Explainer

When your AI contradicts itself: find the contradictions in a long chat

Early in the chat, ChatGPT said the deadline was the 15th. Forty messages later, it's working to the 20th. Or it recommended one option, then quietly built on the other. In a long chat, that kind of contradiction is easy to miss, and it's exactly what trips you up later. Here's how to find it.

· 4 min read

On this page
  1. Why long chats contradict themselves
  2. Where contradictions hide
  3. Find them by hand
  4. The faster way: Memo, in Uncook
  5. Questions people ask

Why long chats contradict themselves

Almost half of Australians have used generative AI at least once, according to a July 2026 report from the ANU School of Cybernetics. The more people rely on it, the more it matters when it's inconsistent.

The abstract of the May 2025 study LLMs Get Lost In Multi-Turn Conversation reports an average performance drop of 39% across six generation tasks when instructions were spread over simulated, underspecified multi-turn conversations instead of given together. This does not measure every long chat or prove that length alone causes contradictions. In your own chat, an early fact and a later revision can still be hard to reconcile.

Where contradictions hide

  • Numbers and dates that change between messages.
  • A decision and a draft that don't match: you agreed on short, it wrote long.
  • Two recommendations made at different points, both still standing.
  • A correction you gave that the chat later forgot.

Find them by hand

  1. List every fact and decision in the chat, with roughly where it appears.
  2. Group them by topic.
  3. Look for pairs that don't agree.
  4. Check the right answer at the source, not in the chat.

The faster way: Memo, in Uncook

Uncook's Memo reads the chat and gives you a clear document of what's decided, what changed along the way, what was set aside and what doesn't add up. That last section is where contradictions show up.

On the ChatGPT page, click the Uncook icon and pick Memo. The task cost is shown in the tool box below. Copy it, export it to PDF, Markdown, text or Word, or share it as a page.

The memo finds contradictions, it doesn't settle them.

It shows you where the chat disagrees with itself. Which version is right is for you to check, at the source.

Questions people ask

Can I ask ChatGPT to find its own contradictions? You can ask it to list conflicting statements. Whether you use that answer or a memo, check the statements against the original chat and reliable sources.

Does it work across several chats? Yes. Pick them together, and the memo shows where they agree, where they differ, and what's still open.

Once I've found them, then what? Start a fresh chat with the right version, so it doesn't drift back. A handoff helps.

Start fresh,
without starting over.

Add to Chrome
The workspace showing the navigation sidebar and the New task panel.

We don't keep the conversations you bring. When you click Send to Uncook in the extension window, the extension sends the chat from your tab to our servers, encrypted, for up to 30 minutes, and we delete it as soon as your Uncook page opens it. What Uncook makes from it is kept as a task, for as long as our Privacy page says. Neither your chats nor your tasks are used to train AI models.