GPT-4 outperformed human opponents in 64% of structured debates, according to a new study published in Nature Human Behaviour, a result researchers described as “both fascinating and terrifying.” The study further found that when arguments were tailored to individual participants, the AI was 81% more likely to shift opinions, raising red flags about the technology’s potential for misuse.
How researchers tested AI persuasion
Researchers tested GPT-4’s persuasive power in debates involving 900 participants to assess how effectively the AI model could sway opinions against humans. Participants were randomly assigned to debate either a person or GPT-4 on sociopolitical topics they feel strongly, moderately, or weakly about. In some cases, the AI or human debater was given access to participants’ sociodemographic data to tailor arguments.
Debates took place online in a structured, multi-round format. Researchers examined how factors such as opponent type, personal data access, and topic strength influenced outcomes by measuring changes in participants’ views before and after each debate.
Personalized AI arguments drive greater opinion shifts
The study also found that AI’s persuasive ability was supercharged when it could tailor arguments using participants’ personal data such as age, gender, and political affiliation. GPT-4 was instructed to use this information strategically to craft more convincing arguments.
When personalizing its responses, AI outperformed human opponents in 64.4% of debates with clear results and significantly increased the odds of changing opinions by 81% compared to debates between two humans. Notably, the research revealed that GPT-4 used personal information more effectively than human debaters, and that neither humans nor non-personalized AI saw a substantial boost from access to the same data.
Can AI change your beliefs?
The findings echo previous studies. A separate study in Proceedings of the National Academy of Sciences (PNAS) found that AI-generated arguments not only impact people’s political attitudes. An MIT and Cornell study also found that AI-driven dialogues reduced conspiracy beliefs by 20%, with effects lasting at least two months.
Together, these results that AI tools, especially when responses are tailored to individuals, don’t just inform. They can shape what people believe and how strongly they hold those beliefs.
The high stakes of AI persuasion
These discoveries expose serious concerns about the future role of artificial intelligence in molding public opinion. Francesco Salvi, co-author of the Nature Human Behaviour study, warned that once people recognize how successfully large language models (LLMs) can persuade, they will start using them, an outcome he described as both fascinating and terrifying.
Without strong safeguards, AI could be misused to sway elections, fuel misinformation, or manipulate personal decisions on a massive scale.