What Your AI Doesn't Know Yet, and When to Ask It to Search

Every AI model stops learning on a fixed date. (A model is the AI system behind the chat, such as the one inside ChatGPT or Claude.)

Up to that day, the model reads an enormous collection of text, and then the reading stops and the model is released. That last day is called the knowledge cut-off, and the model can't know about anything after it.

A guidebook with no print date

Think of a travel guidebook. Someone gathers the facts, the printer runs the copies, and the book never changes again. Open the book two years later and it still says the museum shuts at five, in the same calm print.

An AI model is that guidebook, but with no print date on the cover. Its out-of-date answers sound exactly as sure as its current ones, so how confident a sentence sounds tells you nothing about how old the information is.

Don't bother asking a model for its own cut-off date, because the model may not know. Assume the book is older than it sounds, and check the book when the gap matters.

Which pages go stale

A guidebook's history chapter holds up for decades, while its opening hours are wrong within a year. AI splits the same way, and more goes stale than you'd think.

The stale pages include news, prices and opening hours, who holds a job and the latest version of an app. They also include tax thresholds, visa rules and what a company's policy says today.

Ask yourself one question to sort them. Could this have changed since last year? If so, it's a question for a web search.

For everything else, the history chapter does well. That covers things like how to structure a difficult email, what a word means, how compound interest works, or the plot of a novel from 1850.

Ringing ahead

A careful traveller rings ahead or checks the museum's own website. With AI, the check starts with three words, "search the web". Put those words in your question, because they make a search much more likely.

If your AI has a search or web option you can switch on, switch that on as well.

Search the web. What are the current rules on [the thing] in [your country or organisation]?
Give me the date of each source you use, and link it.
If you can't find anything from the last [three months], say so.

A search can still bring back an old page, which is why the prompt asks for dates. Before you trust an answer built on a source, look at when that source was published.

Fresh isn't the same as right, though. Your AI can still pick a weak page or misread a good one, or mix today's page with the old book in its head.

For work, money, health, travel or anything legal, treat the AI's summary as the place you start. Open the page the AI found and read it yourself, and then ask the question in How Do You Know? before you repeat a word of it.

When you can't tell whether a question counts as recent, make the AI tell you.

Before you answer, tell me if this is something that could have changed in the last year or two.
If yes, search the web first, and tell me what you found and when it was published.
If you're answering from memory, say that at the top.

If you have to open the source anyway, why use AI at all? For a single fact I'd often skip AI, and for "what time does it close" I'd go straight to the page.

AI earns its place when the answer is spread over several pages, or when you need it explained in plain words.

You know today's date, but the model doesn't. Which source you open, and what you pass on with your name attached, stay your call.

This week, take one rule or deadline you've already acted on and ask about it again, with "search the web" at the front.