What is context rot in AI?
Quick answer
Context rot is the way an AI model's answers get worse as the text you give it grows. In a long chat or a huge pasted document, it starts missing details, mixing facts from different parts, or repeating old mistakes, well before it hits its official context window limit.
Last updated
Take the free module →Module 6 of the free ChatGPT for Everyday Work course, about 9 minutes. No account needed.
Updated · By Robert Breen
Why it matters for a small business
A large context window says how much a model can take in, not how well it uses all of it. A 2023 study led by Stanford researchers, "Lost in the Middle", found that models used information at the start and end of a long input more reliably than information in the middle. Later tests, including a 2025 report from the company Chroma that helped popularize the name context rot, found quality slipping as inputs grew.
In practice it looks like this: a chat that has covered three customers quotes the wrong headcount, or an assistant keeps returning to an idea you rejected an hour ago. The fix is rarely a bigger model. It is less text: one task per chat, only the facts the job needs, and a fresh start when things drift.
In a real lesson: Stop Losing Context: When to Start a New Chat vs Keep Going using ChatGPT
Stop Losing Context opens with the problem: some days ChatGPT carries the last customer's headcount into the next quote, and you spend ten minutes arguing with a long, confused chat. The office manager keeps the Oakwood Library quote and its 100-guest change in one chat, because that is one job.
For Pine Ridge Rotary she clicks New chat and attaches only their facts file, so nothing from the library can leak in. The lesson doesn't use the term context rot, but its rule is the cure: same customer and same task, stay in the chat; new customer, new task or a bad start, open a new chat with just the facts for that job.
Try this lesson free or read the step-by-step guide.
Common confusions
Context rot vs context window
The context window is the hard limit on how much text fits. Context rot is the quality drop that starts well before that limit.
Context rot vs model drift
Model drift is a model's behavior changing over weeks or after updates. Context rot happens inside one long chat or input.
Tips
- Give each customer or task its own chat.
- Paste only the pages or rows the job needs, not the whole file.
- When a chat drifts, ask for a short summary, check it, and start a new chat with it.
Related terms
More AI basics terms
Where you use it: free lessons
Frequently asked questions
- Does a bigger context window fix context rot?
- Not by itself. It lets more text in, but the model can still use the middle of a long input less reliably.
- How long is too long for a chat?
- There is no fixed number. Start fresh when answers begin mixing details, missing facts you gave, or repeating rejected ideas.
- Where should I put the most important instruction?
- Near the start or the end of a long prompt, and repeat key rules in your latest message if the chat is long.