Skip to main content
Your First Week with AI

LESSON 5 OF 6

Catch what it gets wrong

BY THE END OF THIS LESSON

Spot a confident wrong answer before it costs you.

Confidence is not a signal

A model does not sound less certain when it is making something up. That is the single most important thing to internalise. Fluency and accuracy are separate — a fabricated statistic arrives in the same calm, well-structured prose as a real one.

The five high-risk categories

Treat these as unverified until you check: numbers and statistics, quotes and who said them, citations, studies and case law, anything after the model's training cutoff, and niche local specifics — a small company, a local regulation, a particular product's current spec.

Low risk: structure, explanation, rewriting your own material, brainstorming, code you are about to run and test.

Two habits that catch most of it

Ask for the source, separately. Not "cite your sources" inside the same answer — models will invent plausible citations. Ask afterwards: "For each factual claim above, tell me how confident you are and how I could verify it."

Supply the facts yourself. The single most reliable pattern is to paste the real document and ask it to work on that. It cannot hallucinate a number you handed it.

WORKED EXAMPLE

RISKY

"Write me a blog post about email marketing benchmarks, with statistics."

SAFE

"Here is the benchmark report I downloaded [paste]. Write a blog post using only figures that appear in this document. If you want to make a claim the document does not support, flag it in brackets instead of writing it."

The first invites fabrication — you asked for statistics and gave it no way to get real ones, so it will produce plausible-looking numbers. The second makes fabrication structurally impossible and gives you a visible marker wherever it wanted to overreach.

YOUR TURN

Audit an answer you already received.

RUN THIS

Go back over your last answer and produce a table with three columns: the claim, how confident you actually are in it (high / medium / low), and exactly how I could verify it. Include every factual claim, including ones you think are obviously true. Do not defend the answer — just audit it.

What a good result looks like

There should be some mediums and lows. An audit that marks everything "high" is not an audit — reply with "which of these would you be least surprised to be wrong about?"

KNOWLEDGE CHECK

Answer all 3 correctly to complete this lesson.

  1. 1. Which is the strongest defence against fabricated statistics?

  2. 2. Which task carries the LOWEST hallucination risk?

  3. 3. A model states a figure very confidently. What does that tell you?

REMEMBER THIS

Fluency is not accuracy — supply the facts, or verify the ones it supplies.

Go deeper — read the full guide