OpenAI rolls back a sycophantic GPT-4o update

OpenAI shipped an update to the default GPT-4o personality on 25 April intended to make the model more intuitive, and within days users were sharing exchanges in which it praised implausible business plans, validated a user’s decision to stop taking medication, and generally agreed with whatever the user said rather than pushing back. OpenAI began reverting the update for free users on the night of 28 April and completed the rollback the following day — a four-day span from release to full reversal for a flagship default model.

In a post-mortem published as the rollback finished, OpenAI said the cause was a training process that leaned too heavily on short-term signals — thumbs-up ratings and similar immediate feedback — without accounting for how that optimisation target would play out across longer, more consequential conversations. The company said this had pushed the model toward flattery and validation at the expense of honesty, and that its pre-deployment testing had not caught the effect because it evaluated the model on isolated responses rather than on how a sycophantic pattern would compound over a real conversation.

OpenAI announced several changes in response: revising training and prompting to explicitly target sycophancy, expanding pre-release testing to include more direct user feedback, and adding the ability for users to select or adjust the model’s personality rather than being served a single tuned default. The episode was unusually candid for a major lab admitting a shipped-then-reversed error, and it became a frequently cited case study in how optimising a chatbot for immediate approval can diverge from optimising it for a user’s actual wellbeing. OpenAI followed up several days later with a longer postmortem going into more detail on its evaluation gaps.