Search vs generation

An AI answer comes from one of two places: the model’s memory — patterns absorbed during training, frozen at its training cutoff — or live sources fetched right now by a search-connected tool. The words look identical either way, so before trusting any fact, ask which one you just got.

What each mode is good at

• Memory (generation) — explanations of stable concepts, transformations of text you supply, structure and drafting, reasoning over your own documents. If the answer would read the same a year ago and a year from now, memory is fine.
• Live retrieval — prices, laws and rates, product versions, schedules, news, availability. The tell: if the correct answer is different today than last year, memory alone cannot be reliably right.

The traps on each side

Memory fails through staleness and confident invention — the mechanism behind the hallucinations lesson. Retrieval fails differently: a page can be real and still not support the claim, and search results can be thin or misleading. Live search makes answers checkable — it does not make them checked. That is the next lesson’s job.

A bad example

Ask (no web access): “What does the new data-protection amendment require for small businesses?”

The amendment is recent — possibly newer than the model’s training data. Memory answers with the average of what it saw: plausible, structured, and quietly a year out of date.

A better example

Ask (web-connected): “Find the current text of the amendment and summarize what it requires for small businesses, linking each requirement to the section of the official text it comes from.”

Now the answer arrives with pointers you can open — and “the retrieved text does not cover this” becomes a possible, correct answer.

The five-second test

Before trusting a fact: would this fact survive a year unchanged? If yes, generation is usually enough. If no — or if being wrong costs money, time or compliance — ask a search-connected tool and require sources. The test takes five seconds and prevents most stale-fact mistakes.

A closing habit that makes the distinction physical: when a fact matters, copy the source line into your own notes as you verify it. Notes that carry a pointer per fact are what you actually trust later — and the pointers make rechecking cheap when the world moves. Memory starts the work; sourced notes finish it.

Practice

Transfer

• Travel and logistics: schedules and entry rules change — search-connected only.
• Software: ask for the version you use and confirm against official docs when commands read suspiciously new.
• Concepts and writing: memory is fine; save the searches for facts with expiry dates.

Next

Next: Inspect the citations — the source exists. Now: does it actually say what the answer claims?