Knowledge Cutoff

The last date covered by a model’s training data. After that date the model has no reliable built-in facts unless tools or retrieval supply them.

How it works

Training datasets are assembled by crawling the web, books, and other sources up to a specific date. The model learns facts from this data during pre-training. Any event, person, paper, or development that occurred after this date is simply absent from the model's knowledge. The model may still attempt to answer questions about post-cutoff topics by extrapolating from prior patterns, which is a primary cause of hallucination on recent events.

Why it matters

Knowledge cutoffs mean LLMs are fundamentally static knowledge stores in a dynamic world. This drives the need for RAG systems that retrieve up-to-date documents at inference time, web-browsing tools that let models search live information, and frequent model retraining or fine-tuning. For applications in fast-moving fields like AI itself, finance, or current events, the knowledge cutoff is a critical practical constraint that shapes system architecture.