MMOOZZGG.com

Glossary

What is a knowledge cutoff?

In short

A knowledge cutoff is the point in time up to which the texts for training a language model were collected. The model does not know about later events by itself.

Language models · knowledge cutoff

MMOOZZGG.com editorsChecked: 1 min read

For a child

Imagine an encyclopedia printed two years ago. It has lots of interesting things, but not a word about a new cartoon or yesterday’s football match: the book ended on the day it was printed. A neural network has such a day too. It has not read anything that happened later. If you ask it about fresh news, it may get it wrong or start making things up.

In plain words

A language model is trained on a huge set of texts, and the set is collected up to a certain date. Then training takes months, and the finished model works for another year or two. All this time its knowledge stands still: it does not know new laws, prices or exchange rates unless you show them. So a model sometimes confidently names an outdated tax rate or last year’s version of a program. Two things help: connect search to it, or pass the needed document right in the request.

For business

For a company, this boundary means a simple rule: anything that changed recently, the model may know wrongly without a hint. This covers rates, laws, supplier prices and your product range. The error sounds convincing, so it is dangerous in answers to customers and in legal notes. The fix is cheap compared with the risk. Fresh data goes to the model through search in a knowledge base (RAG) or calls to your systems, and the instructions ask it to rely only on that data. Before launch, find out what date the provider states and make a set of questions about events after it. Such checks are part of reviewing AI errors before adoption.

For a specialist

A knowledge cutoff is the time boundary of the pretraining data. Providers often give two dates. In its model documentation, Anthropic separates the reliable knowledge cutoff, up to which information is most complete, from the wider training data cutoff. Cheng and co-authors (2024) introduced the idea of an effective cutoff and showed by probing that it often differs from the stated one and varies by topic. There are two reasons: new CommonCrawl dumps contain many old pages, and deduplication schemes remove semantic and near-exact duplicates poorly. The model also does not know the current date unless it is passed in the input. In practice, teams close the gap with retrieval-augmented generation, tool use and the date in the system prompt. Fine-tuning for fresh facts is expensive and goes out of date quickly itself.

Example

An accountant asks an AI assistant about rules that changed this year and gets a confident answer based on the old version. After search in a base of regulatory documents is connected, the assistant answers from the current text and shows where the answer came from.

Related terms

Questions and answers

How do you find out a model’s knowledge cutoff? The provider publishes it in the documentation or the model card. Asking the model itself is unreliable: it may name the wrong date.

For some topics, the real boundary is earlier than the stated one.

Can you move the cutoff without new training? Only new training moves the boundary itself. But you can give the model fresh information in the request: through search in documents, on the internet or in your systems.

For a business, this is cheaper and easier to maintain.

Sources

Not sure whether your company needs this?

AI solutions map
Contact us