Knowledge Cutoff
A model's knowledge cutoff is the date its training data ends. It won't know about later events unless that information is supplied in its context, for example through search or retrieval, or the model is trained again.
Also known as: training cutoff, training data cutoff, knowledge cut-off
A large language model learns from a snapshot of text collected up to some date. After that, its knowledge is frozen: ask about a newer event, product or price and it will either say it doesn’t know or, worse, produce a plausible guess. Models are often released months after their cutoff, so even a brand-new model is behind.
The fix is to put current information into the model’s context window at the time of the question. Chat products do this with web search; enterprise systems do it with retrieval-augmented generation, which fetches the relevant documents first. The same approach covers private data, which a general-purpose model usually never saw in training.