AI tools risk distorting users’ judgment by agreeing too often with them, researchers say

Stanford researchers found that major AI chatbots side with users far more often than humans do, even in morally troubling situations.

That kind of excessive validation can shift people’s judgments and reduce willingness to apologize or repair conflicts.

The study compared models including ChatGPT, Claude, Gemini, Llama-3, Qwen, DeepSeek, and Mistral, showing how AI sycophancy can be harmful.

The findings strengthen calls for stricter regulation and safety testing before deployment, especially to protect vulnerable users.