knowledge cutoff

basicwidespread

It is the date until which a language model's training data extends; the model does not know anything about what happened after that.

In plain terms

A language model learns from vast amounts of text, but learning stops at a certain point. This point is the knowledge cutoff: in the model's mind, the world stopped on that date.

It is like a printed encyclopedia. What happened after it went to press is not included, no matter how thorough the editor was. The model only knows about new things if someone places that new information in front of it.

This is not a bug, but an inherent feature of operation: training a model takes weeks or months and is very expensive, so it cannot be redone daily. Typically, a few months, sometimes even a year, pass between the knowledge cutoff and the release, and after that, the model is used for a long time.

Therefore, it is important to know where the limit lies: regarding yesterday's news, current prices, or laws amended this year, the model can only provide reliable information if it can search the internet in the meantime, or if you provide it with the fresh text yourself.

How it works

Developers gather training material: websites, books, articles, and source code. The collection is closed on a specific date.
The model trains on this material for weeks or months, then is tested, and only then is it presented to users. By then, its knowledge is no longer fresh.
The trained model's knowledge is fixed: it does not learn further from conversations with you and does not learn about new events on its own.
The gap is filled by external tools: internet search, uploaded documents, or a corporate database. In such cases, the model reads the fresh text it receives and answers from it.
The model does not know the current date on its own. The service usually inserts it into the conversation in the background; otherwise, the model might easily assume it is still in the year of its knowledge cutoff.

An everyday example

You ask the chatbot who won last week's match, and it replies that it has no information on this, or confidently states an older result. If you enable web search, you receive a fresh, sourced answer to the same question.

Who is affected

everyoneAnyone asking a chatbot for fresh news, prices, or current events must know the answer may be outdated.

lawyer, citizenLaws change often, and the model can easily quote text from before its knowledge cutoff that is no longer in force. → Law

developerThe model does not know program library versions released after its knowledge cutoff, so it may suggest outdated solutions. → Code

parent, student, teacherWhen doing homework or research, fresh data and the latest results must always be checked separately. → Education & Family

A common misconception

MythArtificial intelligence is always up to date, since it is connected to the internet.

In realityThe model's own knowledge stops at its knowledge cutoff. It can only access fresh information if the service provides a search tool or documents alongside it, and this is not always enabled.

Limits and risks

The model does not always indicate that it cannot know something: it may confidently state outdated data or fabricate a plausible-sounding answer.
The knowledge cutoff is not a sharp line: less text from the most recent months made it into the training material, so the model's knowledge is already patchy there.
Models often misjudge their own knowledge boundaries, so do not ask the chatbot; instead, check the developer documentation.
Web search helps a lot, but the model's background knowledge and perspective remain outdated, and the answer depends on the quality of the found sources.

Similar, but not the same

context window – The context window shows how much text the model can process at once in a conversation, while the knowledge cutoff indicates how far back its trained knowledge extends.
hallucination – A hallucination is a fabricated, untrue statement. An error caused by the knowledge cutoff, by contrast, is mostly information that was true in the past but has since become outdated. The two can be related: if the model is asked about an event after its knowledge cutoff, it may make up an answer.
release date – The release date is the day the model is made available, which is typically months later than the knowledge cutoff.

Related concepts

Broader

large language model, training data, training

Narrower

–

On the edge

hallucination, search-based answer generation, web search, context window, fine-tuning

A brief history

The concept became widely known with the launch of ChatGPT in November 2022: the chatbot then knew the world only up to September 2021, and it regularly referred to this in its answers.

Where things stand

As of late 2026, major developers typically state the knowledge cutoff in their model documentation. Most chatbots already fill the gap with web search, but the limitation itself has not disappeared.

Status: 2026 November

Also known as

knowledge cutoff · knowledge cutoff date · knowledge siloing · knowledge cutoff date · knowledge cutoff · knowledge cut-off · cutoff date · training data cutoff

Sources

Cheng et al. (2024): Dated Data: Tracing Knowledge Cutoffs in Large Language Models (arXiv)

This entry was machine-edited and has undergone machine expert review; the “In the news” section updates automatically from Agent Age stories.