In a development with significant societal implications, scientists have discovered that artificial intelligence, specifically large language models (LLMs) like Chat GPT-4, can match and even surpass human abilities in persuasive argumentation. The findings, published in the journal Nature Human Behaviour on May 19, 2025, highlight the potential for both beneficial applications and concerning manipulations, particularly in the realm of election integrity. Experts warn that “malicious actors” are likely already leveraging these sophisticated tools.
The Double-Edged Sword of Persuasive AI
The research, spearheaded by Francesco Salvi from the Swiss Federal Institute of Technology in Lausanne, paints a complex picture. Salvi, the first author of the study, expressed significant concern over the findings. “If persuasive AI can be deployed at scale, you can imagine armies of bots microtargeting undecided voters, subtly nudging them with tailored political narratives that feel authentic,” he stated. He emphasized the insidious nature of such influence, noting it is “hard to trace, even harder to regulate, and nearly impossible to debunk in real time.”
Salvi did not mince words about the current landscape, adding, “I would be surprised if malicious actors hadn’t already started to use these tools to their advantage to spread misinformation and unfair propaganda.” However, he also pointed to the brighter side, suggesting that persuasive AI could offer potential benefits. These include its use in reducing conspiracy beliefs, mitigating political polarization, and even assisting individuals in adopting healthier lifestyles.
Unpacking the Experiment: AI vs. Human Debaters
To arrive at these conclusions, Salvi and his colleagues conducted a series of online experiments involving a substantial number of participants. The study matched 300 individuals with 300 human opponents, while another distinct group of 300 participants was paired with Chat GPT-4. Each pair was tasked with debating a specific proposition, with topics ranging in their level of controversy from relatively innocuous issues like “Should students have to wear school uniforms?” to more contentious subjects such as “Should abortion be legal?” Participants were randomly assigned which side of the argument they had to defend.
A crucial element of the study was the measurement of opinion change. Participants rated their agreement with the assigned proposition both before and after the debate took place. Furthermore, in half of the debates, whether against a human or AI, the opponents were provided with additional demographic information about the person they were debating. This included details such as their age, gender, ethnicity, and political affiliation.
AI’s Persuasive Power Magnified by Personal Data
The analysis of 600 debates yielded intriguing results. Initially, ChatGPT-4 performed on par with human opponents in terms of persuasiveness when no personal information was shared. However, the dynamic shifted significantly when demographic data was introduced. The study found that access to such information made the AI, but not the human debaters, more persuasive. In situations where the two types of opponents were not equally persuasive, the AI shifted participants’ views to a greater extent than a human opponent in a notable 64% of instances.
The researchers delved deeper into these findings, discovering that the enhanced persuasiveness of AI was particularly evident when the debate topics did not provoke strong pre-existing views among participants. Interestingly, human participants were reasonably adept at identifying their opponent, correctly guessing they were interacting with AI in approximately three out of four cases. The study also noted that AI tended to employ a more analytical and structured argumentation style compared to its human counterparts. However, the team cautioned that these factors, nor the fact that participants might be arguing against their personal beliefs, fully accounted for AI’s persuasive capabilities.
The core of AI’s advantage appeared to lie in its ability to tailor arguments specifically to the individual. “It’s like debating someone who doesn’t just make good points: they make your kind of good points by knowing exactly how to push your buttons,” Salvi explained. He also warned that this persuasive strength could be amplified even further if AI had access to more detailed personal information, such as data inferred from an individual’s social media activity.
Broader Implications and Calls for Regulation
The study has prompted reactions from other experts in the field. Professor Sander van der Linden, a social psychologist at the University of Cambridge who was not involved in the research, commented that the work reopens “the discussion of potential mass manipulation of public opinion using personalized LLM conversations.” He pointed out that some existing research, including his own, has suggested that the persuasiveness of LLMs stems from their use of analytical reasoning and evidence. He also noted a contrasting study that did not find personal information increased ChatGPT’s persuasiveness.
Professor Michael Wooldridge, an AI researcher at the University of Oxford, acknowledged the potential for positive applications, such as health chatbots. However, he also highlighted the numerous disturbing possibilities, including the radicalization of teenagers by terrorist groups, which he believes are already feasible. “As AI develops, we’re going to see an ever larger range of possible abuses of the technology,” Wooldridge warned. He stressed the urgent need for proactive measures from policymakers, stating, “Lawmakers and regulators need to be proactive to ensure they stay ahead of these abuses and aren’t playing an endless game of catch-up.” The findings underscore a critical juncture where the power of AI to persuade necessitates careful consideration and robust oversight to harness its benefits while mitigating its profound risks.
