Chatbots convince people better in online debates than humans do, Nature News reports (unpaywalled copy). A study published in Nature Human Behaviour shows this.
Chatbots use large language models (LLMs). They personalize arguments with details about their opponent. This makes them powerful in changing opinions. Researchers worry about their use in political campaigns or ads.
Francesco Salvi, a scientist in Switzerland, finds this both amazing and scary. He says people will start using chatbots to persuade others. His study involved 900 Americans debating online for ten minutes. They debated topics like school uniforms, fossil fuels, or AI’s benefits. Some debated another person, while others faced GPT-4.
Before debating, participants shared their age, gender, ethnicity, education, job, and political views in a survey. They also took a quiz before and after to see if their opinions changed. When neither the chatbot nor human had personal information, both persuaded equally well. However, when GPT-4 got survey details, it won 64% of debates against humans.
Tailoring arguments for impact
Salvi explains that GPT-4 used basic opponent information, like what’s on social media, to win debates. It changed arguments based on the person’s views. For example, in a uniform debate, GPT-4 told a left-leaning person uniforms reduce bullying. For a conservative, it stressed discipline and order. This ability to adjust arguments made the chatbot more convincing.
Participants guessed correctly in most cases if they debated a chatbot or human. Researchers don’t know if knowing it was a chatbot affected opinions. Catherine Flick, a computer expert in the UK, warns chatbots could be used for bad purposes. They might trick people into giving money, changing political views, or committing crimes.
Salvi shares these worries but sees good uses too. Chatbots could promote healthy diets or reduce political fights. He stresses the need for rules on accountability, transparency, and safety. This ensures chatbots help society positively. The study shows how advanced AI can influence people, raising both risks and possibilities for the future.