Knowledge Cutoff Dates in LLMs

Large language models learn from a snapshot of text collected up to some point. They then go through further training and safety work before release. The knowledge cutoff marks the end of that snapshot. Release often comes months after the cutoff, and a model stays in use long after release, so the gap between what it knows and the present keeps widening. The cutoff is fuzzy rather than a hard line. People keep writing about events long after they happen, so the final months before a cutoff are thinly represented in the training data. This is one reason models are often unsure of their own cutoff, or underestimate it. Vendors usually publish the date in model documentation. Asking the model itself is not a reliable way to find it. Stale knowledge shows up in predictable ways: - outdated version numbers and API syntax - old prices - former officeholders or executives - superseded guidelines - no awareness of newer models, including the model's own successors A model may also assume the current date is close to its cutoff. That distorts its reasoning about ages, deadlines and what counts as recent. The main workarounds supply fresh information when the question is asked. Web search tools let the model read current pages. Retrieval-augmented generation pulls from a document store that someone keeps up to date. Simply stating the current date in the system prompt fixes many date errors. Fine-tuning can add knowledge, but it is slow and costly compared with retrieval as a way to keep facts current. A common misconception is that a chatbot with browsing has a later cutoff. Its training cutoff is unchanged. It is reading new sources, and its answers are only as current and accurate as the pages it retrieves.