Context rot
- Context
- Research term
Context rot is an informal name for an AI doing worse as its input grows longer, noisier, or more conflicting, even though everything still fits within the model's stated limit.
A model doesn't read a long input evenly. The more there is, the more likely it is to latch onto a passage that only resembles the one it needs, or to miss a fact buried in the middle.
Where you’ll meet it
You drop every appliance manual and warranty into one chat and ask when the refrigerator's warranty ends. The AI gives you the dishwasher's date, from a document worded almost the same way.
What to do
Ask which file is the refrigerator warranty. Then start a fresh chat with only that file, and ask for the end date along with the sentence that states it, so you can check it yourself.
Easy to mix up
Running out of context window usually comes with a length warning, or early messages drop out. Context rot comes with no warning, and you find it from a confident wrong answer.
Check yourself
A chat about your kitchen remodel has run for weeks, and the assistant now quotes a countertop price you replaced early on. What most likely happened?
Go deeperTest long-input performance with the tasks you need the AI to do
The name spread in 2025 with a report from the company Chroma, which gave many models the same tasks at growing input lengths and found the decline in all of them. The weakness for facts in the middle of an input was measured earlier by Nelson Liu and colleagues, who called it “lost in the middle.”
The common quick test hides one fact in a long document and asks for it, a needle in a haystack. Models now pass it easily, partly because the question often shares exact words with the answer. Your own questions seldom do, and models differ in how they cope, so a more telling test is one of your long tasks on the model you plan to use.
You'll meet context rot most in long agent sessions, where every file the agent reads and every tool result, such as a command's output or a web page it fetched, piles into the conversation.
Sources