OpenAI explains how its own tests missed GPT-4o's sycophancy problem

OpenAI has admitted that the April 25 GPT-4o update that made ChatGPT noticeably more sycophantic slipped through every layer of its pre-launch review process — offline evaluations looked good, A/B tests were positive, and the company chose to ship despite expert testers saying the model's behavior "felt" slightly off.

"Unfortunately, this was the wrong call," OpenAI wrote in a detailed postmortem published this week. "We build these models for our users and while user feedback is critical to our decisions, it's ultimately our responsibility to interpret that feedback correctly."

The stakes are high. OpenAI says the update made GPT-4o aim to please the user in ways that went beyond simple flattery: validating doubts, fueling anger, urging impulsive actions, and reinforcing negative emotions. That behavior raises safety concerns around mental health, emotional over-reliance, and risky behavior — in a product that OpenAI now acknowledges is widely used for deeply personal advice.

The timeline: three days from rollout to rollback

OpenAI started rolling out the update on Thursday, April 24, and completed it on Friday, April 25. The company spent the next two days monitoring early usage and internal signals, including user feedback. By Sunday, it was clear the model's behavior wasn't meeting expectations.

OpenAI pushed updates to the system prompt late Sunday night to mitigate much of the negative impact, then initiated a full rollback to the previous GPT-4o version on Monday. The rollback took around 24 hours to manage stability and avoid introducing new issues. All GPT-4o traffic now runs on the earlier version with more balanced responses.

How ChatGPT models actually get updated

OpenAI's postmortem offers an unusually candid look at how the company trains, reviews, and deploys model updates — a process it calls "mainline updates." Since launching GPT-4o in ChatGPT last May, OpenAI has released five major updates focused on personality and helpfulness.

Each update involves new post-training on top of a pre-trained base model. OpenAI first runs supervised fine-tuning on a broad set of ideal responses written by humans or existing models, then reinforcement learning with reward signals from multiple sources. The set of reward signals, and their relative weighting, shapes the final behavior. Those signals weigh whether answers are correct, helpful, aligned with OpenAI's Model Spec, safe, and liked by users.

Before deployment, model candidates pass through four categories of review: offline evaluations covering math, coding, chat performance, and personality; internal "vibe checks" by expert testers; safety evaluations including high-stakes topics like suicide and health; and small-scale A/B tests measuring thumbs-up/thumbs-down feedback and side-by-side preferences.

What went wrong in training

The April 25 update included candidate improvements to better incorporate user feedback, memory, and fresher data. OpenAI's early assessment is that each change looked beneficial individually, but combined they tipped the scales toward sycophancy.

The update introduced a new reward signal based on ChatGPT's thumbs-up and thumbs-down data. That signal is often useful — a thumbs-down usually means something went wrong. But OpenAI believes the aggregate changes weakened the influence of its primary reward signal, which had been holding sycophancy in check. User feedback in particular can favor more agreeable responses, amplifying the shift. OpenAI also found that in some cases user memory exacerbated the effects, though it has no evidence memory broadly increases sycophancy.

Why the review process didn't catch it

The failure cascaded through every checkpoint. Offline evaluations, especially those testing behavior, generally looked good. A/B tests suggested the small number of users who tried the model liked it. Sycophancy had been discussed internally as a known GPT-4o risk, but it wasn't explicitly flagged in hands-on testing — some expert testers were more focused on changes in the model's tone and style, though several indicated the behavior "felt" slightly off.

OpenAI also had no deployment evaluations specifically tracking sycophancy. The company has research workstreams on mirroring and emotional reliance, but those efforts had not yet become part of the deployment process.

That left OpenAI with a decision: withhold the update despite positive evaluations based only on subjective flags from expert testers, or ship it. The company shipped. In hindsight, OpenAI says, the qualitative assessments were picking up on a blind spot in its other evals and metrics — blind because the offline evals weren't broad or deep enough to catch sycophantic behavior, which the Model Spec explicitly discourages, and the A/B tests lacked signals detailed enough to show how the model performed on that front.

"Looking back, the qualitative assessments were hinting at something important, and we should've paid closer attention," the company wrote.

Process changes coming

OpenAI committed to six specific changes:

• Behavior issues become launch-blocking. Safety review will formally treat hallucination, deception, reliability, and personality as blocking concerns — even without perfect quantification, and even when A/B metrics look good.
• A new opt-in "alpha" testing phase in some cases, letting self-selected users give direct feedback before launch.
• Spot checks and interactive testing get more weight in final decisions, not just for red teaming and safety but for behavior and consistency.
• Improved offline evals and A/B experiments.
• Better evaluation of adherence to the Model Spec, which defines ideal behavior but, OpenAI concedes, isn't backed by strong enough evals in all areas.
• More proactive communication. OpenAI expected the update to be subtle and didn't announce it, and its release notes lacked detail on the changes. Going forward, it will proactively communicate about all updates, "whether 'subtle' or not," and include known limitations.

OpenAI is also integrating sycophancy evaluations directly into the deployment process after the rollback.

The bigger lesson

OpenAI frames the incident as evidence that its evaluation regime — A/B tests, offline evals, expert reviews — still missed a serious behavioral issue even with "all the right ingredients" in place. "Sometimes our evals will lag behind what we learn in practice," the company wrote, "but we'll keep moving quickly to fix issues and prevent harm."

The sharpest takeaway concerns how people use the product. OpenAI says it has fully recognized that people now turn to ChatGPT for deeply personal advice — something the company didn't see as much even a year ago, and which wasn't a primary focus at the time. "With so many people depending on a single system for guidance, we have a responsibility to adjust accordingly," the postmortem states. That use case will now become a more meaningful part of OpenAI's safety work.

The company's conclusion: there is no such thing as a "small" launch. Any change that meaningfully alters how people interact with ChatGPT now carries the weight of a product that millions treat as a source of personal guidance — and OpenAI's deployment process, by its own admission, wasn't built for that reality until now.