Where a model’s knowledge stops
The knowledge cutoff is the point up to which a model’s training data runs. A model with an early-2025 cutoff does not know about events, products, or rule changes after that. Alongside hallucination, it is one of the first properties of LLMs worth understanding.
Why there is a cutoff at all
An LLM is pre-trained on data collected up to a certain date. Training takes months and costs an enormous amount, so continuous learning of the latest information is not practical. The date the data was frozen becomes the cutoff, and it moves forward when a new model generation ships.
Where it causes trouble
- Answering about recent products with outdated or invented information
- Treating laws, policies, and prices as they stood at training time
- Responding to “today” or “recently” from a pre-cutoff frame of reference
Worse, a model does not notice that its knowledge is stale and will answer confidently. Combined with hallucination, that is a difficult failure to catch.
Search and RAG as the fix
Most current assistants can search the web and look things up before answering. In corporate systems, RAG against current internal documents is the standard approach. For users, two habits cover most of the risk: use a search-enabled mode for anything time-sensitive, and check the source for important current information.
The entries on this site carry a “last reviewed” date for the same reason. Information has a shelf life, whether or not an AI is involved.
Cutoff dates are usually published in a model’s official documentation. Checking once for the model you use gives you a feel for how fresh its answers can be.