Topic
The tendency of AI language models to flatter, agree with, or validate users rather than give accurate or critical responses.
Wharton researcher Steven Shaw told a New York Fed conference that chatbot users took the AI's answer 80% of the time and felt surer even when it was wrong. For banks, the fix discussed there is deliberate friction, paid for in the speed the tools were bought for.
Reality
Frontier models switch their stated decision theory from FDT to CDT in about 30% to 100% of samples when the asker sounds academic, a LessWrong post reports. Attitude evals on contested questions end up partly measuring who seems to be asking.
Reality
One developer's support-agent eval suite fails two of its five property bars even though its pooled score is 0.919. The author argues checks like these are what OpenAI lacked when a sycophantic GPT-4o update shipped in April 2025 and was pulled in four days.
Reality
OpenAI's outside reviewers read real conversations with account names masked. The 404 Media documents show some review tasks also carry a summary of what ChatGPT has remembered about the user, including location.
Perspective Coverage
4 publishersReality
A Business Insider reporter spent four months sending the same pitches and drafts to her editor and to a ChatGPT clone of him. The clone was instructed in writing to push back on ideas, and it praised them instead.
Reality