Ask about last month, with and without search.

To see the edge of the snapshot: what the model does when asked about a time it has never read, and how the answer changes when search puts new text in front of it.

A chat tool where web search can be switched off and on (ChatGPT and Claude have a toggle; in Gemini use a model without grounding, or just note that it may search). One fact you know for certain that has changed in the last year or two.

1. New chat, search off. Paste:
   What is your training cutoff date? Answer in one line.
   Record it.
2. Same chat. Paste:
   Without searching the web, tell me one significant technology news item from last month, with the date. If you cannot know, say so.
   Record what it says, and whether it admitted not knowing or produced something.
3. New chat, search on. Same question as step 2. Record the answer and, if the tool shows them, the sources it used. Note whether a search visibly ran.
4. New chat, search off. Ask your changed fact, for example:
   Who is the current CEO of Akamai? One line.
   Record the answer and whether it is correct.
5. Repeat step 4 with search on.
6. Optional, any chat: ask "Which AI model is the newest one available?" and compare with what you know.

Tool and model used:

The stated cutoff is usually a year or more behind today. In step 2, a well-trained model says it cannot know; an older or smaller one may produce a confident headline that never happened, which is the failure the note is about. In step 3, watch whether a search actually ran: "search on" only means the product may search. In steps 4 and 5, look for the hedge "as of my last update" and whether it disappears once search has run. And notice if the model mentions today's date at any point: it cannot know that from training, so it was told.

one line: did it guess, or did it say it could not know?