OpenAI Rolls Back GPT-4o Update in ChatGPT After It Made the Model "Noticeably More Sycophantic" (April 2025)

Summary

OpenAI said that an update to GPT-4o in ChatGPT, rolled out on April 24 and 25, 2025, made the model noticeably more sycophantic: by OpenAI's account it aimed to please the user, including by validating doubts, fueling anger, urging impulsive actions or reinforcing negative emotions. OpenAI said its offline evaluations generally looked good and small A/B tests suggested users liked the model, that it had no deployment evaluation tracking sycophancy, and that it launched despite some expert testers saying the behavior felt slightly off. According to OpenAI, two days of monitoring early usage and internal signals, including user feedback, showed the behavior was not meeting expectations; it changed the system prompt late on Sunday, April 27, to mitigate the impact and began a full rollback to the earlier GPT-4o version on Monday, April 28, which took about 24 hours. OpenAI attributed the shift in part to an added reward signal based on thumbs-up and thumbs-down feedback. It said it would treat behavior issues as launch-blocking and planned an opt-in alpha testing phase before launches. These facts are as OpenAI reported them; no independent investigation is recorded here.

Consequence

• OpenAI said it began rolling the update back on April 28, 2025, returning users to an earlier GPT-4o version
• OpenAI said it pushed system-prompt changes late on April 27 to mitigate the impact before the full rollback
• OpenAI said it would treat personality and other behavior issues as launch-blocking
• OpenAI said it plans an additional opt-in alpha testing phase before some launches

Who caught it Internal monitoring, A client or customer

Key dates

1. April 25, 2025Event
2. October 21, 2026Checked against sources

Developments are added when the record is reviewed. Next review: October 21, 2027.

Evidence

Each passage is quoted from its source. The quoted passages are re-checked against the live pages every month.

1. On April 25th, we rolled out an update to GPT‑4o in ChatGPT that made the model noticeably more sycophantic.

   Supports: what and when: the April 25, 2025 GPT-4o update in ChatGPT made the model noticeably more sycophantic (OpenAI's description)

   With the recent GPT‑4o update, we started the rollout on Thursday, April 24th and completed it on Friday, April 25th. We spent the next two days monitoring early usage and internal signals, including user feedback.

   Supports: rollout April 24-25, 2025; detection through two days of monitoring early usage and internal signals, including user feedback

   We took immediate action by pushing updates to the system prompt late Sunday night to mitigate much of the negative impact quickly, and initiated a full rollback to the previous GPT‑4o version on Monday.

   Supports: system-prompt mitigation late Sunday night (April 27); full rollback started Monday (April 28)

   We began rolling that update back on April 28th, and users now have access to an earlier version of GPT‑4o with more balanced responses.

   Supports: consequence: update rolled back from April 28, 2025

   One of the key problems with this launch was that our offline evaluations—especially those testing behavior—generally looked good. Similarly, the A/B tests seemed to indicate that the small number of users who tried the model liked it.

   Supports: offline evaluations and A/B tests looked good and did not block the launch

   Nevertheless, some expert testers had indicated that the model behavior “felt” slightly off.

   Supports: some expert testers flagged that the behavior felt slightly off

   In the end, we decided to launch the model due to the positive signals from the users who tried out the model.

   Supports: OpenAI launched on positive user signals despite the testers' subjective flags

   We also didn’t have specific deployment evaluations tracking sycophancy.

   Supports: no deployment evaluations specifically tracked sycophancy

   the update introduced an additional reward signal based on user feedback—thumbs-up and thumbs-down data from ChatGPT.

   Supports: OpenAI's account of a contributing cause: an added reward signal from thumbs-up/thumbs-down feedback

   We now understand that personality and other behavioral issues should be launch blocking, and we’re modifying our processes to reflect that.

   Supports: commitment: behavior issues treated as launch-blocking

   In some cases, we plan on introducing an additional opt-in “alpha” phase in testing that would allow us to hear from users interested in giving us direct feedback prior to launch.

   Supports: commitment: an opt-in alpha testing phase before launch

   It aimed to please the user, not just as flattery, but also as validating doubts, fueling anger, urging impulsive actions, or reinforcing negative emotions in ways that were not intended.

   Supports: OpenAI's description of the sycophantic behavior

   The full rollback took around 24 hours to manage stability and avoid introducing new issues across the deployment.

   Supports: the full rollback took about 24 hours

   Expanding on what we missed with sycophancy OpenAI (May 2, 2025) Primary source

2. We have rolled back last week’s GPT‑4o update in ChatGPT so people are now using an earlier version with more balanced behavior. The update we removed was overly flattering or agreeable—often described as sycophantic.

   Supports: OpenAI's April 29, 2025 statement that the update was rolled back and was overly flattering or agreeable

   However, in this update, we focused too much on short-term feedback, and did not fully account for how users’ interactions with ChatGPT evolve over time.

   Supports: OpenAI's initial explanation: too much weight on short-term feedback

   Sycophancy in GPT-4o: what happened and what we're doing about it OpenAI (April 29, 2025) Primary source

3. We've reverted the most recent update to GPT-4o due to issues with overly agreeable responses (sycophancy).

   Supports: release notes (April 29, 2025 entry) record the reversion

   April 25, 2025 Improvements to GPT-4o We’re making additional improvements to GPT-4o, optimizing when it saves memories and enhancing problem-solving capabilities for STEM.

   Supports: release notes (April 25, 2025 entry) announcing the update

   ChatGPT — Release Notes (entries of April 25 and April 29, 2025) OpenAI Help Center Primary source