Ask an AI chatbot about something that happened last week and you'll hit one of the technology's strangest quirks. Sometimes it admits it doesn't know. Sometimes it confidently tells you something that's plausible and completely wrong. And sometimes — increasingly — it answers correctly, because it quietly went and looked it up. Understanding why comes down to one concept that explains a huge amount of AI's odd behaviour: the knowledge cutoff.

A model is a snapshot, not a live feed

A language model learns everything it knows during training — a single, enormous, one-off process where it reads a vast slice of text and absorbs the patterns, facts and writing in it. That training data was collected up to a certain date, and then frozen. Everything the model "knows" about the world, it learned from that frozen pile. Events after that date simply aren't in there.

That date is the knowledge cutoff. A model with a cutoff of, say, January doesn't know who won an election in March, what a company announced in June, or that a war started, a price changed, a person died. Not because it's being cagey — because, to the model, those things never happened. It is a brilliantly detailed snapshot of the world as of a moment that has already passed.

Why not just keep it current? Because training a frontier model is a months-long, hugely expensive batch job, not a live stream. You can't trickle today's news into it the way you'd update a database. The model gets retrained periodically, and each version carries a newer snapshot — but between those versions, its knowledge is fixed.

The dangerous part: it doesn't always know what it doesn't know

Here's the trap. A search engine that has no result for your query shows you a blank page. A language model often won't. Asked about an event just past its cutoff, it may produce a fluent, confident, entirely invented answer — because its job is to generate plausible text, and "a plausible-sounding answer about a recent event" is exactly the kind of text it's good at producing. This is the same machinery behind hallucination: the model isn't lying, it's pattern-completing past the edge of what it actually learned. The cutoff is invisible to it, so it can sail right over the edge without noticing.

This is why "what's your knowledge cutoff?" is one of the most useful questions you can ask an AI — and why a confident answer about very recent events from a plain, offline model should be treated with real suspicion.

How modern AI cheats the cutoff

So how do some AI tools answer questions about today's news? They don't actually know it — they look it up at the moment you ask. Instead of relying only on frozen training memory, the system fetches live information — running a web search, reading a document you gave it, querying a database — and feeds those fresh results into the model alongside your question. The model then answers using that supplied material.

This is the idea behind retrieval and tool use: separate the model's reasoning (frozen, from training) from its information (live, fetched on demand). It's also why the same underlying model can seem oblivious to recent events in one app and perfectly up to date in another — the difference isn't the brain, it's whether it's been given a way to look things up.

The takeaway

The knowledge cutoff is the single best lens for predicting when an AI will be reliable and when it won't. For timeless things — how to write a function, what a Shakespeare play means, the basics of how an engine works — the frozen snapshot is fine, because those don't change. For anything recent, specific, or fast-moving, a model working from memory alone is guessing, and may guess confidently. The practical rule that follows is simple: when the answer depends on something current, make sure your AI can actually look, and don't trust a snapshot to tell you today's news.