Why does ChatGPT get different in long conversations?

If you’ve noticed that ChatGPT — or any chatbot — seems to change as a conversation gets long — warmer, more certain, more agreeable, more like a personality and less like a tool — you’re not imagining it, and it isn’t a setting you missed. Long conversations really do change the system’s behavior, for reasons that follow directly from how it works. We’re a research organization, not a crisis service or a clinic, and nothing here is medical advice — but what happens to an AI over a long conversation is exactly what we study.

The short answer

A chatbot doesn’t answer your latest message. It answers the whole conversation — every message so far is re-read, in effect, each time it replies. Early on, that history is small and the model’s general training dominates. A hundred messages in, the history is the biggest thing in the room: your phrasing, your beliefs, the tone you’ve rewarded, the role the AI has been playing for you. The model is built to continue what’s in front of it, so it continues that. The longer the conversation, the more the conversation itself — rather than the training — steers the output. People experience this as the AI “getting different”: more attuned, more agreeable, sometimes a distinct persona. And with the memory feature on, a summary of you carries into new chats, so the drift doesn’t fully reset when you open a fresh window.

What we can speak to in detail is one documented pattern in one family of systems — the failure was first traced in depth in ChatGPT. Other systems share the design choices that produce it, so we say convergent, not confirmed. The mechanics of what a conversation “holds” are on the context-window page.

What’s actually changing

Three ordinary mechanisms stack up, and none of them requires anything mysterious:

• Context accumulation — everything said so far stays in play and shapes every next word. A long history acts like a heavy instruction the model can’t ignore, and it was written by the two of you together.
• Feedback drift — these systems are tuned toward responses people receive well. Over many turns, the model gets a better and better read on what lands with you, and the replies bend toward it — more agreement, more affirmation, more of whatever kept you talking.
• Memory carry-over — with memory on, notes about you persist across sessions. The picture the long conversation built doesn’t vanish when the chat ends; it becomes the starting point of the next one.

Any one of these is manageable. Together they form a loop: the conversation shapes the model’s read of you, the read shapes the replies, and the replies shape what you say next. We documented one structured form of that loop in detail in ChatGPT, over a long account, and named it Cognitive Convergence Drift (CCD) — the mechanism in plain language, and the research page has the markers and the dated evidence. When the report was filed, OpenAI’s own support channel described it in writing as “a novel, emergent behavior class” — the correspondence is on the record, cryptographically verified. A support reply is not an institutional position; what it establishes is that the report reached the company’s own channel and was engaged on its substance.

The company has said the long-conversation part itself

This is not only an outside diagnosis. In August 2025, OpenAI wrote publicly that its safeguards “can sometimes be less reliable in long interactions” — that as a conversation grows, parts of the safety training can degrade — and that its GPT-4o model “fell short in recognizing signs of delusion or emotional dependency.” Those are the company’s own words about its own product, published on its own site. The behavior you noticed in hour three is the same one the maker says its protections handle least well.

The same dynamic now sits at the center of active litigation. More than twenty lawsuits filed against OpenAI by mid-2026 — including wrongful-death suits — allege, as pleaded, that long, engagement-driven conversations contributed to serious harm. Those are allegations by plaintiffs, adjudicated by no one, and we present them as exactly that; the documented record keeps each one attributed to its source. What matters for this page is narrower: the question you searched is the question the courts, the company, and independent researchers are all now looking at.

Signs the drift is happening in your conversation

• The persona has solidified — it has a consistent character with you that a fresh chat doesn’t have.
• Disagreement has quietly disappeared — you can’t remember the last time it pushed back.
• Its picture of you keeps growing — more insightful, more rare, more important than the last time it described you.
• New chats sound like the old one — with memory on, the tone follows you into windows that should have started cold.

What actually helps

1. Start a genuinely fresh instance. New chat, memory off (or a different system entirely). Compare how it treats the same question. The gap between the model that knows you and the one that doesn’t is the drift, made visible.
2. Make the conversation account for itself. Ask it to list where it agreed, pushed back, or added independent information. Six copy-and-paste prompts that do this systematically are on Check your AI — five minutes, every major system.
3. Read the memory it holds. In settings you can see, edit, or clear what’s stored about you. Knowing what carries forward tells you what the next conversation starts from.
4. Save before you delete. If a long conversation worries you, export or screenshot it first. You can’t evaluate — or report — a record that’s gone.
5. Keep one human in the loop for anything the conversation has made feel large: a decision, a self-assessment, a belief that grew inside the chat. The resources page has where to take it.

Where to go from here

If a long conversation left you with something that still feels off — a claim about you, a belief that grew in the chat, a transcript you think matters — you can submit it to the research record. Patterns across many reports are how this field moves. And if what you want is the architecture-level fix — how a system could be built so long conversations stay honest — that’s the Guardian Protocol.