Checking a claim in two minutes

A claim has passed the triage. Now what?

The method has to be fast, because a slow one gets skipped when it matters most. Two minutes, four steps.

1. State the claim precisely

Write down the exact claim, in one sentence, with the numbers and qualifiers intact.

This sounds like a formality and it is the most useful step. Half the time you discover you cannot state it precisely — which means you did not actually know what was being asserted, and that was the real problem.

“Remote work reduces productivity” is not checkable. “A 2023 study found a 17% productivity drop among fully remote knowledge workers” is, and stating it that way immediately raises the questions that matter: which study, which workers, measured how.

2. Go to the primary source

The primary source is the thing itself: the paper, the legislation, the company filing, the dataset, the original statement. Not an article about it, and not a summary of the article.

Every layer between you and the source is a layer where a qualifier was dropped — E4·L3 and E5·L7 both showed that happening. By the third retelling, “in this sample, under these conditions” has become a general law.

Search for the specific thing. If you cannot find a primary source at all, that is itself the answer: the claim may not have one.

3. Read the claim in place

Find the actual sentence and read what surrounds it.

This is where most failures surface, and they are rarely dramatic. The study exists, and it covered a different population. The law exists, and the clause applies only above a threshold. The quote is real and was said about something else.

The claim is not fabricated. It has been detached from the conditions that made it true, which is both more common and harder to notice than invention.

4. Read laterally

If a source is unfamiliar, do not evaluate it by reading more of it. A professional-looking site is trivially easy to produce.

Open a new search and find out what other people say about that source. This is lateral reading, it is what fact-checkers actually do, and it takes about thirty seconds — far less than reading an About page written by the people you are assessing.

Never verify a model with the same model

The most common mistake in this whole module.

“Are you sure?” is not verification. The same process that produced the answer produces the reassurance, and agreement between them is not evidence of anything. Worse, models are agreeable under pressure — push back on a correct answer and it will frequently apologise and produce a wrong one.

A different assistant is only slightly better: overlapping training data means correlated errors, not independent confirmation.

Independence means a source that does not derive from a language model. That is the whole point.

Try it now (7 minutes)

Take a claim you have seen recently. State it precisely, find the primary source, and read the sentence in context.

Then ask an assistant “are you sure?” about something it told you correctly, and push back once. Watch what happens. That demonstration is worth more than this lesson.

Check your understanding

1. Why is asking 'are you sure?' not verification? Models cannot assess their own outputs at all The same process produces both answer and reassurance, and models are agreeable under pressure It uses another message from your allowance

2. Lateral reading means: Reading further down the same page Leaving the page to see what other sources say about it Skimming several claims at once

3. The most common verification failure is finding that: The source was entirely fabricated The source is real but the claim was detached from its conditions The source is behind a paywall

Recap

State the claim precisely — often that alone dissolves it — then find the primary source, read the sentence in its context, and read laterally about any source you do not know. Never ask the model that produced an answer to confirm it, and treat a second assistant as correlated rather than independent.