Mistakes

Verification mistakes

The ways people over-trust AI output — and the cheap checks that prevent it.

Knowledge cards
Mistake
Check Against the Source, Not the Summary
The mistake of accepting Claude's summary, quote, or paraphrase of a document as accurate because it reads well — without checking the specific claims back against the source. A fluent summary can still misstate, over-generalise, or invent a detail the original never contained.
L2 · Context engineering
Mistake
Don't Outsource the Final Judgement
The mistake of letting Claude make the call rather than inform it — shipping its output because it seems reasonable, without the review and judgement that the decision actually needs. The model is there to draft, analyse, and pressure-test; the decision and the accountability stay with you.
L2 · Context engineering
Mistake
Don't Ship Code You Only Read
The mistake of accepting AI-generated code because it looks correct on the page — well-structured, sensibly named, plausible logic — and shipping or committing it without actually running it, testing it, or exercising the edge cases. Readable is not the same as working.
L2 · Context engineering
Mistake
Don't Trust Confident Specifics Without Checking
The mistake of taking Claude's most confident-sounding specifics — exact figures, quotes, citations, API names, legal or medical claims — at face value because the fluent, assured tone reads as authority. Confidence is a property of the prose, not evidence the fact is right.
L2 · Context engineering
Verification mistakes

Fluent is not the same as correct

The most dangerous mistake with any capable model is trusting an answer because it reads well. Claude is careful and tends to stay grounded, but it can still be confidently wrong, and its polish makes it easy to skip the check. Treat fluency as a neutral signal, not evidence — a well-written paragraph tells you nothing about whether it is right. The people who get burned are the ones who mistake a polished answer for a verified one, especially on facts, figures, and code they did not test.

Cheap checks that prevent expensive errors

Verification does not have to be heavy. Ask Claude how it reached a claim, and the reasoning often exposes a weak link. Ask 'what would break this?' to surface the edge cases. For facts, request sources; for code, run it or write the test. These checks cost seconds and catch the errors that fluent writing hides. Build them into how you work — a quick verification pass on anything that matters is far cheaper than shipping a plausible mistake.

Frequently asked
Does Claude make mistakes?

Yes. It is careful and tends to stay grounded, but it can still produce confident errors, especially on facts, figures, and untested code. Fluency is not evidence of accuracy.

How do I verify Claude's answers cheaply?

Ask how it reached a claim, ask 'what would break this?', request sources for facts, and run or test code. These checks take seconds and catch errors that polish hides.

What is the biggest verification mistake?

Trusting an answer because it sounds confident and well-written. Building a quick check into your workflow for anything that matters prevents most over-trust errors.