Why AI gets stuck on its first idea
Why does AI keep trying the same wrong fix?
Every failed attempt stays in the chat, and AI keeps building on it.
The loop
You: That didn't work.
It: Apologies! Here's the corrected version.
Same problem. Again.
Five rounds later, it's suggesting the thing you rejected in round two.
Titian, Sisyphus (1548–49). Pushing the same boulder up the same hill, forever.
Every wrong attempt stays on the desk
Every failed try is still in the conversation, and it re-reads all of them each time (see "Why your AI usage runs out so fast").
AI is also built to continue patterns. Five attempts heading in the same direction make a strong pattern, so the sixth tends to head the same way.
The name for it: context poisoning
Context poisoning: a mistake gets into the conversation and the model keeps building on it.
When Google had its Gemini model play Pokémon, it sometimes became convinced of something untrue about the game and spent hours chasing it. One time it went after an item that doesn't exist in the version it was playing.
The wrong belief was written into its own notes, and it kept reading it back.
It happens on long tasks too
In 2026, several research teams separately found the same thing in AI agents working on long jobs. The agent commits to an early theory and keeps pushing it, long after the evidence says to stop.
Drew Breunig calls this context inertia.
The fix they all landed on is to wipe the conversation regularly and start again, keeping only the useful results.
Basically, "have you tried turning it off and on again?"
Arguing makes it worse
"No, that's wrong, stop doing that" puts more text about the wrong idea on the desk.
Pushing harder rarely helps. Clearing the desk does.
Try this: the two-strikes rule
If two fixes for the same problem have failed, stop.
- Ask: "Summarise the problem, everything we've tried, and why each attempt failed."
- Open a fresh chat.
- Paste the summary and add: "Don't try these again."
You keep what you learned and lose the clutter.
The two-strikes rule, as a picture
What a good handover looks like
Problem: The monthly report shows the wrong total for March.
Tried: 1) Changed the date filter: same total. 2) Removed duplicate rows: same total.
What we know: The raw data adds up correctly. So the problem is somewhere in the report step.
Don't suggest the date filter or duplicates again.
Sources
- Drew Breunig, "How contexts fail and how to fix them", Jun 2025: dbreunig.com/2025/06/22/how-contexts-fail-and-how-to-fix-them.html
- Drew Breunig, "Context inertia", Oct 2026: dbreunig.com/2026/10/14/context-inertia.html
- Drew Breunig, "The state of context engineering", Oct 2026: dbreunig.com/2026/10/24/the-state-of-context-engineering.html