How current is ChatGPT?

Ask a chatbot about this morning’s news and you may get a confident answer that is months out of date. That isn’t random. Every AI chatbot has a built-in horizon called the knowledge cutoff, and whether an answer is current depends on which side of that horizon the facts fall — and on whether the chatbot went out to the live web before answering. This page explains where the cutoff comes from, when a chatbot can and can’t browse, and how to tell which kind of answer you’re looking at.

The short answer

A chatbot like ChatGPT learned what it knows from an enormous collection of text gathered up to a certain date — its knowledge cutoff. Events after that date are simply not in its memory. Many chatbots can now supplement that memory with a live web search, and when they do, the searched parts of the answer can be genuinely current. When they don’t search — and much of the time they don’t — you are hearing from the snapshot, however old it is. Both kinds of answer arrive in the same fluent, confident voice, so the reliable habit is to check which one you got.

Why there’s a cutoff at all

An AI model isn’t connected to a rolling feed of the world. It is trained: built in one long, expensive process from a fixed collection of text, and that collection stops at a date. Think of an encyclopedia printed on a particular day. Everything up to the printing date can be in it; nothing after can be. Between one version of a model and the next, its built-in knowledge stands still — it doesn’t quietly absorb the news overnight.

Two useful details follow from this. First, the cutoff is a real, stated fact about each model — the company publishes it, and you can ask the chatbot its cutoff — most major products will tell you, though models sometimes misstate it, so the company’s published model documentation is the authoritative source. Second, the model has no clock of its own. It doesn’t inherently know today’s date; the product usually supplies the date behind the scenes. So a chatbot can know what day it is and still know nothing about what happened last week — those are two different mechanisms, and only one of them updates. The how AI is trained page walks through where that built-in knowledge comes from.

When it can see the live web — and when it can’t

Many chatbots can now run a web search in the middle of answering. That changes what “current” means, but only for the parts it actually looked up:

• Searching — the chatbot pauses, runs a search, and reads a handful of live pages before answering. You’ll usually see a visible sign of it (a “searching the web” indicator) and the answer arrives with source links. Those cited parts can be as current as the pages it read.
• Not searching — the chatbot answers from training memory alone. No indicator, no links, no sources. Whatever it says about recent events is coming from the snapshot, which ends at the cutoff.
• The mix — a search fetches a few specific pages, not the whole web. One answer can be current on the fact it looked up and months old on everything surrounding it. The links show you exactly which parts were checked live.

Not every product, plan, or setting includes browsing, and a chatbot with browsing available won’t always choose to use it. The presence of the feature isn’t the same as its use on the answer in front of you — which is why the check below is about the individual answer, not the product.

How to tell whether an answer is current

The tone of an answer carries no timestamp: stale information is delivered in exactly the same confident voice as fresh information. So don’t read the confidence — read the signals:

1. Look for sources. If the answer includes links to live pages — especially dated ones — those parts were checked against the web. No links usually means no search.
2. Ask it to search. A plain request works: “search the web for the latest on this.” If the product can browse, this is the direct way to get a current answer instead of a remembered one.
3. Ask its cutoff. “What is your knowledge cutoff?” usually tells you where the snapshot ends — but models sometimes misstate it, so check the company’s published model documentation if it matters. Anything that happened after that date, it can only know by searching.
4. Treat time-sensitive facts as live-check facts. Prices, laws, software versions, schedules, medical guidance, who holds an office, anything described as “the latest” — these change fast enough that a training snapshot is the wrong source even when it sounds sure. Verify them against a live page before you act on them.

One more pattern worth knowing: asked about something after its cutoff, a model doesn’t always say “I don’t know.” It may extrapolate from what it does know and present the result as fact. That failure mode has its own page — why AI makes things up — and recent events are one of the places it shows up most.

Where to go from here

Spot something here that’s out of date or could be clearer? Tell us — this is an education resource, and it only earns trust by being checkable.