Knowledge cutoff

A large language model learns from a snapshot of text collected up to a point in time. Training then takes months, so a model is usually released some time after its cutoff and stays in use long after that. Its built-in knowledge ages every day it is in service.

The cutoff is not the same as the model's release date, and it is not a hard wall. Material from the last months before the cutoff is often thin, because the web had not yet written much about recent events, so the model's grip on that period is weaker than on earlier years.

How it works

Builders gather training data up to a chosen date and then freeze it. Everything the model learns about the world, from prices to who runs a company, comes from that frozen set. After training, the model has no internal clock and no feed of news. Asked about something newer, it may say it does not know, or it may guess from older patterns and present the guess as fact.

Many assistants work around the cutoff by searching the web at question time and writing the answer from what they find, an approach related to retrieval-augmented generation. The model's language skills still come from training, but the facts come from the retrieved pages. When search is switched off, unavailable or not triggered, the answer falls back on the frozen training knowledge.

Providers usually publish a model's training cutoff in its documentation. The model itself is not always a reliable source for its own cutoff, because it learned about itself only from text written before it existed. A newer model does not fix this permanently either: from the day it is released, its own cutoff starts to age in the same way.

Example

A travel agency moved offices and changed its phone number last spring. A customer asks an AI assistant for the agency's contact details. With web search on, the assistant reads the current website and gives the new number. With search off, the model answers from training data gathered before the move and confidently gives the old address.

The owner tests this by asking two different assistants the same question, once in a mode that searches and once without, then noting which details are out of date. The fix is on the agency's side: update the website, its Google Business Profile and directory listings so that any search finds the new details consistently. Within a month or two, search-enabled answers usually reflect the change; answers from training alone will not until a newer model is trained.

Why it matters for your business

Customers increasingly ask AI tools about businesses, and a model past its cutoff can repeat old prices, closed locations or services you no longer offer. You cannot change a model's training, but you can make current facts easy to find and consistent everywhere, which is what search-enabled assistants read. When you use AI for your own work, check anything time-sensitive, such as laws, rates or software instructions, against a current source.

Common mistakes

• Trusting an AI answer about recent laws, prices or events without checking whether it searched the web.
• Assuming a newly released model knows about things that happened in the months just before its release.
• Updating your website but leaving old details on directories and profiles that AI tools also read.
• Asking a model what its own cutoff is and treating the answer as definitive.

Questions people ask

How do I find an AI model's knowledge cutoff?

Check the provider's documentation or model card, which usually states the training data cutoff. Asking the model directly is less reliable, because it may be uncertain about its own training. Test it with a question about a recent event you know the date of, with web search turned off.

Does web search remove the knowledge cutoff?

It works around it for facts the search finds, but it does not update the model itself. The assistant reads current pages and summarises them, so answers can be up to date. If the search misses the right page or is not triggered, the answer may still rely on older training knowledge.

Why does ChatGPT not know about something recent?

Usually because the event happened after the underlying model's knowledge cutoff and the tool did not search the web for that answer. Asking it to search, or pasting in a current source, usually solves it. Some very recent events also have little written about them online yet.