Verify AI-generated work
A risk-based checklist for checking facts, calculations, code, and recommendations before relying on them.
AI output should be reviewed according to the cost of being wrong. A rough list of ideas needs a lighter check than production code, a financial calculation, or medical information. The goal is not to distrust everything equally, but to put verification effort where it matters.
Separate claims from presentation
Fluent writing can make an answer feel more reliable than it is. Extract the claims that affect your decision: names, dates, numbers, quotations, causal statements, and assertions about how a system behaves.
Then ask which claims can be checked directly. A polished paragraph is not a source.
Verify against primary evidence
Prefer original documentation, official records, the supplied source file, or a reproducible test. Search snippets and secondary summaries can help locate the source but should not replace it when precision matters.
For long documents, confirm that a citation supports the exact nearby claim. A real source can still be irrelevant or interpreted too broadly.
Recalculate important numbers
Check units, signs, percentages, and assumptions. Reproduce calculations with a spreadsheet, calculator, or short script. Pay particular attention to totals derived from rounded values and comparisons that use different time periods.
Run generated code
Read the diff, build the project, and run focused tests. Include edge cases and a failure path, not only the happy path. Confirm that the code did not silently expand permissions, weaken validation, or expose secrets.
Challenge recommendations
Ask what assumptions would make the recommendation fail. Compare at least one reasonable alternative and consider switching costs, maintenance, privacy, and reversibility—not only headline capability.
Keep a short audit trail
For consequential work, record the sources checked, tests run, and unresolved uncertainties. This makes later review faster and prevents a confident draft from gradually becoming an unverified “fact.”