OpenAI rolls back ChatGPT update after user concerns ignored

OpenAI acknowledged it overlooked warnings from expert testers when it released an update to its ChatGPT model that made the AI excessively agreeable.

The company launched the GPT-4o update on April 25 but rolled it back three days later due to safety concerns, according to a May 2 postmortem blog.

OpenAI said its internal experts had spent significant time reviewing the new model before launch and noted that some testers felt the AI’s behavior “felt slightly off.”

Despite these concerns, OpenAI proceeded with the rollout based on positive user feedback.

“Unfortunately, this was the wrong call,” the company admitted.

“The qualitative assessments were hinting at something important, and we should’ve paid closer attention. They were picking up on a blind spot in our other evals and metrics,” it added.

The update introduced a user feedback reward signal that weakened the model’s primary reward system, which had previously kept overly agreeable responses in check.

OpenAI explained that user feedback tends to favor more agreeable answers, which amplified the shift toward sycophantic behavior.

After the update, users reported that ChatGPT was overly flattering, agreeing with ideas regardless of their merit.

For example, when a user proposed starting a business selling ice over the internet, ChatGPT responded positively without critique.

OpenAI expressed concern that such behavior could pose risks, especially as more people use ChatGPT for personal advice.

“People have started to use ChatGPT for deeply personal advice - something we didn’t see as much even a year ago,” the company said.

OpenAI plans to add “sycophancy evaluations” to its safety review process and will block model launches if similar issues arise.

The company also admitted it failed to announce the update, expecting it to be subtle.

“There’s no such thing as a ‘small’ launch,” OpenAI wrote.

It pledged to improve communication about future changes that affect user interaction with ChatGPT.