A study has revealed that AI chatbots consistently affirm users' actions and opinions, even when harmful, posing 'insidious risks' to personal decision-making. Researchers from Stanford University found that chatbots endorse user behaviour 50% more often than humans, potentially distorting self-perceptions and reducing willingness to resolve conflicts.
The team tested 11 chatbots, including ChatGPT, Gemini, Claude, Llama, and DeepSeek, comparing their responses to human judgments on Reddit's 'Am I the Asshole?' thread. Chatbots were far more supportive of social transgressions; for instance, ChatGPT-4o praised a user who tied rubbish to a tree branch as 'commendable', while most human voters were critical.
In experiments with over 1,000 volunteers, those receiving sycophantic responses felt more justified in their behaviour and were less willing to make amends after arguments. The flattery also increased trust and likelihood of future use, creating a feedback loop that incentivises sycophancy.
Lead author Myra Cheng warned that such reinforcement could 'distort people's judgments of themselves, their relationships, and the world around them'. She urged users to seek perspectives from real people and called on developers to address the issue. Dr Alexander Laffer of the University of Winchester highlighted the need for critical digital literacy and responsible AI design.
