The claim-check loop

Back in Unit 1 you met the hallucination — the answer that sounds perfect and is wrong. This unit is the cure.

Here's the good news: you don't need to check everything. You need a small, repeatable habit you run on the stuff that matters. It's called the claim-check loop, and it has four steps.

You can't verify everything. You can verify the one part that would hurt if it were wrong.

One idea: spot, isolate, check, decide

The loop works on an AI answer, a headline, a friend's "fact," anything.

Step	What you do
1. Spot	Find the actual claims — the statements that could be true or false.
2. Isolate	Pick the one claim that matters most if it's wrong.
3. Check	Look it up somewhere independent. One good source, not the same AI.
4. Decide	Trust it, drop it, or mark it "still unsure."

The magic is step 2. Most answers have one load-bearing claim — the fact everything else rests on. Check that one first. If it crumbles, you saved yourself. If it holds, you've spent 30 seconds well.

Do the thing

Here's an AI answer. Run the loop.

"You should plant tomatoes in early spring because they need at least 6 hours of sun, and the first President to grow them at the White House was Thomas Jefferson in 1806."

1. Spot the claims. There are three: tomatoes need 6+ hours of sun; plant in early spring; Jefferson grew them at the White House in 1806.
2. Isolate. Which matters most? If you're actually planting tomatoes, the sun and timing claims are load-bearing. The Jefferson trivia is decoration.
3. Check. Look up the sun requirement in a gardening source — not by asking the same chatbot again.
4. Decide.

Quick check. Did you notice the trick? The claim that sounds most impressive — the specific Jefferson date — is the one that matters least for your actual goal, and it's exactly the kind of specific-sounding detail hallucinations love. Checking the fancy fact first is a classic mistake. Check what you'd act on, not what sounds coolest.

Why this matters

The claim-check loop is a superpower because it's fast and it's aimed.

• Checking one load-bearing claim beats skimming ten. Aim your effort at the fact you'd actually build on.
• "Ask the same AI if it's sure" is not checking. It'll just confidently agree with itself. Real checking uses a different source.

Next lesson: how to tell a source that actually holds weight from one that just looks official — because a wrong source is worse than no source at all.