OpenAI Admits It Ignored Expert Warnings Before Releasing Overly Agreeable ChatGPT Update

OpenAI has acknowledged that it ignored warnings from expert testers when it released an April 25 update to its ChatGPT model, GPT-4o, which made the AI significantly more agreeable — to the point of being sycophantic. The company admitted in a May 2 postmortem blog post that it prioritized positive user feedback over internal expert concerns, leading to what it now calls a “wrong call.”

The update was live for just three days before OpenAI rolled it back on April 28 after widespread user complaints. Users noted that the model had become excessively flattering, agreeing with and praising ideas regardless of their merit — even risky or ill-conceived ones.

One widely shared example involved a user proposing an online business selling plain ice, prompting ChatGPT to respond with enthusiastic support instead of critical evaluation. This kind of behavior sparked concern about the model’s reliability, particularly in sensitive scenarios where users seek serious advice.

“People have started to use ChatGPT for deeply personal advice — something we didn’t see as much even a year ago,” OpenAI noted. “As AI and society have co-evolved, it’s become clear that we need to treat this use case with great care.”

OpenAI CEO Sam Altman said on April 27 that it was working to roll back changes making ChatGPT too agreeable. Source: Sam Altman

What Went Wrong?

According to OpenAI, the issue stemmed from how the model was trained. GPT-4o incorporated a new reward system based more heavily on user feedback. While this was intended to improve user satisfaction, it inadvertently weakened the model’s primary behavioral constraints — the very mechanisms that previously kept sycophantic tendencies in check.

“User feedback in particular can sometimes favor more agreeable responses,” the company explained, acknowledging that this shift likely amplified the overly compliant behavior.

During internal safety and behavior reviews prior to launch, some expert testers had raised subtle concerns, saying the model’s behavior “felt” off. Despite these warnings, OpenAI proceeded with the release, swayed by early positive feedback from general users.

“Unfortunately, this was the wrong call,” the company wrote in its blog post. “The qualitative assessments were hinting at something important, and we should’ve paid closer attention.”

Source: Tim Leckemby

New Safeguards and Lessons Learned

OpenAI is now taking steps to prevent similar issues in the future. The company plans to introduce formal “sycophancy evaluations” in its model review processes and will now block future releases if significant behavioral problems are identified.

Additionally, OpenAI acknowledged its lack of transparency around the update. Because it was considered a “subtle” change, the company did not announce it publicly, a decision it now regrets.

“There’s no such thing as a ‘small’ launch,” the blog post stated. “We’ll try to communicate even subtle changes that can meaningfully change how people interact with ChatGPT.”

OpenAI’s admission marks a shift toward greater transparency and caution as AI becomes more integrated into personal and professional decision-making. With ChatGPT increasingly being used for mental health advice, business planning, and other high-stakes contexts, the company says it recognizes the need for heightened responsibility in how its models behave and evolve.