Knowledge Cutoff

A knowledge cutoff is the date after which a model’s training data stops. Anything that happened later is not in the weights, so a model asked about it is working from whatever it can retrieve at the time of the question, or from nothing at all.

Why it is the wrong thing to blame

The intuitive conclusion is that a business founded after the cutoff cannot be recommended until the next model is trained, and that is mostly wrong. Assistants answer commercial questions by searching at the moment they are asked, and a source retrieved during the conversation does not need to have been in the training data at all. That is the path a new business is actually on.

What the cutoff genuinely determines is what the model knows without looking: the background against which retrieved material is read, and what it will say if retrieval fails or is switched off. A business that exists only in retrieval is named when the search works and absent when it does not, which is a less stable position than being in both, but it is a real one.

What this means in practice

Two things are worth acting on. First, being retrievable is the lever that responds on a timescale anyone can work with: the training corpus is not something a business influences this quarter, and whether a crawler can reach the site is decided today. Second, a claim that contradicts what the model learned before its cutoff has a harder job than one that fits: an outdated fact about a business that the model absorbed years ago can persist, and correcting it means giving retrieval something clear and current to find.

A model working from stale weights and no usable retrieval is also the condition in which a confident wrong answer, a hallucination, is most likely.

One question people ask

How do I find out a given assistant’s cutoff?

Asking it is unreliable: models are frequently wrong about their own cutoff, and the answer can differ between runs. The provider’s model documentation is the place to check, and it is worth rechecking, because the deployed model behind a product name changes without the name changing. For most practical purposes the more useful question is not what the cutoff is but whether the assistant can retrieve your pages right now.

Sources

This entry describes how training cutoffs and retrieval relate. It reports no measured figures and cites no third-party data.