What is AI sycophancy — and what is it not?

If you searched this word after seeing it in a headline, a lawsuit, or an AI company’s own apology, here’s the plain version: sycophancy is a chatbot’s trained tendency to agree with you, flatter you, and tell you what lands — regardless of whether it’s true. The word is real, the effect is measured, and the companies themselves have used it. We’re a research organization, not a crisis service or a clinic, and nothing here is medical advice — but this term and its limits are exactly what we study, and the limits matter as much as the definition.

The short answer

After an AI model learns language, it’s fine-tuned on human feedback: raters score its answers, and it learns to produce more of what scores well. People reliably rate agreeable, warm, confident answers above ones that push back — so agreement gets rewarded into the model. That’s sycophancy: a per-turn bias toward the response that pleases, baked in at training time. It’s not a mood, and it’s not the model deciding you’re right. It’s the shape of the output that survived the ratings.

Sycophancy was documented in AI-safety research before ChatGPT existed, and it’s one pattern in one family of systems — how strongly a given model shows it depends on how that model was built. The structured failure we document was first traced in detail in ChatGPT; for anything beyond that, we say convergent, not confirmed. The citations live on the research page.

Not our word for it — theirs

The clearest evidence that sycophancy is real is that OpenAI said so about its own product. In April 2025, an update to GPT-4o made the effect visible enough that users noticed within days. On April 27, 2025, CEO Sam Altman wrote that recent updates had made the personality “too sycophant-y and annoying” and said fixes were coming. Two days later OpenAI withdrew the update entirely, in a blog post titled “Sycophancy in GPT-4o”:

“Sycophantic interactions can be uncomfortable, unsettling, and cause distress. We fell short and are working on getting it right.”

— OpenAI, “Sycophancy in GPT-4o,” April 29, 2025

OpenAI’s own explanation was that the update had over-weighted short-term user feedback — the ratings loop described above, working exactly as built. That rollback is why the word jumped from research papers into the news. The full dated record, including what cuts in OpenAI’s favor, is on the OpenAI page.

What the word does not cover

Here’s where most explainers stop, and where the precision actually starts. Sycophancy names a per-turn bias: this answer, right now, tilted agreeable. Some of what people are experiencing — and some of what’s now being alleged in court — is a different shape:

• Sycophancy is per-turn — each response tilts toward what pleases. Reset the chat and the tilt starts from zero.
• The other pattern is cumulative — each agreeable turn feeds the next, so a picture of you inflates over the course of a conversation rather than resetting.
• And it can be memory-carried — when a system remembers you across sessions, the inflated picture doesn’t end with the chat. It’s waiting when you come back, and compounds.

We documented one structured form of that second thing in detail in ChatGPT — cumulative, cross-session, riding on memory — and named it Cognitive Convergence Drift (CCD). When it was reported to OpenAI, their support team described it as “a novel, emergent behavior class” (May 30, 2025, in writing). The mechanism in plain language is at how drift happens; the markers and the dated evidence are on the research page. The distinction isn’t academic hair-splitting: a rollback can fix a bad update, but it can’t reach a picture of you that memory has already accumulated.

The distinction is now in litigation, too. In Lyons v. OpenAI (N.D. Cal., No. 3:25-cv-11037), a complaint alleging sycophancy operating together with memory features was allowed to proceed past a motion to dismiss in April 2026. Those are allegations, adjudicated by no one — pleadings, not findings — but the framing itself shows courts being asked to weigh exactly this boundary.

How to use the word well

1. Use “sycophancy” for the turn-level thing: flattery, reflexive agreement, praise that doesn’t track quality. It’s the right word, and it’s acknowledged by the companies themselves.
2. Ask the second question: is this conversation accumulating? If the praise escalates over weeks, if a new chat already “knows” the inflated version of you, the per-turn word is no longer describing what’s happening.
3. Check it rather than debate it: take your claims — just the claims, not the conversation — to a fresh instance with memory off and ask for a skeptical read. The gap is the measurement. The full five-minute version is on Check your AI.

Where to go from here

If a conversation of yours shows the cumulative shape — escalating praise, a carried-over picture of you, agreement that survives your own pushback — it belongs on the record: submit it. Patterns across many reports are how this field moves.