1. Home
  2. Glossary
  3. Knowledge cutoff
AI glossary · Models & training

Knowledge cutoff

Knowledge cutoff: A knowledge cutoff is the date after which a language model's training data stops. Unless it is given live information through search or documents, the model knows nothing about events, prices, or rules that changed after that date.

Models are trained on a snapshot of text collected up to a certain point, then released months later. Everything after the snapshot is a blank. Ask about a tax threshold, a software feature, a regulation, or a company that changed after the cutoff, and the model will answer from stale information, often without warning you.

Two things make this worse. Models are frequently unsure of their own cutoff date and may state it wrong. And they rarely say "I do not know about that yet." Instead, they generate a fluent answer that was true as of their training, or that sounds like it could be. Stale facts are one of the most common ways AI output is confidently wrong.

The fix is to bring the current information to the model. Most assistants now offer web search, which fetches live pages and cites them. You can also paste the current document, policy, or price list into the prompt and tell the model to answer only from it. When a question depends on anything that changes yearly, such as limits, rates, deadlines, or product features, treat the model's unaided answer as a guess.

Example at work

A financial analyst asks for the current contribution limit on a retirement account. The assistant gives a figure from a prior year. She re-asks with web search turned on and the request "cite the IRS page and its date," gets the current limit with a source, and confirms it on the IRS site before it goes into a client memo.

Why it matters

Stale knowledge does not look stale. If you know every model has a cutoff, you will build the habit of supplying current sources or turning on search for anything time-sensitive, which prevents a whole category of embarrassing errors.

Related terms