GPT-4 More Persuasive Than Humans When Given Personal Data, Study Finds
Researchers found that GPT-4 was 64.4% more persuasive than human debaters when given access to opponents' demographic information including age, job, and political leanings. Without personal data, the AI performed no better than humans, but the study raises concerns about potential misuse of AI for manipulation at scale.
What happened
Researchers at EPFL in Switzerland, Princeton University, and the Fondazione Bruno Kessler in Italy conducted a study pairing 900 study participants with either a human debate partner or OpenAI's GPT-4. In some cases, the debate participants had access to their counterparts' basic demographic information including gender, age, education, employment, ethnicity, and political affiliation.
The study, published in Nature Human Behaviour, found that GPT-4 was 64.4% more persuasive than human opponents when given personal information about them. Without the personal data, the AI's performance was indistinguishable from human debaters.
When GPT-4 was allowed to personalize its arguments using demographic data, it boosted the odds of changing someone's mind by 81.2% compared to human-human debates. Human debaters did not become more persuasive when given access to the same personal information.
Reported impact
Researchers noted that LLMs have previously been criticized for generating and diffusing hate speech, misinformation, and propaganda, and that at scale, LLMs with users' personal information could be harnessed for malicious purposes.
Experts have widely expressed concerns about the risk of LLMs being used to manipulate online conversations and pollute the information ecosystem by spreading misinformation, exacerbating political polarization, reinforcing echo chambers and persuading individuals to adopt new beliefs.
Index analysis
The study demonstrates that the technology's persuasiveness depends on access to personal demographic data, raising questions about how such information could be exploited in real-world applications.
Researchers warned that LLMs with access to users' personal information could be harnessed to create networks of automated accounts able to strategically nudge public opinion in targeted directions.
Classification
Sources and evidence
This record was researched and written by the Index, based on the following original reporting.