Knowledge cutoffs: why your AI does not know last week

A knowledge cutoff is the date after which a model saw no training data at all. Everything it knows unaided was collected before that date and frozen when training ran, and nothing about using the model updates it afterward. So when you ask about something from last week and the assistant is not connected to a live search tool, you get one of three things: an honest admission that it does not know, an answer drawn from before the cutoff and served as though it were current, or a fluent guess assembled from patterns. The third is the one that costs you, because on the screen it looks every bit as certain as the other two.

The cutoff is also not a clean wall. It is more like a fog bank that thickens as you approach the edge, which is why a model can be sharp about something from four years ago and muddled about something from four months before its own cutoff.

Where the cutoff comes from

Training a large model starts with collecting an enormous snapshot of text. Once collection stops, that snapshot is fixed. The expensive part of training runs afterward, and after that comes the tuning and testing that turns a raw model into an assistant. All of that takes time, which is why a model released in one year commonly carries a cutoff from the previous one. The mechanics of that pipeline are covered in how models are trained.

The edges are fuzzy for a second reason. Later tuning stages use smaller batches of newer material, so a model sometimes knows a scattering of facts past its stated cutoff without knowing the surrounding context. That produces the odd effect of a model naming something that happened after its cutoff while getting every detail about it wrong.

The last few months are the haziest part

The web does not write about an event all at once. A news story appears, then analyses, then corrections, then reference pages, then arguments about it, spread over months and years. An event from five years before the snapshot is represented by thousands of documents that broadly agree. An event from six weeks before the snapshot is represented by a handful of first drafts, some of which were wrong.

The practical consequence: treat roughly the final six to twelve months before any stated cutoff as partially known rather than known. The model may know that something happened, but get the date, the outcome, the numbers or the people wrong. This is the same mechanism behind invented facts generally, applied to the thin end of the training data.

Why it cannot tell you its own cutoff

The cutoff date is a fact about how the model was built, not something written in the text it read. Unless the product deliberately tells the model its cutoff in a hidden instruction you never see, the model has to infer the date from the material it remembers, and it infers badly. Common failure patterns include naming a date several months off, naming the cutoff of an earlier model in the same family, or confidently stating a date that was simply a frequent phrase in its training text.

Models also do not know today's date unless the app inserts it into every request, which many do and some do not. If the model is wrong about both today's date and its own cutoff, it cannot reason about what it is missing.

A better test takes thirty seconds. Ask about two or three things you can date precisely yourself, spaced a year apart, and see where the accuracy falls off. That tells you what you actually need to know, which is not the label on the model but the edge of its useful knowledge.

What a web connection fixes, and what it does not

The blending problem deserves attention. Retrieval typically pulls a small number of pages into the request, and the model then writes one coherent answer. It does not mark which sentence came from where. A paragraph can contain a fee published yesterday followed by an eligibility rule remembered from three years ago, in the same calm voice, with a single citation link at the end that only supports the first half. The broader differences between this and an ordinary search are covered in AI search versus a search engine.

Telling memory from retrieval in under a minute

1. Look for the product's own signals first: a searching indicator, footnote markers, or a sources panel. No indicator usually means no search.
2. Ask directly, in the same chat: "Did you search the web for this, and what date does this information reflect?" A model that did not search will often say so when asked plainly.
3. Open the first citation and search the page for your key figure. If it is not there, the figure came from training.
4. Check the page's own published or updated date, not the date the chatbot implies.
5. If there is no usable source, re-ask with an explicit instruction: "Search the web, quote the exact line, and give me the publication date."
6. If it still cannot produce a source, treat the answer as pre-cutoff information and verify it the ordinary way, using the routine in fact checking an AI answer.

Which questions are actually time sensitive

Some things move constantly and some things do not, and sorting your question into the right pile saves most of the effort.

Demand a source for prices, fees, interest and tax rates, benefit thresholds, laws and filing deadlines, opening hours, transit schedules, software menu names and settings paths, medical guidance, sports and election results, who currently holds a job or office, and whether a company or service still exists.

You can relax for how a process works in principle, grammar and translation, historical events long settled, mathematics, how to structure a letter or a plan, and explanations of concepts. These were stable before the cutoff and they are stable now.

What to check first

Before you rely on anything with a date attached, ask two questions: does this fact change, and did the answer come with a source I opened? If the fact changes and you did not open a source, you have an unverified claim rather than an answer, no matter how well written it is. Re-ask with an explicit request to search and quote, then read the page yourself. Thirty seconds of checking is the whole defense, and it works regardless of which assistant you use or what its cutoff turns out to be.