Skip to content
Back to AI principles

AI principles · Chapter 6 of 8

Checking what you get

A polished answer can hide a wrong figure, a made-up source or a missing fact, and a model may agree with you just because you said it, so you check what comes back and refine until it passes.

14 minutes0 of 3 goals done

Goals

  • Find the four errors planted in a polished report, and say which kind each one is.To do
  • Choose the follow-up that fixes the most, and see what a second check finds.To do
  • Pass the check.To do

Step 1 of 4: The read

Finished is not the same as right

Chapter 5 was about how to ask. This one is about what comes back. An answer can look finished, with a title, headings and a number in a box, and still be wrong. Judging it means checking three things.

Facts: the names, numbers, dates and sources in it. Check the ones that matter against something you trust, such as your own notes, and work out key numbers yourself. Reasoning: whether each conclusion follows from the facts given. Gaps: what is left out that would change the conclusion.

Polish lowers our guard

That finished look works against you: people tend to check finished-looking answers less. As chapter 2 showed, smooth words show what was likely, not what is true. A model can even write a source that does not exist, such as a study with a believable name: chapter 3 called this confabulation. If you cannot open a source and find the claim in it, treat the claim as unconfirmed.

It may just agree with you

Models tend to agree with the person asking, even when that person is wrong. This is called sycophancy. A likely reason is training: models learn partly from people’s ratings of their answers, and people tend to rate agreement well. So if your question holds a guess, such as “this is why sales rose, right?”, the answer may hand your guess back to you as a fact. Ask it to test the guess against the facts instead.

Describe, check, refine

The habit that helps is a loop. Describe what you want, as in chapter 5. Check what comes back. Then refine: name each problem and ask again. Vague feedback such as “make it better” tends to change the wording, not the facts. Check the revision too, because a new version can fix one error and add another. Use the result only once it passes.

In the visual

You run a small café and ask an AI assistant for a short report on May’s numbers, with your notes and your own guess. The report was written for this chapter, with four errors planted on purpose: a wrong figure, a made-up source, missing context and flattering agreement. Find them and say which kind each one is. Then choose one follow-up and see what a second check finds.

Step 2 of 4: Try it: Spot the planted errors

Spot the planted errors

Find and name the four errors planted in a polished report, then choose one follow-up and see what a second check finds.

Mode:Re-createdRe-created: the café, its notes, the report, the answers to the four follow-ups and what each check finds were written for this chapter, in the way a chat assistant often answers, with four errors planted on purpose. The study the report cites does not exist. No model runs here, and nothing you do leaves the page.

Step 3 of 4: Check

Check

Put the steps of working with AI on a task in order.

Drag the items, or use the arrow buttons, to put them in order.

  1. Position 1: Check the answer’s facts, reasoning and gaps.
  2. Position 2: Describe what you want, with the facts only you know.
  3. Position 3: Name each problem and ask for a revision.
  4. Position 4: Use the result once it passes your check.
  5. Position 5: Check the revision the same way.

Step 4 of 4: What you can do now

What you can do now

  • Check an AI answer’s facts, reasoning and gaps against a source you trust.
  • Spot a made-up source, a wrong figure and missing context, even in a polished report.
  • Notice when an answer only repeats your guess, and ask it to test the guess against the facts.
  • Write a follow-up that names each problem, and check the revision before you use it.