Skip to main content
ChatGPT: Beginner to Advanced

LESSON 5 OF 6

Check its work

BY THE END OF THIS LESSON

Catch the specific ways a confident answer goes wrong.

Fluency is not accuracy

2 options, side by side.
Fluencycalm, well-organised prose
Accuracyno tell in the writing

A fabricated detail arrives in exactly the same calm, well-organised prose as a real one. There is no tell in the writing, which is why proofreading does not catch it.

What catches it is knowing which categories are risky and checking those, rather than reading the whole answer more carefully.

Where to look

5 of 6 to check carefully.
  • Check this: Figures and statistics
  • Check this: Quotes and who said them
  • Check this: Citations and studies
  • Check this: Anything recent
  • Check this: Niche specifics
  • Safe: Rewriting material you supplied

Treat as unchecked until verified: figures and statistics, quotes and who said them, citations and studies, anything recent, and niche specifics — a small company, a local rule, one product’s current behaviour.

Lower risk: structure, explanation, rewriting material you supplied, and anything you can run and test yourself.

Two habits that do most of the work

A prompt, split into 2 parts: Separately — For each factual claim above, how could I verify it?; Supply — Supply the facts yourself whenever you can..
Two habits
SeparatelyFor each factual claim above, how could I verify it?
SupplySupply the facts yourself whenever you can.

Ask afterwards, separately: "for each factual claim above, how confident are you and how could I verify it?" Asking inside the same request invites plausible-looking citations alongside the answer.

And supply the facts yourself whenever you can. Working from a document you provided removes the whole category — it cannot invent a figure you handed it.

WORKED EXAMPLE

INVITES INVENTION

Write a section on email benchmarks for our industry, with statistics and sources.

REMOVES THE CATEGORY

[attaches the benchmark report] Write the section using only figures that appear in this document. If you want to make a claim the document does not support, write it in brackets as a note to me instead of stating it.

The first asks for statistics it does not have, so it produces plausible ones. The second makes fabrication structurally impossible and turns the gaps into a visible to-do list rather than invisible errors.

YOUR TURN

Make an answer you already have declare its own weak points.

RUN THIS

Here is an answer you gave me earlier: [paste it] List every factual claim in it. For each, say how confident you are, whether it came from something I supplied, and exactly how I could check it. Be blunt rather than reassuring.

What a good result looks like

Expect figures, dates and named sources to cluster at the low-confidence end. If everything comes back as high confidence, reply "which of these would you not bet on?"

KNOWLEDGE CHECK

Answer all 2 correctly to complete this lesson.

  1. 1. Why does reading an answer more carefully not catch fabrication?

  2. 2. Why ask for verification separately rather than in the same request?

REMEMBER THIS

Check the risky categories, ask for verification separately, and supply the facts when you can.