Why AI does not know about recent events

What’s actually happening

A model is trained once. Training consumes a body of text that ends on a particular date, the run finishes, and what comes out is a fixed set of numbers. Nothing after that date is inside it. Using the model does not update it; the next version does.

That much is well known. What is less well known is how soft the published date is.

Vendors publish two different dates and mean two different things. Anthropic’s comparison table carries both a “Reliable knowledge cutoff” and a “Training data cutoff”, and defines them in a footnote: “Reliable knowledge cutoff indicates the date through which a model’s knowledge is most extensive and reliable. Training data cutoff is the broader date range of training data used.” For Claude Haiku 4.5 the two are five months apart — February 2025 and July 2025. So there is a stretch of time the model has seen something about, and about which it is explicitly not underwritten.

Some cutoffs are a range, not a line. Google’s model card for Gemini 3.6 Flash puts it in one sentence:

The knowledge cutoff date for Gemini 3.6 Flash is March 2026 – users can expect updated information for some domains while in others they may experience the model's knowledge is limited to January 2025.

Fourteen months of difference depending on the topic.

And some models state no date of their own at all. Two of the Gemini entries in our dataset inherit theirs through a chain of model cards that defer to an older card, which is recorded in the dataset rather than smoothed over.

The demo above puts all of it on one axis: fourteen models, each date linking to the vendor page it came from. Pick a date and the bars that end before it are the models for which that event does not exist.

The last piece is the one that causes the actual damage. A model has no way to notice that a question is about a period it was not trained on — there is no clock inside it, and often no date in the request either. So the answer arrives in the same tone as every other answer. The gap in knowledge is invisible from where you are sitting.

What to do about it

Paste the material instead of asking about it. The reliable route for anything recent is to put the text in the request. Then it is in front of the model regardless of when training stopped.

Turn on search when the question is time-sensitive. If your app can retrieve, use it, and check what it retrieved. Retrieval changes the input; it does not change what the model knows.

Say the date out loud. “As of July 2026” in the question, and dates on any event you mention, give the model something to work from — it does not otherwise know what day it is.

Do not ask the model for its own cutoff and stop there. That claim comes from training or from a system prompt, and both go stale. The vendor’s model page is the source of truth, and it is one click away.

Be most careful where facts expire quietly. Prices, versions, laws, officeholders, availability and interest rates all keep their old values with no visible mark of age. Those are the answers to check first.