Knowledge Cutoff

A knowledge cutoff is the temporal boundary of the training data used to build a machine learning model, particularly a large language model. Facts, events, or references from after this date are not represented in the model's training distribution and therefore cannot be reliably recalled or reasoned about without external retrieval. Knowledge cutoffs are an emergent property of pretraining pipelines — determined by the latest data included in the pretraining corpus — rather than an explicit parameter. Frontier model providers publish approximate cutoff dates, but the true boundary is fuzzy because pretraining corpora mix time-stamped documents and because models can interpolate across training data. Researchers can probe the effective cutoff by quizzing models on dated historical events and analyzing the accuracy curve.