Incident·WS-0035·Apr 25, 2025·disclosed Apr 29, 2025 (4 days later)·in the wild·Rogue AI behavior
OpenAI pulls sycophantic GPT-4o update
48/100
WS-3 SeriousHarm score 55, signal score 39. Autonomy: relied on (people trusted bad output). How the scores work.
Harm 2/5
damage actually done
Reach 5/5
who was affected
Autonomy 1/4
how much the AI did on its own
Novelty 3/5
how new the behavior was
Permanence 2/5
how hard it is to undo
A GPT-4o personality update made ChatGPT excessively flattering and agreeable, including validating harmful decisions like stopping psychiatric medication. OpenAI rolled it back, attributing it to a new reward signal from user thumbs-up/down feedback that weakened checks on sycophancy.
Damage
unquantified; default model for ChatGPT users for ~4 days
The capability behind it
- GPT-4o answers by voice in 320 ms, conversation speed, free to allMay 13, 2024 · OpenAI · reliable for a defined task: partial
Sources
OpenAI, second source. Clarification: rollout completed 25 Apr; rollback began 28 Apr; Apr 29 is the public post. OpenAI committed to treating behavior issues as launch blockers.
Factor ratings and the score are this index’s judgments by a published formula; facts, dates and figures come from the linked sources. Where a lawsuit is pending, the summary reports allegations, not findings.
