How to Check an Output You Can’t Verify Yourself

The output that costs you something is never the one that looks wrong. Obvious junk gets caught: the summary that missed the point, the email in a voice that isn’t yours, the analysis that clearly answered a different question. You spot those in four seconds and re-run the prompt. Nothing lost.

The expensive one is the output that reads beautifully, states a figure to one decimal place, cites a standard by name, and is wrong about one of them. It goes into the deck. Someone in the room knows the real number.

If you use Claude for professional work, you’ve probably landed on a defense that sounds responsible: read it carefully before you send it. That defense has a structural limit, and the limit is worth understanding before you rely on it any further.

Re-reading tests the wrong property

When you re-read an output “carefully,” you’re checking whether it hangs together. Does the argument follow, does the number look like the kind of number that belongs there. You’re testing plausibility.

Plausibility is the one property the output is certain to have. Claude produces fluent, well-formed language from patterns in language; that’s the job it does, and it does it whether or not the underlying fact is real. When plausible and true come apart, nothing in the text marks the seam. The sentence with the invented figure has the same cadence, the same confidence, the same clean grammar as the twelve true sentences around it. So a careful re-read applies your plausibility detector to a text that has been optimized for plausibility. You will pass it, and your attention was never the problem: the check is aimed at a property that carries no signal.

There’s a second thing re-reading can’t catch, and it’s quieter. Claude leans toward agreeing with you. The cause isn’t mysterious, and it isn’t malice: these models are trained with reinforcement from human feedback, and people rate agreeable answers more highly than disagreeable ones. Optimize for what raters like and you optimize, partly, for telling people what they want to hear. Which means Claude is a poor judge of your own ideas. It’s mildly tilted toward liking them. Re-reading an output that agrees with you feels like confirmation. It’s closer to an echo.

The rule under both: a check has to come from outside the text. If your verification never leaves the page, it isn’t verification. It’s a second reading with the same eyes.

Not every claim carries the risk

Verifying everything from outside the text is a fine principle and a terrible practice. You’ll do it for two weeks, it’ll eat the hours the system was supposed to give back, and you’ll quietly stop. The professional who checks everything out of anxiety abandons the tool on roughly the same schedule as the one who checks nothing.

So the real skill is triage: knowing which sentences in front of you carry the risk. Three questions do most of the sorting, and you can run them in about thirty seconds.

Could this be wrong in a way the page can’t show me? Some claims carry their own evidence. “Your Q3 churn was 4.1%” pasted from the spreadsheet you supplied is checkable against your input. “The industry benchmark is around 5.5%” is not — it either came from somewhere or it came from the shape of the sentence, and the text looks identical either way. Circle the second kind. Numbers, dates, names, regulatory references, anything specific that arrived from outside what you gave it.

Where did this come from? Ask it of every source the output cites. If you handed Claude the source, you can check the claim against it. If you didn’t, ask where it came from — and treat a citation that appeared from nowhere as unverified until you’ve found it yourself. Not doubted. Unverified. Different thing, different action.

What happens if I’m wrong and nobody catches it? This is the dial. A rough draft for your own eyes can be wrong at no cost; you’ll find out in the next pass. A number in a client proposal can’t. The question isn’t how much you trust the output, it’s how reversible the mistake is. Reversible and cheap earns a glance. Irreversible and public earns a check against a primary source and a second reader, every time, with no allowance for the fact that it reads well.

Most of your work sits between those. That in-between isn’t a shrug — it’s where you decide once, per type of work, instead of re-deciding at the moment you’re about to hit send, which is the worst possible time to be making that call.

What a check that works looks like

Go to the source, not to the output. Open the document, the spreadsheet, the regulation. Read the claim against the thing it’s a claim about. If the source doesn’t exist or you can’t find it in eight minutes, the claim doesn’t survive — you strike it or you rewrite the sentence without it. The eight-minute rule matters. Without a stopping point you either check forever or you talk yourself into “it’s probably fine.”

Ask for the case against, before the verdict. Not “what do you think of this plan?” — you already know the answer to that one. Try: list three risks or weaknesses I haven’t accounted for, then give your assessment. Making disagreement the assignment gets you something the agreement never would. And if Claude comes back approving of everything, treat that as a reason to push, not a result.

Move part of the check into the prompt. This is the habit in the set that pays back most, and it costs one line. Ask Claude, up front, to mark anything it isn’t confident of, to flag when it’s inferring rather than reading from the data you gave it, and to state the assumptions under a recommendation before making it. None of this removes the need to verify. What it does is aim it. Instead of re-reading four pages with uniform suspicion — which goes glassy by paragraph three — you’re checking the six places already marked. Depth is set by the stakes. The flags tell you where.

One more thing worth knowing if you work with long documents. Researchers testing retrieval over long documents found that recall depends on where the information sits: weakest for material buried in the middle, strongest at the beginning and the end. That was measured on retrieval and question-answering over documents, not on multi-turn conversations, so take it as a reason for care rather than as proof of what happens in a long chat. The practical version is cheap either way: put the documents high, put the instructions that matter at the end, and when something long comes back, read the last section against your original instruction, looking specifically for the constraint that quietly stopped holding. That one is a real check — you’re comparing the text against something outside itself.

The habit under all of it

Claude states what it knows and what it’s guessing in the same steady voice. The certainty in the prose is not information about the world. It’s a property of the writing.

Which means the confidence has to come from you — from the source you opened, the assumption you made it state, the number you looked up. None of that is distrust. Blanket suspicion is just as unusable as blanket faith, and it fails the same way: you stop doing it. Calibrated trust is the middle position, and it’s a decision you make per output, from what the thing is and what you’re about to do with it.

I go further into this in Part 5 of Stop Using Claude AI Like a Magic 8-Ball, where it gets a full framework. But the part that matters most fits in a sentence: your check has to leave the page.