI persuasion
The comprehensive study was conducted between December 2023 and February 2024. It was funded by the Swiss National Science Foundation and the European Union, and it is published as a preprint on arXiv.
The broader concern is already visible in public warnings about large language models. OpenAI CEO Sam Altman recently warned:
"I expect AI to be capable of superhuman persuasion well before it is superhuman at general intelligence, which may lead to some very strange outcomes."
The study gives that concern a specific shape. It indicates that the persuasive power of AI may depend heavily on personalization, not just on the raw quality of the model. A chatbot that knows little about its audience may be only modestly more persuasive than a human. A chatbot with even basic personal context may be far more effective.
For online platforms, the implication is direct: AI persuasion is not only a future capability to watch. It is a design, moderation, and trust problem that may grow as personalized chatbots become easier to deploy.