Use AI like a talented intern, not a source of truth.

01 / The useful mental model

Fluency is a finish, not proof.

AI can sound certain while being wrong. Treat the polished sentence as a draft—not as evidence.

A language model builds a response from patterns in its training and the context you give it. It is very good at making a plausible next move. That is not the same as checking a live source, opening a database, or knowing that a claim is true.

So change the question from “Is this answer confident?” to “What would I need to see before I rely on it?” The answer can still be useful: a thinking partner, first drafter, patient tutor, or tireless way to explore options. You stay the editor and the person accountable for the result.

02 / Choose your checking level

Match the check to the stakes.

Not every idea needs a research project. But the cost of a wrong answer should set the depth of your check.

For low-stakes work—brainstorming a dinner, exploring names, getting unstuck—use the answer as raw material and keep moving. For medium-stakes work—business claims, product comparisons, a public post—ask for the assumptions and verify the important points against primary sources.

For health, legal, financial, safety, or reputational decisions, slow down. Use qualified people and authoritative records; never let a fluent summary become the final decision-maker. If the model cannot point to a source you can inspect, the claim is still unverified.

• Low stakes → explore freely.
• Medium stakes → check claims against the source.
• High stakes → bring in an expert and primary evidence.

03 / The working loop

Make the conversation a loop.

The best result usually arrives after the first answer, when you give the model something real to react to.

Start with the outcome. Generate a first pass, mark what feels generic or unsupported, then add the missing context and tighten the constraints. Ask the model to expose its assumptions, separate known facts from guesses, and tell you what would change its recommendation.

A small correction beats a grand, mysterious prompt. “Keep the layout, but make the lighting quieter” gives the next round a usable direction. Your taste is the steering wheel; the tool supplies more roads to choose from.

• Generate → select → diagnose → add context → verify → finalize.