Which claims need checking

You cannot check everything, and L1 argued you should not try. So the skill is deciding, quickly, what earns the effort.

Two questions and a short list of usual suspects will handle nearly all of it.

The two questions

If this is wrong, what happens?

Nothing — you were curious. Something small — a slightly worse decision. Something serious — money, health, a legal position, someone’s reputation, your own credibility in front of people who matter.

How easily could it be undone?

A wrong fact in a draft you will revise is nearly free. A wrong figure in a board paper, a filed document, a published article or a message already sent is not.

High consequence and hard to reverse is where your budget goes. That combination is rarer than it feels, which is exactly why this triage is worth doing rather than defaulting to anxiety.

The claims that fail most often

Independently of consequence, some categories are simply less reliable. Treat these as needing a check whenever you intend to use them:

Citations and references. The classic failure. A plausible author, a plausible journal, a plausible year, and no such paper. Never repeat a citation you have not opened.

Numbers. E5·L5 covered why. Anything computed in prose, any statistic, any percentage without a stated base.

Quotations. Who said what, in what words. Real people get given paraphrases, and paraphrases get sharpened into things that sound better than what was actually said.

Attributions. Which company did what, which country passed which law, which study found which result. Adjacent-but-wrong is the common failure here.

Legal, medical, financial specifics. Not general explanation — specifics. Notice periods, dosages, thresholds, deadlines, eligibility. These vary by jurisdiction and change, and confident wrongness is expensive.

Anything recent. Training data ends somewhere. Beyond that, an assistant either lacks the information or retrieved it — and E4·L4 covered how retrieval goes wrong.

Being told what you wanted to hear

One more trigger, and it is the one people skip.

If an answer is unusually convenient, check it. It confirms your position, it says the difficult thing is unnecessary, it means the deadline is fine.

This is not because models flatter you — although they somewhat do. It is because a convenient answer is the one you will not scrutinise, so an error there travels furthest before anyone catches it.

What “checking” costs

Less than people assume, which is why the triage matters more than the technique.

Opening a citation: thirty seconds. Confirming a number against a source: a minute. Checking a legal specific for your jurisdiction: a few minutes.

Compare that to being wrong in public. The economics are not close for anything that passed the two questions.

Try it now (6 minutes)

Take the last substantial answer you got from an assistant. List every distinct factual claim in it — there will be more than you expect.

Mark each against the two questions. Check the ones that qualify. Note how few of them there were.

Check your understanding

1. The two triage questions are: Is it interesting, and is it well written? What happens if it is wrong, and how easily could it be undone? Is it recent, and is it long?

2. Which category should always be checked before use? General explanations of how something works Citations and references Suggestions for what to read next

3. Why check an answer that is unusually convenient? Convenient answers are usually fabricated It is the answer you will not scrutinise, so an error there travels furthest Models are programmed to flatter users

Recap

Ask what happens if a claim is wrong and how easily it could be undone, and spend your attention there. Always check citations, numbers, quotations, attributions, legal and medical specifics, and anything recent — plus any answer that is suspiciously convenient, because that is the one you will not question on your own.