Knowledge Cutoff
A knowledge cutoff is the point in time after which a language model has seen no training data, so events, publications, products and facts from later dates are simply absent from what it learned. A model is a frozen snapshot of text up to that moment; it does not update as it is used, and a conversation with it months or years later is a conversation with the world as it looked at the snapshot. This is a structural feature of how language models are built, not a bug in a particular system, and it is the reason that questions about recent news, current prices or the newest software version are where such models most often go wrong.
The cutoff is not the same as the release date, and it is not as sharp a line as the name suggests. Between the end of data collection and public release come cleaning, training, evaluation, safety work and deployment, which has often taken many months, and then the model stays in service for a year or more. The average conversation therefore happens well over a year after the cutoff. When ChatGPT launched in late November 2022, for example, its underlying models drew on data that ended in 2021. Over a model's working life, the gap between what it knows and what is true keeps widening, which researchers studying temporal generalization, such as Lazaridou and colleagues in their 2021 paper, found measurably degrades language model performance on text from after the training period.
Models often guess the current date wrongly because nothing in a plain language model carries a clock. Without a date in the prompt, it has only its training data to go on, and the safest-looking guess is a year close to the latest it saw plentifully, so it tends to assume that little time has passed. A related quirk is that models frequently underestimate their own cutoff. The amount of text discussing any period keeps growing for years afterward as articles, analyses and references accumulate, so the final months before the cutoff are thinly represented compared with how they will eventually be covered, and the model sees them only faintly. The effective cutoff, the point after which knowledge becomes unreliable, therefore sits earlier than the nominal one.
The standard remedy is not to retrain but to give the model fresh information at the moment of use. Retrieval-augmented generation, described by Lewis and colleagues in 2020, fetches relevant documents from a current index and places them in the prompt, so the answer is grounded in text the model never trained on. Search tools, API calls and a simple system-prompt line stating today's date handle the clock problem and a surprising share of the recency problem. Fine-tuning is a poor fit for this purpose, since injecting new facts through a small training set is unreliable, and continued pretraining on newer data is expensive and must be repeated. The cost of freshness shifts from training time to inference time.
Several misconceptions and pitfalls cluster around the topic. One is that a model with web access knows everything current; it knows only what its retrieval returns, which may be sparse, wrong or low quality, and it can still blend search results with stale memory. Another is that a model will say when it does not know; confident fabrication about post-cutoff events is common, because the model predicts plausible text and not a statement about its own limits. A third is the belief that the cutoff only matters for news. Software APIs, medical guidance, legal rules, prices and research literature all change, and code from deprecated library versions is a daily example. Evaluators also worry about the reverse problem, benchmark contamination, where test material predating the cutoff may have leaked into training data.
Good practice treats the cutoff as a property to be disclosed and designed around. Developers state it in documentation or model cards, put the current date into prompts, route time-sensitive questions through retrieval or tools, and test explicitly with questions about events just after the cutoff. Users can ask a model for its cutoff, but the answer is an estimate, not a measurement, and recent topics deserve verification against a live source.
Read a knowledge cutoff as the date on a photograph of the world: everything in the picture is sharp, everything after the shutter clicked is missing, and the model does not feel time passing, so it must be told what day it is and handed fresh evidence for anything that has changed since.