Check the work

01What this module is about

AI writes well. That's the trap. A smooth, confident answer feels right, so it doesn't get checked, and the one wrong date goes out to forty people.

This module shows you where AI mistakes hide, how to make fewer of them happen, and how to catch the rest quickly. You don't need to check everything. You need to check the things that matter, every time.

By the end you'll have a short checklist for your own project, and you'll have used it.

02Why a polished answer can still be wrong

AI sometimes writes things that sound right but aren't. This is called a hallucination. Anthropic says even its most advanced models can do it. OpenAI, Google and xAI all say the same about their tools, and all four tell you to check important answers.

The usual suspects:

• Facts that aren't in your files. A date, a name or a price that it filled in because it seemed likely.
• Small changes to real facts. 22 March becomes 29 March. The car park becomes the church.
• Wrong sums. Totals and percentages that look right and aren't.
• Sources that don't back it up. A link or quote that doesn't say what the AI claims it says.
• Skipped instructions. The word limit, or the word you banned.

The tone doesn't tell you anything. A wrong answer sounds just as sure as a right one.

03Four habits that catch most mistakes

1. Give it permission to say "I don't know". Anthropic says this one change cuts false information a great deal. If it isn't allowed to say it's unsure, it fills the gap with a guess. Your first starter line from Module 1 does this.

2. Ask where each fact came from. Ask it to point to the line in your files, or the source, for every fact. Anything it can't back up gets removed. This is Anthropic's own advice for checking claims.

3. Turn your problems into a checklist. You've been keeping a problems list since the introduction. Each problem becomes a yes-or-no question you can check in seconds.

4. Ask for an honest report at the end. What did it do? How does it know it's right? What didn't it check? What did it assume? The answers tell you where to look.

One more trick for anything important: ask the same question in two fresh chats and compare. Anthropic suggests this too. Where the answers differ, at least one of them is guessing.

04Text to copy

Your starter lines from Module 1 already tell the AI to say when it isn't sure, and to say where each fact came from. That cuts down mistakes. It doesn't end them. Anthropic says so too. So you still check.

To check an answer's facts, paste this in the same chat:

Check your answer. For every fact, name, date and number, show the line from my files or the source it came from.
If you can't find one, remove that fact and mark the gap with [ ].

To get an honest report at the end of a task:

Before I use this, tell me:
What you did.
How you know it's right.
What you didn't check.
What you assumed.

Here's a checklist for the walking club newsletter. Yours will be different.

Does every walk date match the Reference file?
Does every meeting point and time match the Reference file?
Is it under 300 words?
Is it free of the banned words?
Is it signed off the right way?

06Common problems

• It made up a fact anyway. Check that the starter lines are in your tool's Instructions box, and that you started a new chat. It can still happen even then, which is why you check.
• It says it checked, but it didn't. Ask it to show the check: "List every date and the line it came from."
• The checklist takes too long. Cut it to the five checks that matter most. A short checklist that gets used beats a long one that doesn't.
• You can't find a source to check against. Then the fact isn't checked. Leave it out, or mark it as unconfirmed.

07Keeping it current

• Add to your checklist whenever a new kind of mistake gets through.
• Remove checks that never catch anything.
• Check harder when it matters. A note to yourself needs a glance. Anything that goes to other people needs the full checklist.