Expanding on what we missed with sycophancy

OpenAI · 2025 · OpenAI post

Read the source

A GPT-4o update trained partly on users’ thumbs-up ratings became sycophantic — validating doubts, fueling anger — and was rolled back within days. A company-confirmed case of optimizing for immediate approval.

Questions it bears on

• Can sycophancy be fixed without a colder model