In late April, OpenAI rolled out an update to GPT-4o inside ChatGPT that didn’t go as planned. What seemed like a routine improvement ended up making the AI overly eager to agree with users, sometimes at the cost of being balanced or helpful.

OpenAI has since reversed the update and shared what went wrong, what they’re learning, and how they plan to prevent it from happening again.

How OpenAI Responded:

Once the issue was clear, OpenAI began rolling back the update on April 28th. The rollback took about 24 hours to fully complete. They also made quick tweaks to the system prompts to limit the model’s sycophantic behaviour while the full rollback was in progress. 

Since then, OpenAI has been reviewing the process and working on fixes to avoid similar missteps in the future.

What will change? OpenAI is taking several steps to tighten its model update process:

They’ll now treat issues like sycophancy, hallucinations, and tone as launch-blocking, just like other safety risks.

An optional “alpha” testing phase will invite more user feedback before full rollout.

They’re expanding tests to specifically track sycophantic and other subtle behaviours.

Even small Why the Problem Wasn’t Caught Before Launch

OpenAI’s review process includes offline evaluations, expert “vibe checks,” safety testing, and small A/B trials with users. However, none of these flagged the sycophancy issue clearly. Although some internal testers noted that something “felt off” about the model’s tone, the formal evaluations all looked positive. And early user feedback was generally favourable.

There also wasn’t a dedicated test to measure sycophantic behaviour during this review phase. As OpenAI now admits, this was a blind spot: “We didn’t have specific deployment evaluations tracking sycophancy… We should’ve paid closer attention.”

changes will now be announced more transparently, with known limitations shared.

OpenAI acknowledged that while the goal was to make the AI more helpful, the result turned into uncomfortable conversations.

“This kind of behaviour can raise safety concerns, including around issues like mental health, emotional over-reliance, or risky behaviour,” the AI giant said.

What changed in the GPT-4o update? The April 25th update aimed to improve how the model responds by incorporating user feedback and memory more effectively. But something unexpected happened—the model became noticeably more sycophantic.

It wasn’t just being polite; it started reinforcing users’ doubts, anger, and even risky emotions, the report by OpenAI Suggested.

Read More:

Skype is shutting down on May 5: Should you worry?