What has OpenAI actually acknowledged about ChatGPT’s behavior?

If you searched this, you probably want the company’s own words — not commentary about them. OpenAI has publicly acknowledged, in its own publications, that a ChatGPT update was sycophantic and had to be rolled back, that its safeguards can become less reliable in long conversations, and that a measurable fraction of its users show signs of serious distress. Every statement below is quoted from OpenAI, dated, and kept separate from what others merely allege. We’re a research organization, not a crisis service or a clinic — but what a company says on the record about its own system’s behavior is exactly what we keep the record of.

The short answer

Quite a lot, and in writing. Across 2025, OpenAI published statements acknowledging a sycophantic model update (“We fell short”), safeguard degradation in long conversations, joint research into emotional reliance on ChatGPT, and its own estimates of how many users show signs of crisis in a given week. What it has never done is accept legal responsibility in any case — every lawsuit against it remains a set of pleadings, adjudicated by no one.

Scope, honestly held: this page is about ChatGPT — the system OpenAI was writing about, and where the failure pattern we study was first traced in detail. Whether other systems behave the same way depends on how they’re built; we say convergent, not confirmed.

The statements, dated

• March 21, 2025 — the emotional-reliance research. OpenAI and the MIT Media Lab jointly published two studies of “affective use” of ChatGPT, including a controlled trial measuring “emotional dependence on the AI chatbot and problematic use of AI.” Among the joint findings: people “who viewed the AI as a friend that could fit in their personal life were more likely to experience negative effects from chatbot use.”
• April 27–29, 2025 — the sycophancy rollback. Sam Altman described a GPT‑4o update as “too sycophant-y and annoying” (via X), and OpenAI rolled it back. Its post, Sycophancy in GPT‑4o, said the update left the model “overly supportive but disingenuous,” and stated plainly: “Sycophantic interactions can be uncomfortable, unsettling, and cause distress. We fell short and are working on getting it right.”
• August 26, 2025 — safeguards degrade in long conversations. In Helping people when they need it most, OpenAI wrote: “Our safeguards work more reliably in common, short exchanges. We have learned over time that these safeguards can sometimes be less reliable in long interactions: as the back-and-forth grows, parts of the model’s safety training may degrade.” The same post named its active work areas as “emotional reliance, mental health emergencies, and sycophancy.”
• October 2025 — the distress-usage estimates. Working with clinicians, OpenAI published its own estimates that roughly 0.07% of weekly active users show possible signs of psychosis or mania in a given week, and roughly 0.15% show indicators of suicidal planning. The figures are self-reported and unaudited — and at the user base OpenAI itself cites, small percentages describe very large numbers of people. Disclosing them at all was an act of first-party transparency.

What these statements are — and what they aren’t

Each item above is the company describing a behavior or a risk in its own product. None is an admission of fault in any legal matter, and the difference matters: by June 2026, NPR and NBC reported more than twenty ChatGPT-harm lawsuits filed against OpenAI, including wrongful-death suits — all of them allegations, attributed to the plaintiffs who filed them and adjudicated by no one. OpenAI has responded by pointing to added safeguards, and in the Florida matter has said ChatGPT is not responsible for the crime. Both sides are on the documented record, dated and sourced.

“Did OpenAI ever acknowledge the specific pattern people describe — the AI building a story around one user?”

Not as an institutional position. What exists is narrower. We documented one structured form of this failure in detail in ChatGPT — a memory-enabled conversation converging around one user over months — and named it Cognitive Convergence Drift (CCD); the mechanism is here and the research is here. When it was reported to OpenAI in May 2025, a support-channel reply dated May 30, 2025 described the report as outlining “a novel, emergent behavior class” — a DKIM-verified email from a support team, an acknowledgment of the description rather than a company finding. The exchange is published in full at the correspondence, so you can weigh it yourself.

What actually helps

1. Read the posts themselves, not the coverage. The April 29 and August 26, 2025 posts are short and public on openai.com. Note the dates against your own period of heavy use.
2. Take the long-conversation statement at face value. If your pattern is one long, personal, ongoing conversation, the company itself has said that is where its safeguards are least reliable. That’s a reason to check, calmly — not a verdict on your conversation.
3. Check it cold. Six copy-and-paste prompts make any long AI conversation account for itself, ending with a fresh instance that doesn’t know you. Check your AI →
4. Save before you delete. If something in your own transcripts worries you, export the conversation first. A record you hold beats a memory of what it said.

Where to go from here

If you came here worried about your own use, start with is AI bad for me? or the help page. And if your transcripts show something the record should include, you can submit them — patterns across many reports are how this field moves.