What is a knowledge cutoff — and why the model lies about yesterday

Ask a brand-new model about something that happened last week. Often you get one of two things: an honest "I don't know," or — far worse — a confident, smooth, completely made-up answer. Both have the same cause, and it's called the knowledge cutoff.

What a knowledge cutoff is

A model doesn't learn from "the whole internet right now." It learns from a snapshot of data gathered up to some point. That point is the knowledge cutoff. Anything that happened after simply never reached the model.

It's like a person who spent six months on an expedition with no signal: they come back and have no idea about the news, the new slang, or who won the championship. Not because they're slow. They just weren't there.

How it works

A language model is a "cast" of the texts it was trained on. Gathering data, training, testing — that's months. So even a fresh model shipped today usually "remembers" the world up to some month in the past, not up to today. The gap between the cutoff and the release date is normal, not a bug.

One key detail: the cutoff is about training, not the conversation itself. Within a single chat the model handles whatever you just gave it perfectly — that's its context. But that knowledge lives only until the chat ends and never gets written back into the model. The moment it generates a reply (inference) adds no new facts to its memory.

Why this matters to you

Here's the real trap. The problem isn't that the model doesn't know something — that's half the trouble. The problem is that it doesn't know that it doesn't know. Ask about yesterday's release, a stock price, or the latest version of a library, and it'll hand you a plausible answer with the same confidence it has about times tables. Only it's invented. This is a close cousin of a hallucination: the model always wants to give a smooth answer, even when staying quiet would be more honest.

What to do about it in practice:

• Don't ask for fresh facts from memory. News, prices, "what's the latest version" — that's not a job for its memory.
• Hand it the data in the request. Paste the text, document, or search results right into the prompt, and the model works with that instead of inventing. When this happens automatically (the model pulls in fresh documents itself), it's called RAG.
• Double-check anything about "now." If the answer hinges on "latest," "recently," "current version" — verify with the source. The model errs most on tool versions.

Where you'll notice it

You'll notice the moment you ask about something new. Sometimes the model warns you honestly: "my knowledge is limited to such-and-such date." Sometimes it doesn't, and it gives itself away indirectly: it "hasn't heard" of a thing that's already routine to you, or it mixes up versions.

Good interfaces add web search to the model — exactly to close the gap between the cutoff and today. If your tool has search, turn it on for anything fresh. And don't confuse the cutoff with fine-tuning: fine-tuning changes the model's behavior on new examples, but doesn't make it all-knowing about yesterday. Moving the knowledge boundary takes retraining on fresher data — and that's a new version of the model.

FAQ: Can I find out a model's exact cutoff date?

Sometimes yes — it's listed in the docs, or the model names it when asked directly. But don't fully trust that self-report: the model can be wrong here too. More reliable: test it. Ask about a couple of events with known dates and see where the boundary falls.

FAQ: Why not train the model in real time?

Because training is a long, expensive process on massive hardware, not jotting a line in a notebook. Running it for every new event is impossible. So "freshness" comes from search and feeding documents into context, not from retraining.

FAQ: If the cutoff is in the past, is the model useless for new topics?

No. Reasoning, explaining, writing code — it does all that regardless of the date. It only falls short where you need a specific fresh fact. Give it that fact in the request and it's back on its feet.