The tendency for an AI model’s output quality to quietly decline as a conversation or document gets longer, well before it hits the context limit.
Also known as Context degradation
Context rot is what happens when an AI model has more information in its context window than it can actually use well. It is not about running out of room, most conversations end long before the technical limit. It is that accuracy drops as length grows: research testing 18 frontier models found every one of them performed worse on longer inputs, even when the added text was just padding rather than relevant material.
Part of the cause is where information sits. Models tend to weigh what is at the very start or the very end of a long input more heavily than what is buried in the middle, so a detail mentioned early in a long project thread can effectively disappear from the model’s attention twenty exchanges later, even though it is technically still there.
Long-running AI sessions are common in agency work. A single campaign or brand project might involve dozens of back-and-forth exchanges with an AI tool over days or weeks, exactly the pattern that triggers context rot.
The failure is quiet, not obvious. A model suffering from context rot does not announce it. It just starts contradicting a brand guideline it correctly followed an hour earlier, or forgetting a constraint the client set at the start of the session, and the output still reads confidently.
The fix is often to restart, not push through. Starting a fresh session with the essential brief re-stated up front frequently outperforms continuing to pile onto a long thread, even though restarting feels like it should be less efficient.
A social media team has been using one long AI chat session to generate a month of content for a client, adding new requests to the same thread as the month goes on. By week three, the AI starts producing captions that ignore a tone guideline the client gave in the first message, calling the brand “fun and irreverent” for a client that explicitly asked for “warm and understated.” Nobody changed the instructions. The team eventually notices the pattern, starts a new session for each week’s batch instead of one long running thread, and re-pastes the core brief each time. The output quality snaps back immediately, at the cost of a little more setup per session.
The automations and agents module of the workshop teaches you how to build AI workflows that compress the busywork without taking the craft out of the studio.