# Check before you trust: a five-minute routine

A seven-line checklist you run on any AI answer before you act on it.

Start here · v1.1 · draft · https://hayzlabs.com/start/check-before-you-trust/

In short:
- Before you act on an AI answer, ask the same chat to list every number, date, name, quote and citation in it.
- For each claim you will act on, open the source and find the sentence that says it, and run any code or sums yourself.
- Do not ask the chat if it is sure, and if checking takes longer than the task, do the task yourself.

This is step 5 of [Start here](https://hayzlabs.com/start/start-here/). The checklist is our routine, and no study tests it.

## Steps

1. **Pick an answer (2 minutes).** Use the last answer you acted on, or one you are about to act on. It should contain at least one number, name, date or source.
2. **List what is checkable (3 minutes).** Paste the prompt (the message you type to the AI) below into the same chat. It lists the claims, so you do not have to hunt for them. Delete any line that is not in the answer.
3. **Run the seven lines (10 minutes this first time, about 5 later).** Work through the checklist below on that one answer. Open at least one source.
4. **Save the checklist (2 minutes).** Copy the seven lines into a plain text file called `CHECKLIST.md`, beside the files from [Tell it who you are, once](https://hayzlabs.com/start/tell-it-who-you-are/), where you will see it before you press send.

```text title="List the checkable claims · v1.0"
List every number, date, name, quote and citation in your last answer,
one per line. Do not add anything that is not in the answer. Do not say
whether any of it is true.
```

> [!OUR RULE]
> **The seven lines.**
>
> 1. Stakes first. If this is wrong, can I undo it? Private and easy to undo, skim it. Shared with others, check it. Public or hard to undo, get a second person.
> 2. Mark every number, name, date, quote and citation.
> 3. For each one I will act on, open the source and find the sentence that says it.
> 4. Test one claim I can check in a minute. If it fails, check more of them.
> 5. For code or arithmetic, run it or work it out myself.
> 6. Do not ask the same chat "are you sure?". Use a source, a test or a person.
> 7. If checking takes longer than doing the task myself, do the task myself.

Line 3 matters most. A source that exists may not say what the answer says it does.

<!-- WIDGET sourceCheck -->

## A worked example (illustrative)

An assistant summarises a supplier's terms for an email to a colleague: "Net 30 payment, a 2% late fee, and free returns within 14 days (see clause 8)." The email is shared, so line 1 says check it. The prompt lists four claims: net 30, 2%, 14 days, clause 8. Opening the contract, net 30 and clause 8 hold, but clause 8 says 10 days, not 14, and the late fee appears nowhere. We fix the email before sending and drop the late fee. Total time: about six minutes.

## When to skip it

> [!OUR RULE]
> If you cannot check the answer and a mistake would be costly or hard to undo, use the assistant to draft your questions, and take the answer from a professional or a primary source. That applies most to health, money and legal decisions.

## Why line 6 is there

In a 2025 test of eight AI search tools, answers to queries asking which news article an excerpt came from were wrong more than 60% of the time (1,600 queries; [Tow Center, 2025](https://www.cjr.org/tow_center/we-compared-eight-ai-search-engines-theyre-all-bad-at-citing-news.php), field study). ChatGPT got 134 of 200 wrong but signalled doubt only 15 times, so the lack of a warning is no sign that an answer is right.

> [!LIMIT]
> This figure comes from citation queries on news, so it shows that errors are common there. It gives no error rate for your tasks. For a legal, medical or money decision, a qualified person makes the check.

## How to tell it worked

- The claim list holds only what was in the answer, and you opened a source for each claim you will act on.
- For each of those claims you can point to the sentence in the source that says it, or you changed or dropped the claim.
- `CHECKLIST.md` exists where you work, and you ran it on a second answer in under five minutes.

## What you keep

`CHECKLIST.md`: the seven lines, saved where you work, and one answer you have checked end to end.

## Next

- To decide whether a task should go to AI at all, see [When not to use AI](https://hayzlabs.com/guides/when-not-to-use-ai/).
- To get a second opinion that pushes back, see [Getting disagreement instead of agreement](https://hayzlabs.com/guides/getting-disagreement-instead-of-agreement/).

## Sources

1. Tow Center for Digital Journalism, Columbia Journalism Review. "We compared eight AI search engines. They're all bad at citing news." 2025. Field study. [Link](https://www.cjr.org/tow_center/we-compared-eight-ai-search-engines-theyre-all-bad-at-citing-news.php)