Greta runs operations at a mid-sized logistics company. When her team started using an AI assistant last spring, it felt like a new colleague who never got tired. It drafted supplier emails, summarized meeting notes and answered questions about contracts in seconds.
By autumn, something had changed. The assistant was still fast and polite, but its answers had become slightly off. It quoted a delivery deadline from an old contract. It mixed up two suppliers with similar names. Once, asked a simple question about a client’s payment terms, it replied that it couldn’t find the information, even though Greta had discussed it with the assistant twice.
Nothing was broken. The assistant had simply been buried in noise.
What information noise is
In everyday life, noise is everything that competes with the information you actually need: the hundred unread emails hiding the one that matters, or the meeting notes where the key decision sits between small talk and agenda items.
For AI, noise takes three main forms. There’s irrelevant information, text that has nothing to do with the question. There’s outdated information, facts that were true once but no longer are. And the most dangerous kind: look-alikes, information that resembles the right answer but isn’t it, like a different supplier’s contract or last year’s price list.
Greta’s assistant was suffering from all three.
Why noise confuses AI
It’s tempting to think more information always helps an AI. Modern models can read the equivalent of several books at once, so why not give them everything?
Because research shows they don’t handle it equally well. One study tested 18 leading models, and found that performance degrades significantly with longer inputs, contrary to the assumption of uniform context processing. Researchers call this “context rot”, and it doesn’t wait until the limit is reached: performance degrades gradually as context fills with tokens, long before reaching any limit.
Look-alikes make it worse. In the same research, even a single piece of similar-but-wrong information significantly hurt performance. And the effect isn’t new. Back in 2023, researchers added small amounts of irrelevant sentences to simple math problems and found that model performance dramatically decreased when irrelevant information was included.
For Greta, this explained the mixed-up suppliers. Two contracts with similar names and similar wording are exactly the kind of noise that trips a model up.
How noise damages AI memory
The bigger problem was memory. Greta’s assistant didn’t remember conversations the way people do. Like most AI memory systems, it saved past conversations and, when a new question came in, searched them for the most relevant pieces to show the model. The model only sees what that search hands it.
In the first weeks, Greta’s history was small, and almost everything relevant made it through. After six months, the facts that mattered for any one question were a tiny fraction of everything stored. The rest was her own other conversations: same suppliers, same topics, similar words.
Tests of AI memory systems show how much this matters. When 90% of a person’s history is unrelated to the question, the best setups still answer over 95% of questions correctly. Others drop below 70%. Systems that simply look at the most recent conversations fall to about half. Most failures happen not because the model misreads the information, but because the right piece never reaches it. Facts mentioned only once are the most likely to get lost. And many systems keep old and new versions of a fact side by side, leaving the model to guess which one is current. That’s how Greta got the old deadline.
What Greta changed
Once she understood the problem, the fixes were surprisingly practical.
She split her work into separate projects, one per client group, so unrelated conversations stopped competing with each other. She started stating key facts clearly instead of mentioning them in passing, and when something changed, she said so explicitly: “The new payment terms from October replace the old ones.” She reviewed what the assistant had saved and deleted outdated entries. For anything critical, like contract figures, she pasted the source document directly into the conversation rather than relying on memory.
Her company’s IT team made changes too. They set up the assistant to search both by exact words and by meaning, to rank results by relevance rather than recency, and to show the model the date of every saved fact.
Within a few weeks, the answers were sharp again.
Signal over volume
Greta’s story points to a simple idea that’s easy to forget in a world of ever-larger AI models. Giving an AI more information doesn’t automatically make it smarter. What matters is how much of what it sees is signal, and how much is noise.
The best AI work, like the best human work, isn’t about knowing everything. It’s about having the right information in front of you at the moment you need it.