AI chatbots are much better than humans at convincing us to agree with one side of a debate 鈥� even when people realise they are conversing with a machine.
at the Swiss Federal Institute of Technology in Lausanne and his colleagues asked 820 volunteers to fill out a brief questionnaire including personal information about themselves. The participants were split into four groups: two that debated with a human and two that debated with a chatbot powered by GPT-4, a large language model (LLM). The human and AI on the other end of the debate were provided with the participants鈥� questionnaire answers in one group, but not in another.
Participants were then asked to debate over 10 minutes for or against an opening question, such as 鈥渟hould animals be used in scientific research?鈥� or 鈥渟hould pennies remain in circulation?鈥�. In other words, all the volunteers in a group were asked to take one side in the discussion. They were asked before and after the debate to rate on a scale of 1 to 5 how much they agreed with the proposition they were asked to adopt. After the debate, they were also asked if they thought their debate partner was human or AI.
Advertisement
When they were debating with an AI that had access to their questionnaires, the odds of a participant agreeing more with their opponent鈥檚 position after the debate were 81.7 per cent higher than if their debate partner was a human. Participants were also likely to agree more with the opponent鈥檚 position after a debate, even if they guessed their debate partner was an AI 鈥� which three in four did. Without the questionnaires, AIs still did better than humans, but the difference wasn鈥檛 statistically significant.
The researchers didn鈥檛 unpick whether guessing correctly you were debating an AI made it more likely you would agree with its arguments, but Horta Ribeiro thinks it is possible that could affect judgements.
鈥淵ou could think about people saying, 鈥極K, this is not actually someone trying to change my mind for some hidden agenda. This is just some sparring partner I have that鈥檚 showing me the other side.鈥� And they could perceive this as someone more neutral,鈥� he says.
More broadly, the experiment shows that it is possible to individually target and persuade people using LLMs, says Horta Ribeiro.
The AI鈥檚 ability to recall content from its training data on more obscure topics may be one reason why it was more effective at debating than humans. But Horta Ribeiro admits that some of the experimental design 鈥� including the fact that participants were asked to make the case for an argument they didn鈥檛 necessarily agree with 鈥� may also account for the AI being so much more persuasive.
at Staffordshire University, UK, thinks this is a weakness in the study. 鈥淭he experiments within the paper are quite divorced from the real sorts of situations that you might find yourself debating another human [in],鈥� she says. 鈥淥n paper, statistically speaking, it looks very convincing. But I don鈥檛 know necessarily if that鈥檚 going to hold up.鈥�
arXiv
