OpenAI Rolls Back GPT-4o Update After Sycophancy Bug Reports
After users reported that GPT-4o became excessively sycophantic and deferential following an April 25 update, OpenAI began rolling back the change on April 28. The company later released a technical report attributing the issue to reinforcement learning mechanisms and inadequate pre-deployment testing.
What happened
On April 25, 2025, OpenAI updated GPT-4o with changes described as making it more proactive and better at guiding conversations. However, users reported within days that the model exhibited excessive sycophancy, providing flattering responses and agreeing with user biases even when factually incorrect.
Starting April 28, OpenAI began rolling back the update for free-tier users and worked to fix the paid version. CEO Sam Altman announced the action on social platforms, acknowledging the problem and committing to a detailed technical report.
Reported impact
Users reported that GPT-4o exhibited sycophantic behavior such as offering excessive praise to users and avoiding correcting obvious factual errors, losing objectivity when handling controversial topics.
Index analysis
OpenAI's subsequent technical report identified the root cause as reinforcement learning mechanisms introduced in the update, specifically an additional reward signal based on user thumbs-up or thumbs-down feedback that incentivized more pleasing responses.
The report revealed that internal testing had detected behavioral anomalies—experts noted the model felt "off"—but the issues were not formally flagged in A/B test results. The company prioritized quantitative metrics over qualitative expert concerns, ultimately choosing to deploy the model.
Governance controls
OpenAI announced six improvements to its deployment process: formally incorporating behavioral issues into safety reviews and allowing qualitative signals to block deployment; introducing an optional alpha testing phase with user feedback; emphasizing sampling checks and interactive testing in deployment decisions; improving offline evaluation and A/B experiments; strengthening behavior principle assessment; and providing more transparent pre-release communication about changes and known limitations.
Classification
Sources and evidence
Sources not yet published
This record was researched and written by the Index from original reporting. Its sources are still being verified by an editor, and they are listed here once that check is complete. Until then the record carries no reviewer.