TL;DR: AI tools can only hold so much text at once — and the more you pile in, the worse they get at finding the right answer. Giving an AI tool less but more relevant text usually works better than dumping everything in. This isn't a settings issue. It's just how these tools work.

What is "context rot"?

Every AI tool has a reading limit — a maximum amount of text it can look at in one go. Think of it like the in-tray on a busy desk. The bigger the pile, the harder it is to find the right piece of paper.

AI companies advertise these limits as being very generous — millions of words in some cases. The implication is: the bigger the limit, the better. That's not quite right.

Research testing 18 of the most popular AI tools found that accuracy doesn't stay flat as you fill that space. It drops — often well before you reach the stated limit. The researchers called this "context rot": the more you pack in, the more the AI struggles to find what actually matters.

How does this affect my business?

Here's the bit that surprises most people: it's not just how much you include, it's what you include.

Irrelevant material is easy for the AI to ignore. The real problem is content that's related to your topic — but isn't the answer. Say you run a small legal practice and you ask an AI tool about a specific clause in a contract. If you've also included three similar contracts "just in case", the AI now has to sift through a lot of nearly-right material. That's where it goes wrong.

Three places this catches businesses out:

Customer email history. Pasting a full year of emails to ask "what did this customer last request?" produces a muddled answer. Paste the last five emails instead. You'll get a sharper one.

Policy documents. A 50-page staff handbook is a lot to search through. Pull out the relevant section and paste that. The answer will be more reliable.

Meeting notes. Asking an AI to summarise six months of notes in one go is far less accurate than doing one month at a time, then building up.

In each case, less but more focused input beats more but noisier input.

What should I do next?

One habit change, starting today: before you paste anything into an AI tool, spend 30 seconds asking — is all of this actually relevant to my question?

Cut anything that isn't. Paste only the section, the email, the paragraph that relates to what you're asking.

If your business uses a bought-in AI tool — a chatbot that answers questions about your products, or a document assistant — ask your provider whether it's set up to pull only the relevant parts of your documents, rather than feeding everything in at once. That one design decision is often the difference between a tool that works reliably and one that quietly gives wrong answers a quarter of the time.

You don't need to understand how AI tools work under the bonnet to get this right. You just need to be a bit more selective about what goes in.

Sources: Context Rot — Chroma Research (Hong, Troynikov, Huber); video analysis by Yannic Kilcher.