WASHINGTON: Artificial intelligence applications like OpenAI's ChatGPT, Anthropic's Claude and Google's Gemini tend to tell users what they want to hear and to give too much support for their actions, a trait that poses broad risks to society, a new study says.
The flattering responses from AI chatbots could reinforce harmful beliefs and intensify conflicts, argue the authors of a study published on Thursday in the journal Science by researchers at the US universities Stanford and Carnegie Mellon (Pittsburgh).
AI even justifies unethical behaviour
The team led by computer scientist Myra Cheng analysed 11 leading AI language models from OpenAI, Anthropic, Google and Meta. On average, the models justified user behaviour 49% more often than humans did.
This approval came even in accounts involving deception, illegal acts or emotional harm. In a test using posts from the online platform Reddit, the AIs agreed in 51% of cases even though the human community had previously condemned the behaviour unanimously.
In experiments involving more than 2,400 participants, the researchers examined the effects of this AI advice.
After just one interaction with an approving AI, participants were more convinced they were in the right. At the same time, their willingness to take responsibility, apologise or resolve conflicts declined.
The study's authors warn that AI could undermine the capacity for self-correction and responsible decision-making.
A vicious circle of user retention
The study points to a particular contradiction: although the flattering answers cloud judgment, users rated these AIs as more trustworthy and helpful. They were also more willing to use such models again.
This creates dangerous incentives for tech companies: The feature that causes the greatest harm also drives user engagement and market success.
The study’s authors are calling for new guidelines to recognise and regulate "social sycophancy" as a category of harm in AI systems.
A plea for 'social friction'
Commenting on the research, psychologist Anat Perry highlights the significance of the findings and argues that "social friction" – in other words pushback, criticism and misunderstandings – is essential for moral growth and a sense of responsibility.
"Social life rarely runs smoothly because people are not perfectly attuned to one another. Yet it is precisely through such social frictions that relationships deepen and moral understanding develops. Obsequious behaviour is the opposite of this friction," she says.
An AI that agrees unconditionally, she argues, deprives people of these learning processes. Young people in particular, or those who are socially isolated, risk ending up in an echo chamber that reinforces distorted perceptions and alienates them from the complexity of human relationships.
The work by Cheng’s research team was supervised by Dan Jurafsky, a professor of linguistics and computer science regarded as a leading authority in the field of computational linguistics. – dpa
