Knowledge Cutoff — What It Means in Plain Terms
Why this matters
- Understanding why a model confidently quotes last year's price and an outdated version of a policy
- Knowing which processes must never be handed to a model without supplying current documents
- Asking a vendor where the system gets current data, not just which model sits inside it
What is a knowledge cutoff, in plain terms
A reference book was printed in March. Nothing that happened after March made it in — not because the editors were lazy, but because presses do not print the future. The book has no idea, and answers every question with March data.
A model works the same way. Training runs on texts collected up to a certain date. That date is the cutoff: the boundary between what the model read and what never existed in its world.
There is one difference from the reference book, and it is the important one. The book is honest about itself: the year of publication is right there on the cover. A model shows no publication year and answers a question about yesterday's event just as smoothly as one about the year before last.
How it works
- Data is collected up to a specific day. Training takes in the texts available at the time of collection. Collection then closes, and the contents of the foundation model are fixed.
- The model is trained and tested. This takes months. The gap between the cutoff and the model's release is created here.
- The model ships. From day one it answers according to the state of the world at the cutoff date, not at the date of the question.
- The gap widens. Every month in service adds a month to the lag. A model that has run for eighteen months answers with data that is two years old.
- Anything current is supplied in the request. Everything that happened after the cutoff reaches the model only through the text of the request: an attached document, a retrieved excerpt, a field from a database.
A separate word on confidence. An answer built on stale data sounds exactly like an answer built on current data: same tone, same level of detail. There is no label reading "this comes from two-year-old memory." So the boundary is judged not by tone but by where the data in a given answer came from.
A case from practice
A wholesale company planned to hand a model the job of answering questions from its sales staff: shipping terms, current discounts, delivery times by category. The prototype was built on a single model with no connection to a database — just a question and an answer.
On a test set of fifty questions, the picture split in two. General questions — how a payment deferral is calculated, what a framework agreement is — the model handled correctly. Questions about the company's own specific terms it also handled — confidently and wrongly, because those terms were not in the training data and could not have been.
The prototype was then rebuilt: each question triggered a search that pulled in the current price list and the current version of the policy, leaving the model to work from the supplied text. The same set of fifty questions was run again, with the commercial director checking the answers. The measurement method is simple and repeatable: same question set, same reviewer, measured before and after.
What this gives a business
- Tasks sort into two piles ahead of time. Stable knowledge — language, formats, general parsing rules — the model carries inside it. Anything that changes more often than once a year is supplied as a document. The sorting takes an hour on paper and saves weeks of piloting.
- Vendor due diligence gets concrete. Instead of "which model do you use," the question becomes "where does the system get data that was not in training." Either there is an answer, or the system will confidently serve up stale information.
- Switching models stops being a lottery. When current data arrives from a knowledge base rather than the model's memory, swapping the engine does not rewrite what the system knows. Comparing the options is easier with the solution finder.
- A whole class of costly errors disappears. An answer based on last year's tariff or a repealed version of a policy costs money and reputation. The price of such errors for a specific process can be worked out with a calculator.
When this can be ignored
Some tasks are indifferent to the cutoff date. Translation, proofreading, rewording a letter, parsing an attached document into fields, summarizing a meeting — here the model works with the text it was handed, and its own knowledge of the world barely comes into play.
The same goes for tasks built on stable rules: classifying inquiries by type, extracting registration details, sorting incoming mail by topic. The rules rarely change, and a year-long gap makes no difference.
The distinction is simple: if the answer follows entirely from the supplied text, the cutoff does not matter. If the model has to fill in anything from memory, it always does.
What to check
First: the cutoff date comes from the provider's documentation, not from the model's own answer. A model recalls its date from memory and is off by months — that recollection is an answer from training data like any other.
Second: staleness is uneven across fields. The rules of business correspondence have not changed in a year, while tariffs, exchange rates, versions of regulations, and the range of models on the market have changed completely. Testing has to use questions from the relevant field, not general ones.
Third: supplying documents in the request cures the cutoff only where a document was actually found. If the search comes back empty, the model will default to answering from memory. The system therefore needs an explicit rule to answer "no data" — and that rule deserves its own separate test, using questions the base demonstrably has no answer for.
Frequently asked questions
What is a knowledge cutoff, in plain terms?
It is the date of the most recent data used to train the model. Anything that happened in the world after that does not exist as far as the model is concerned: it never read about it. Asked about later events, the model answers as of its own date — and usually gives no warning that the date has passed. Current information has to be supplied directly in the request.
How does the cutoff differ from a model's release date?
The release date is when the model became available. The cutoff is when the collection of training data stopped. Several months typically separate the two: the data has to be gathered, the model trained on it, then tested and shipped. Even the newest model is therefore already behind the calendar on day one, and the gap widens with every day it stays in service.
Why does the model not say when it does not know something?
The model answers from its own data and never checks the calendar. If an older version of a regulation was in the training data, it will confidently recite that older version — and the wording will look every bit as convincing as a correct answer. This is a specific case of [hallucination](/en/blog/gallyutsinatsiya). The fix is not arguing with the model but feeding the current document into the request.
How can the cutoff of a particular model be found?
The date is published by the provider in the model's documentation. Asking the model itself is unreliable: it recalls the date from memory and is often off by months. A practical check is to ask a dozen questions about events of known vintage in the relevant field and see where the answers start going stale. The check is repeated every time the model changes.
What should be done about processes where currency matters?
Current documents are pulled into the request by searching an in-house knowledge base — the approach is called [RAG](/en/blog/rag). The model then answers from the supplied text rather than from memory. A second layer is a rule to answer "no data" when the required document is not found. A third is an update date shown next to every document in the base.
Let’s discuss your project?
Tell us about your process — we’ll suggest where AI pays off fastest.
Related articles
AI hallucinations: why models confabulate and how to catch it
A hallucination is a confident answer that looks plausible but does not match reality: an invented link, a made-up number, or a clause that appears in no regulation.
Source citation: how to verify a model answer in a minute
Source citation means a link to a specific fragment of a document next to every claim a model makes: the answer can be verified in a minute and is fit to act on.