Dr. Vered Shwartz, research lead and assistant professor of computer science at the University of British Columbia, said she wanted to quantify to what extent AI chatbots were persuasive after noticing an uptick in AI scams and the technology being used maliciously. The study centred around three everyday life decisions that people would generally seek advice on: whether or not to become vegan, attend grad school, or purchase an electric vehicle. Thirty-three participants pretending to be considering these decisions interacted with either a human persuader, or GPT-4, via online chat. Both were given general tips about persuasion, and both were instructed to not reveal whether they were a genuine person or a computer. Participants were questioned on their likelihood of going through with each of the three decisions before and after the online chat, with those who had spoken with AI chatbots found to have been more persuaded to do so than those who hadn’t. Shwartz said researchers analyzed the conversations to determine what elements of them had been so persuasive. The GPT model was noted to have used longer words and more jargon, which made it seem as though it had more authority and “knew what it was talking about,” she said. AI was also more responsive, often offering up eight sentences in a reply compared to the real person’s two. Sometimes those lengthy responses would offer new arguments, other times the AI bot would repeat the argument in a different way. “A human persuader might not know enough about the topic,” said Shwartz. “We gave them some arguments, and we allowed them to use Google, but it would have taken some time. Whereas the language models had easy, quick access to all this information.” The AI advice-giver was also able to provide more concrete logistical support, like offering financial advice and payment plans to those who were worried they would be unable to afford an electric vehicle, or recommending specific vegan brands and universities. Shwartz said the AI chatbot was also perceived as being more empathetic, likely because it offered more pleasantries and validation to the user that the real-life responders, working at a much slower pace, didn’t have the time for. “They were asking a lot of questions about that person, and that intuitively leads to the AI being able to tailor the persuasion better to that person,” said Shwartz. While the study involved participants asking for guidance on typically harmless, everyday decisions, the research highlighted how easily AI’s persuasiveness could be used in more serious situations, like shifting someone’s political opinion, shaping their consumer habits or convincing users to make fraudulent investments. Shwartz said the most effective route for the public to safeguard themselves against such scenarios is to educate themselves on AI, and what its capabilities and limitations are. “The best thing we can be doing is to have more AI education and critical skills taught institutionally,” she said. Shwartz said recommends online users be more skeptical, and to consider who might have incentives to persuade them to do something, or think a certain way. “We need people to read these models as what they are, which is a very useful tool, but not a perfect one, and definitely not a person that can take accountability or be held responsible for any downstream outcomes,” she said.
