An AI knowledge cutoff is the date up to which a model's built-in knowledge is expected to reflect its training data. It is a rough boundary, not a guarantee that the model knows every earlier fact or that an AI product cannot fetch later information.

How a knowledge cutoff works

During training, a large language model adjusts its internal values using patterns found in its training data. Once that trained model is deployed, answering your question normally uses those values without updating them. New events do not enter the model simply because time passes or someone mentions them in a chat.

The reported cutoff summarizes how recent the model's learned information is meant to be. It does not mean that every source was collected on one day. Training data comes from many sources, collected and filtered at different times. Some information from before the cutoff may be absent, rare, contradictory, or too weakly represented for the model to reproduce reliably.

Researchers therefore distinguish a reported cutoff from an effective cutoff. The reported cutoff is the date supplied by the model provider. The effective cutoff is the period through which the model actually demonstrates reliable knowledge for a particular source or topic. Research on dated datasets found that those boundaries can differ.

This makes a cutoff more like a fading edge than a sealed wall. Knowledge can become patchier near the edge, and coverage differs by subject.

A worked example

Suppose a model's documentation gives March 31 as its knowledge cutoff.

  • You ask about a widely reported law passed on February 1. The model may answer correctly because the event happened before the cutoff and appeared often enough in its training data.
  • You ask about an obscure local notice published on February 1. The model may not know it, even though it predates the cutoff.
  • You ask about an election result announced on April 15. The model could not have learned that result from training data ending in March. Without another source, it may admit the gap, make an inference, or invent a plausible answer.
  • You paste the official April 15 result into the conversation and ask the same question. The model can now answer from the text you supplied. Its cutoff is still March 31.

Web search and retrieval systems work like the last case. A tool finds relevant material and adds it to the information the model can read for that request. This can produce a current answer without changing what the model learned during training.

Why the cutoff matters

A cutoff matters whenever an answer depends on time. News, laws, prices, officeholders, product specifications, medical guidance, and software documentation can all change after a model was trained.

Check the provider's documentation for the exact model you are using. Models from the same provider can have different cutoffs, and a product can switch its underlying model. Do not rely on the model to identify its own cutoff correctly.

For high-stakes or fast-changing questions, a cutoff date is only the first check. You also need to know whether the product can retrieve current sources, whether it actually did so for this answer, and whether those sources are authoritative. Search reduces the recency gap, but weak or misread sources can still produce a wrong answer.

A later cutoff can improve the odds that recent information was included. It does not prove that the model knows the fact you need. Test important, time-sensitive knowledge in your own subject area.

Common misconceptions

"The model knows everything before the cutoff"

No. The date describes a training boundary, not complete coverage. A pre-cutoff fact can be missing or unreliable.

"The model cannot discuss anything after the cutoff"

It can reason about newer information that you or a tool supplies. It may also infer likely developments, but an inference is not evidence that it learned the later event.

"Browsing updates the model"

Browsing gives the model temporary material to read for a request. It does not rewrite the trained model or move its cutoff.

"The cutoff is the same as the context window"

They limit different things. The knowledge cutoff concerns when built-in knowledge was learned. The context window limits how much information the model can process in the current request.

Where to go next

The practical distinction is between information learned during training and information supplied when the model answers. If you need the foundation for that distinction, start with What Is an LLM?. Then use Context Window vs. Knowledge Cutoff to separate the two limits directly.