Ten things you
must know.

Every answer is a prediction of what text comes next — there is no database of facts behind it, and no internal flag that separates remembering from inventing. Everything else on this page follows from that.

The tone never wavers, right or wrong. Fluent, precise-sounding prose is what these systems produce — it tells you nothing about whether the content is true.

A model has no memory of last week’s chat and no idea what’s in your files. It sees a fixed window of text — what you put in front of it — and nothing else.

The single biggest reliability upgrade: stop asking what it remembers and show it the source. Answers from a supplied document can be checked; answers from memory are guesses with good manners.

Vague instructions get vague work. State the role, the task, the format, and what to do when it doesn’t know — the skill is precision, not magic words.

Pasting client material into a consumer tool is a disclosure to a third party on that tool’s terms. Enterprise route, UK/EU endpoint, zero data retention — or it doesn’t get client material.

Fabricated authorities have already ended up before the courts, and the responsibility landed on the professionals who filed them, not the software. If you can’t find it on an official source, it doesn’t go in the document.

There is no “the AI” — there are tiers and settings. Cheap and fast is right for volume triage; the frontier model on high effort is right for the advice that carries your name.

The newest systems don’t just answer — they do things: search, draft, file, send. Capability and risk rise together; delegation is fine, abdication is not.

Confidentiality, competence, accuracy — no regulator has granted an AI exemption. The work still leaves under your name; the tools just change how quickly you can produce it — in both directions.