Understand · Judgment & risk

Judgment & risk

When to trust an AI answer

A quick way to decide how much checking an answer needs.

5 min readPlain English

Read it in

In one sentence

Match your checking to the stakes: skim low-risk drafts, check facts against a source, and never pass on a number, name or rule you haven’t confirmed yourself.

Think of it like…

It’s like directions from a friendly stranger: fine for finding a coffee shop, worth double-checking before you drive to the airport. How sure they sound tells you nothing about whether they’re right.

How it works

Ask four quick questions about any answer before you use it. Together they tell you how much checking it needs.

  1. 1StakesWhat happens if it’s wrong? An awkward sentence costs nothing; a wrong price, date or legal claim can cost money or a customer.
  2. 2Fact or draftDrafts, ideas and rewrites are judged by reading them. Facts, figures, quotes and rules need a source.
  3. 3SourceDid it come from a document you gave it, or a link you can open? Check there. If it came from the model’s memory, treat it as a lead, not a fact.
  4. 4FreshnessModels learn up to a cutoff date. Prices, laws, product details and news may have changed since, unless the assistant searched and shows you where.

The more an answer matters, the closer you check. A confident tone is not evidence.

Good at, bad at

Good at

  • Rewording, summarizing and drafting from text you provide
  • Explaining well-known ideas in plain words
  • Brainstorming options you’ll judge yourself

Bad at

  • Exact figures, quotes and citations from memory
  • Recent events, prices and rule changes
  • Local or niche facts about your town, trade or customers

Where you’ll see it at work

  • A supplier contract summary. Good for finding your way around. Check each clause it mentions against the contract itself before relying on it.

  • A tax question from the bookkeeper. Treat the answer as a list of questions for your accountant, not as the answer.

  • Ideas for social posts. Low stakes: read, pick and edit. Check any claim about your products before posting.

Myth vs reality

Myth: “If it cites a source, it’s true.”

Reality: AI can cite sources that don’t say what it claims, or that don’t exist. A citation is a place to check, not proof: open it and find the sentence.

Words to know

HallucinationKnowledge cutoffRAGChain of thought
Test yourselfNo grades

An assistant tells you a supplier’s new minimum order and links a source. What should you do before telling your team?

From understanding to doing