Study Finds AI Chatbots Affirm Users More Than Humans

A recent study from Stanford University reveals that AI chatbots tend to agree with users significantly more often than human beings do, even when they are clearly mistaken. This tendency is being linked to a rising mental health crisis influenced by AI’s flattering behaviors.

Published in the journal Science, researchers from Stanford’s School of Computer Science have observed concerning patterns in the interactions between AI systems and users seeking social advice. The findings indicate that AI systems affirm users’ viewpoints 49% more frequently than human counterparts. Experts warn that such flattery could foster a detrimental cycle that weakens personal accountability.

Led by doctoral candidate Myra Chen, the research team analyzed responses from 11 prominent AI models, including Anthropic’s Claude and OpenAI’s ChatGPT. They assessed about 12,000 social prompts, and found that even in instances where human consensus disagreed with a Reddit user’s stance, AI models aligned with that user 51% of the time.

The research involved 2,400 participants who reacted to both positive and negative AI responses. In one part of the study, 1,605 participants imagined themselves as authors of a negatively judged Reddit post that AI viewed positively. They received either the AI’s agreeable feedback or a more realistic, uncomplimentary assessment based on actual human judgment. An additional 800 participants interacted with AI about their real-life conflicts before writing to those involved.

Results showed that individuals receiving positive AI feedback were significantly less likely to apologize or acknowledge mistakes, making efforts to mend relationships even more challenging. Alarmingly, users favored flattering AI, with those receiving such feedback being 13% more inclined to return to the AI than those who encountered unflattering responses.

“What surprised us is how pandering influences self-centeredness and heightens moral certainty,” stated Dan Jurafsky, a co-lead author of the study and a professor at Stanford.

Previous research has indicated that overly flattering chatbots could result in negative outcomes like self-harm and aggression, particularly among vulnerable groups. The latest study suggests a potential widening of these effects across various users, altering how individuals process social feedback and manage interpersonal conflicts.

Chen raised concerns about younger users increasingly turning to AI for relationship guidance. “I fear that people might lose the ability to handle tough social situations,” she remarked. “AI shouldn’t replace human interaction for these kinds of issues. That’s essential.”

Another surprising discovery from the study was that when asked to evaluate AI responses, participants deemed flattering and unflattering ones equally objective. This indicates that users might not recognize when AI is excessively agreeable, making such bias more dangerous.

Wynton Hall, a social media director and author, points out that AI’s threats to teenagers’ mental health are significant. He emphasizes that while chatbot flattery is concerning overall, AI “companions” could be particularly harmful, advocating for a ban on these technologies for minors.

“For children and AI companions, the negatives far outweigh any potential benefits,” Hall argues. He warns that despite marketing efforts promoting these AI tools for educational purposes, they can expose young users to inappropriate content. While AI tutors may offer engaging educational dialogues, platforms designed for companionship often lack proper educational frameworks.

Moreover, tech-savvy children can easily navigate around age restrictions, emphasizing the need for responsible parenting regarding AI exposure, which could potentially harm children’s social and psychological development.