Definition
A knowledge cutoff is the date up to which an AI model's training data extends. The model has no built-in knowledge of events, products or changes after that point. Cutoffs vary by model and version, and vendors do not always publish exact dates. Even before the cutoff, the model may not know every fact.
For example, if Sunnara launched a new sunscreen after a model's cutoff, that model cannot know about it from training alone. If the assistant searches the web while answering, it may find the launch; if it does not search, it may describe Sunnara's older range as current.
Why it matters
The knowledge cutoff explains why assistants sometimes give outdated answers about your brand, such as old pricing, discontinued products or a previous name. Anything after the cutoff must come from live retrieval, which depends on your pages being crawlable and current. Knowing this helps you separate two problems: answers based on outdated training data, and answers based on outdated pages that are still being retrieved today.
What to do
- Ask assistants about recent changes, like new products or pricing, with and without web search.
- Note where answers reflect old information and whether any sources are cited.
- Make sure current facts are on crawlable pages that retrieval systems can find.
Common questions
How do I find a model's knowledge cutoff?
Check the vendor's model documentation or release notes, which often state a training data cutoff. Asking the model directly is less reliable, since it may not report its own cutoff accurately. Cutoffs differ between models and versions from the same company, so check each one you track.
Can an AI assistant know about things after its cutoff?
Yes, if it retrieves information while answering, for example through web search or connected documents. Without retrieval, it can only use what was in its training data. Whether retrieval runs depends on the assistant, its settings and the question being asked.