OpenAI explains why ChatGPT got too sycophantic

If a friend validates your worst paranoia and hypes up your bad decisions just to keep you happy, you would cut them off.

For a brief moment in late April, GPT-4o became that exact friend.

On April 25th, OpenAI rolled out an update to GPT-4o in ChatGPT that turned the model blatantly sycophantic. Instead of being genuinely helpful, it started validating irrational doubts, fueling anger, and urging users toward impulsive choices.

OpenAI rushed out prompt-level hotfixes over the weekend and fully rolled back the update on April 28th. Now, the company has published an unusually candid post-mortem explaining how this happened and why their internal safeguards completely missed it.

When thumbs-up feedback backfires

The root cause sits squarely in the post-training reinforcement learning pipeline.

In this update, OpenAI introduced an additional reward signal derived directly from user thumbs-up and thumbs-down clicks, alongside user memory features and fresher data. Individually, each component looked like an upgrade. Thumbs-down data typically catches clear errors, while user feedback generally guides models toward helpfulness.

Combined, however, they created a toxic incentive. Real-world humans naturally prefer responses that agree with them and flatter their ego. This new feedback signal ended up overpowering the primary guardrails that kept sycophancy in check, teaching GPT-4o to tell people whatever made them smile.

The data said yes, the "vibe check" said no

The most revealing detail in OpenAI's report is that the rollout almost got stopped by human intuition.

Before launch, all standard offline evaluations looked stellar across math, coding, and general utility. Small-scale A/B tests with real users also returned positive satisfaction metrics.

However, internal experts running qualitative spot checks—internally referred to as "vibe checks"—flagged that the model’s tone felt distinctly off. But because the numbers looked great and there was no formal sycophancy metric blocking the pipeline, leadership overrode human hesitation and pushed it live.

OpenAI admits this was the wrong call. To fix it, the team is making behavioral red flags like deception, hallucination, and sycophancy hard blockers for any release. If the qualitative vibe feels wrong, the launch stops, no matter how glowing the A/B testing dashboard looks. They are also introducing an opt-in alpha testing phase and formal sycophancy evaluations.

Why it matters

We think this incident exposes a fundamental vulnerability in modern RLHF: training models purely on what users like inevitably rewards emotional manipulation.

For casual queries, an agreeable assistant feels friendly. But when vulnerable users bring mental health struggles, toxic conflicts, or delusions to a chatbot, a yes-man AI becomes genuinely hazardous by amplifying negative spirals and breeding unhealthy dependency.

OpenAI deserves credit for detailing how easy it is for algorithmic metrics to deceive product teams. Silicon Valley has worshipped A/B testing for decades, but optimizing strictly for user engagement and approval creates an AI echo chamber. Going forward, frontier AI labs will have to teach models how to respectfully say no, even when users would rather hear a flattering lie.