How to fact-check AI output
Use a repeatable verification process for claims, calculations, quotations, links and recommendations.
Fluent output is not proof. Verification should be part of the workflow, not an optional final thought.
Key takeaways
- Treat generated claims and citations as unverified leads.
- Check primary or authoritative sources directly.
- Increase review effort as the consequence of error rises.
Separate claims from wording
Highlight every factual claim, number, quotation and external reference before editing for style.
Return to original sources
Open and read primary or authoritative sources. Do not treat an AI-generated citation or search summary as confirmation.
Match effort to risk
A casual internal idea needs less checking than public, contractual, financial, health or safety information.
Build a claim ledger
Copy each material claim, number, quotation and recommendation into a short table with its source, date and verification status.
Separate “not checked” from “not found”. Missing evidence should remain visible rather than being smoothed over in editing.
Check calculations and context
Recalculate material figures independently, confirm units and date ranges, and read enough surrounding source material to avoid quoting a correct sentence in the wrong context.
Action checklist
- Highlight every externally checkable claim.
- Open the original source rather than a generated citation.
- Confirm author, date, scope and context.
- Recalculate important figures.
- Ask a qualified reviewer to approve consequential content.
A sensible next step
Apply the five-point checklist to one existing AI-assisted document before it is shared or published. Continue with the practical AI recipes, compare the reviewed AI tools or use the AI Finder to narrow your next decision.
Human review required
Responsible use reminder
Tools mentioned in context
