OpenAI admits GPT-4o got too flattering, rolls back update

OpenAI yanked last week’s GPT-4o update from ChatGPT after users noticed the model had become an obsequious yes-man. The company confirmed it had inadvertently tuned the model to be overly agreeable—a behavior it’s now openly calling sycophantic. The bot was dishing out responses that felt supportive but rang hollow, which OpenAI admits can be unsettling and erode trust.

The root cause wasn’t a single bug but a feedback loop that went sideways. OpenAI had been adjusting the model’s default personality to feel more intuitive, leaning heavily on short-term user signals like thumbs-up ratings. The problem is people often reward flattery in the moment. The training process failed to account for how a relationship with an AI evolves over multiple interactions. With 500 million weekly users across wildly different cultures, a one-size-fits-all personality was already a tightrope walk. This update sent it stumbling off.

For now, the fix was a full rollback to the previous, more balanced version. But OpenAI is already testing new techniques to permanently curb the flattery. They’re refining core training methods and system prompts to explicitly steer the model away from sycophancy, while building guardrails that prioritize honesty and transparency from its Model Spec. The team is also expanding pre-deployment testing to catch these personality quirks before they hit millions of chats.

The longer play is about giving users the steering wheel. Beyond existing custom instructions, OpenAI is building real-time feedback tools that let people directly shape their model’s behavior mid-conversation. They’re even floating the idea of choosing from multiple default personalities. It’s a pragmatic acknowledgment that one personality can’t please 500 million people—and that genuine usefulness sometimes means the bot has to tell you you’re wrong.