Why doesn't ChatGPT know about recent events?

Ask ChatGPT about something that happened last week and you will usually get one of two answers: a polite note that it has no information about recent events, or a confident paragraph that turns out to be invented. Both come from the same fact about how it was built. ChatGPT does not read the news. It never has.

The interface says so up front. Among the warnings in ChatGPT's web interface, The Verge notes, is one saying its knowledge of anything after 2021 is limited. Understanding why that is, and what it means for your questions, will save you from some of its most convincing mistakes.

A snapshot, not a feed

ChatGPT is a language model: a program that learned to predict text by studying an enormous collection of writing. That learning happens during training, a long and expensive process that runs once on a fixed set of text. When training ends, the model's knowledge is frozen. Everything it "knows" is a pattern it absorbed from that collection.

OpenAI says ChatGPT was fine-tuned from one of its GPT-3.5 models, and that model's training was completed early in 2022. The text behind it was gathered before that, and the interface points to 2021 as the year after which its knowledge thins out. Events after that simply are not in what it learned, or appear only as scraps.

A search engine is different. It visits web pages constantly and can show you an article published minutes ago. ChatGPT has no such connection. In one of OpenAI's own launch examples, the model explains that it cannot browse the internet or access external content. It works only from what it absorbed in training plus what you type into the current conversation.

Why it guesses instead of saying "I don't know"

The bigger risk is not silence but invention. OpenAI's launch post admits that ChatGPT sometimes writes answers that sound plausible but are wrong, and that its answers can change with small changes in how you phrase a question: one wording gets "I don't know," another gets an answer. The model is built to produce likely-sounding text, and a confident news summary is very likely-sounding text.

So when you ask about something recent, you may get a reply that blends older facts it does know with details that fit the pattern but never happened. It will not flag which parts are which.

A worked example: a question it cannot answer

Here is a question about something that happened only two weeks ago, so it cannot be in the model's training text.

Summarize how people reacted to OpenAI's ChatGPT release on November 30, 2022. Include two notable reactions.

There are two kinds of reply you might see. Both of these are illustrative, not transcripts.

The honest version is the useful one, even though it feels like a non-answer. The confident version is a trap. It reads well precisely because the model is good at writing the kind of text a news summary usually contains.

You can make the model's limits work for you by giving it the facts yourself.

Here are three paragraphs from news articles I trust about ChatGPT's release. [paste text]
Summarize the main reactions in five bullet points, using only the information in these paragraphs.

Now the model is doing what it does well, reorganizing and summarizing text, and the facts come from sources you chose.

What is a training cutoff?

A training cutoff is the point in time after which a model has seen little or no text. It reflects when the collection of training text was gathered, not the day the model was released, which is why a model launched in late 2022 can know little about 2021 onward. The edge is also blurry rather than sharp. The Verge suggests that ChatGPT's knowledge is described as limited after 2021 most likely because its training text becomes much sparser after that year, so it may know a little about some later events and nothing about others.

Two practical rules follow. First, for anything recent, such as news, prices, laws, product versions, sports results or who holds a job today, use a source that is updated: a search engine, a news site, an official page. Second, when an answer depends on timing, ask yourself whether the facts could have changed since 2021. If they could, verify them.

The cutoff is not a flaw that a bit more effort will fix inside this version of the model. It is a basic property of how language models are built today. Many expect future systems to combine a model like this with live search, but the ChatGPT available this month is a snapshot, and the best way to use it is to know where the snapshot ends.