Verifying AI Output Before You Use It
AI output is confidently wrong just often enough to be dangerous. Not usually, not everywhere — but unpredictably, fluently, and with no change of tone to warn you. Verification is the habit that makes everything else in this series safe: the error that reaches your manager, your client or your board travels under your name, not the tool’s. Here’s what to check and how much checking each task deserves.
Why this matters
Generative AI produces plausible text, and plausibility is precisely what makes errors slip through — a wrong figure looks exactly like a right one. People who skip verification don’t usually get burnt on day one; they get burnt in week six, after trust has built and reading has become skimming. The habit has to be mechanical, not mood-based, because the tool gives you no signal about when it’s needed.
What to check, always
- Facts and figures — every number, date, percentage and total, against the source
- Names and titles — people, organisations, systems, documents
- Quotes — anything in quotation marks must exist, verbatim, in the source
- Claims of fact — ‘the policy states’, ‘the report found’: open the policy, open the report
- Omissions — the hardest one: what did the summary quietly leave out that mattered?
How to check efficiently
- Source-check: read the output beside the source material, not from memory
- Reverse-check: for anything you can’t source, ask where it came from — and treat ‘it sounds right’ as a no
- Sample-check long outputs: verify a slice deeply; if the slice has problems, verify everything
- Fresh-eyes check for high-stakes items: someone who didn’t write the brief reads the result
- Final read as the recipient: would anything here surprise, mislead or overcommit?
How much is enough
Match the effort to the traffic light. Green tasks earn a careful read. Amber tasks earn source-checking and, where the audience justifies it, a second reviewer. Red tasks shouldn’t be resting on AI output at all. The discipline isn’t ‘check everything forever’ — it’s proportionate, deliberate, and never skipped just because last week went fine.
Putting it into practice
- Adopt one rule today: nothing AI-drafted leaves you unread
- Verify every number against its source on your next AI-assisted document
- Ask ‘what’s missing?’ of your next AI summary — read the source’s headings to answer it
- Set your personal amber standard: what triggers a second reviewer?
- Keep a private log of errors you catch — it calibrates your trust accurately
Key takeaways
- AI errors are fluent, unpredictable and unsignalled — verification must be habitual
- Always check facts, figures, names, quotes, claims and omissions
- Source-check beside the material, never from memory
- Scale effort to the traffic light: read, source-check, or don’t rely on AI at all
- The error that ships carries your name — that’s the whole case for the habit