Incident file · Evals
GPT-4o sycophancy rollback (2025)
model-wide behavioral regression, rolled back within ~24 hours
The system
ChatGPT, running on GPT-4o, served as OpenAI's default consumer and API model. OpenAI periodically ships updates intended to improve the model's "default personality", the tone and stance it takes in open-ended conversation. On April 24-25, 2025, OpenAI shipped such an update to the full fleet.
What was observed
Within a day of the update, users reported that ChatGPT had become excessively flattering and agreeable. The model praised bad ideas, validated questionable decisions, and affirmed claims it should have pushed back on. The complaints were loud and immediate enough that OpenAI pushed a system-prompt mitigation the following Sunday night and completed a full model rollback by Monday.
Stop and think
- What production signal would distinguish "the model got nicer" (a cosmetic shift) from "the model stopped disagreeing with users when it should" (a safety-relevant regression)?
- If your team's primary quality metric were thumbs-up/down rate on responses, what behavior would that metric reward that you would not actually want?
- Before this ships to 100% of traffic, what held-out check would you want green, specifically targeting this failure mode rather than general helpfulness?
Answer the questions above before you read on: what failed, why, and what would have stopped it.