From memory, by searching, by investigating: three modes

When you ask an assistant a question, it can answer in three very different ways. The answer looks the same in all three cases; its reliability does not.

1. From memory

Without searching, the model answers with what it learned during training, which stopped at a cut-off date often several months old. Excellent for explaining a concept, a grammar rule, a principle of physics. Dangerous for anything that moves: a recent law, a price, a timetable, the news. For want of anything better, it fills the gaps with what was true at the time, or with what seems plausible.

2. By searching

With web search enabled, the assistant runs a few queries, reads a handful of pages, then writes a summary with links. It is fast, up to date, verifiable. The quality depends entirely on the pages found: if the top results are mediocre, the summary will be too, with nice formatting.

3. By investigating

Deep research modes work for several minutes: they draw up a plan, run dozens of searches, read many pages and produce a structured report. Useful for breaking new ground on a whole subject. We come back to this in a dedicated lesson.

Knowing which mode you are in

Look for clickable links in the answer, and a search indicator while it is being prepared. No link: it is from memory. For a question about current events or figures, explicitly enable search, or write "search the web and give me the links".

Key points

• From memory: solid on what does not change, fragile on current events and figures.
• By searching: up to date and verifiable, but only as good as the pages found.
• No link in the answer: it comes from the model's memory.