Knowledge cutoff

A cutoff date is a property of the pretraining corpus, not of the model's behaviour, and it is fuzzier than the single date on a model card suggests. Web-scale corpora are assembled from crawl snapshots such as Common Crawl, filtered and deduplicated, and then frozen months before training finishes; the model is then post-trained, evaluated and released, so the gap between cutoff and general availability is commonly several months to a year, and the model may stay in service for a year or more after that. Some providers now publish two dates: Anthropic's model documentation, for example, lists a "reliable knowledge cutoff" (the date through which knowledge is most extensive and reliable) separately from a later "training data cutoff" (the broader range of data used).

The distinction matters because coverage of any period keeps growing for years after it happens: news analysis, documentation, forum answers and encyclopaedia edits about an event accumulate long after the event itself. The last months before a cutoff are therefore thinly represented, and the model's knowledge of them is patchy. Cheng et al. ("Dated Data", 2024) probed models against time-stamped versions of the same resources and found that effective cutoffs often differ substantially from reported ones, and differ between sub-resources, attributing this to old content reappearing in newer crawl dumps and to deduplication schemes that interact badly with near-duplicates.

Models also do not reliably know their own cutoff. Asked directly, a model may name a date earlier than the real one, because the text it saw about itself or about the most recent period is sparse. It has no clock either: unless the system prompt or a tool supplies today's date, it will reason as if the present were somewhere near its training period, which produces wrong ages, wrong "latest version" answers and outdated regulatory status. Deployments therefore inject the current date and, where recency matters, give the model search or retrieval.

The cutoff interacts with neighbouring concepts in specific ways. Retrieval-augmented generation and tool use do not move the cutoff; they place newer text in the context window for the current request only. Fine-tuning on recent data can add some recent knowledge but is an unreliable way to update facts. Continued pretraining or a new model version is the only way to shift the cutoff itself. For security work the practical rule is that anything version-, vulnerability- or law-specific should be checked against a current primary source, because the model's parametric knowledge is guaranteed to be stale for anything that changed after its cutoff.