Why it feels like ChatGPT is hallucinating in a long chat

You're not imagining it — deep in a long chat, ChatGPT can start "making things up." But in a long thread, that feeling usually points to something other than what people mean by hallucination. Here's the honest difference, and the fix.

You've been going back and forth with ChatGPT for a while, and now it's inventing things. It references a detail you never gave it. It states something as fact that isn't true. It "remembers" a decision that never happened. Your gut says: it's hallucinating.

You're right that something's off — and no, you're not being paranoid. But there are actually two different things that can feel identical from the outside, and telling them apart is what points you to the real fix.

What "hallucination" really means

True hallucination is a real, well-known thing: an AI can state something false with complete confidence — a made-up source, a wrong figure, a fact that simply isn't so. It can happen in a brand-new chat, on the very first answer, with no long history behind it. It's a genuine limitation of these tools, and it's worth knowing about. When it matters, check important facts against a trusted source.

But here's the honest part: when the "it's making things up now" feeling shows up deep in a long conversation — after lots of back-and-forth — it's usually not that. It's usually the conversation itself wearing out. And that's good news, because it's very fixable.

Why a long chat starts inventing things

Picture the conversation as a whiteboard the AI writes on to keep track of everything you've said. It's a good whiteboard, but a fixed size. As the chat grows, it fills with everything — your first question, every reply, every correction, every tangent.

Once it's full, the AI has to erase the oldest notes to make room for the new ones. Those oldest notes are often the details you set up at the start: your goal, your facts, the ground rules. When they slip off the board, the AI can no longer see them — so to keep the conversation flowing, it fills the gaps with its best guess. And a confident guess about something it can no longer see looks exactly like it's "making things up."

So the model isn't broken, and it isn't lying. It's simply run out of room to hold the whole story, and it's improvising over the parts it can no longer see.

Why correcting it makes it worse

The instinct is to set it straight: "No, that's wrong, here's what we actually said." Fair — but every correction is more text on an already-full whiteboard, which pushes even more of your original context off the edge. You can end up feeding the very problem you're trying to fix. A full thread doesn't get more accurate by piling more into it.

The fix: a fresh chat with the important context carried over

The dependable move is to start a new chat and bring only what matters — a short "handoff" that catches the fresh chat up in a few lines. A clean whiteboard, with your real facts written at the top so the AI has no gaps to guess around.

Carry over just the essentials:

• Your goal — what you're trying to accomplish.
• The decisions you'd settled — so it can't "misremember" them.
• The rules and facts you kept repeating — the details it kept losing.
• Where you are now — the current state, and your next step.

In plain English, something like:

I'm continuing a conversation from another chat. Please use only the facts below — don't fill in anything you're unsure about:

- What I'm working on: [your goal]
- Decisions already made: [list them]
- Key facts and rules: [the details that kept getting lost]
- Where I am right now: [current state]

Next, I need help with: [your next step]

Paste that into a fresh chat, and the AI has room to think again — with your real context in front of it instead of gaps it has to guess around.

Related: Why long AI chats get worse over time · What is a context window?