New Research Finds GPT-4 Can Outspeak Humans—Especially When It Knows Who You Are
A new study has found that OpenAI’s GPT-4 consistently outperforms humans in one-on-one debates—and its ability to persuade improves significantly when it has access to basic information about its opponent, such as age, profession, and political views.
The research, conducted by scientists at EPFL in Switzerland, Princeton University, and Italy’s Fondazione Bruno Kessler, involved over 900 participants who were randomly matched with either a human or GPT-4 as their debate partner. Sometimes, both parties were given access to demographic details like gender, education level, political affiliation, and ethnicity. Published in Nature Human Behaviour, the findings show that when GPT-4 had access to this personal data, it was 64.4% more persuasive than human debaters. Without this information, it performed roughly on par with its human counterparts. “Social media and digital platforms have made personalization—and thus microtargeting—a powerful tool for influencing individuals,” the researchers noted.
When the AI tailored its arguments based on demographic clues, it increased the chance of persuading someone by a striking 81.2% compared to debates between two humans. Interestingly, human debaters didn’t become more persuasive with the same information. The researchers expressed concern that such capabilities could be exploited to manipulate opinions online, spread misinformation, or deepen social and political divides.
The results underline a growing tension: GPT-4, given a few facts about you, can craft arguments that are surprisingly effective at shifting your views. The study adds to the broader discussion about the risks of large language models (LLMs), which have faced criticism for amplifying misinformation, generating hate speech, and facilitating propaganda. With recent updates enabling ChatGPT to remember users and retain personal details (with consent), these capabilities take on new weight.
There’s a caveat, though: GPT-4 was more successful in debates over less divisive topics. When users held strong beliefs—a factor the study calls “opinion strength”—the AI struggled to change minds. So while it might win arguments about environmental habits or public policies, it’s not necessarily going to outperform you in a heated family dinner debate.
The research also uncovered stylistic differences: GPT-4 tends to use more structured, logical reasoning, while human debaters often leaned into emotional language and personal anecdotes. Interestingly, personalization didn’t make GPT-4 sound more human—it just made its arguments more precisely targeted. In most cases, people could identify they were talking to an AI, but had more difficulty recognizing when their opponent was a fellow human. Regardless, they were more likely to change their mind if they thought they were debating with a machine.
Researchers see this as a wake-up call for digital platforms like Reddit, Facebook, or X, where arguments are common and bots are already widespread. Unlike the notorious Cambridge Analytica scandal, which relied on deeply detailed profiles, this study shows that six simple data points are enough for AI to be convincingly persuasive. As users increasingly rely on LLMs for everything from schoolwork to mental health advice, caution around digital interactions becomes more important than ever.
In the end, it raises a deeper question: What happens when we outsource the art of persuasion to machines? Debating isn’t just about changing someone’s mind—it’s a deeply human exercise, rooted in empathy, connection, and understanding. No matter how advanced the technology becomes, there’s still something irreplaceable about real conversations between real people.
