AI Output Verification Checklist

Set the stakes of the output in front of you, run the checks for that level and no fewer, and tell the person who receives it what you checked. Quick for a draft, thorough for claims, numbers and decisions.

The failure patterns are named in a standard, the NIST Generative AI Profile (NIST, 2024). Mixed research: one field experiment found people using AI on a task beyond its reach got the right answer less often (Dell'Acqua et al., 2026). The four-level ladder is practice, not research.

Use it when:

• AI-assisted work is about to go to colleagues, customers, authorities or a decision meeting.
• The output contains facts, figures, citations, legal points or anything about a person.
• You review AI-assisted material that someone else hands you.

Skip it when:

• The output is a private brainstorm you will rework anyway: a quick read is enough.
• The output keeps failing your checks: the task may not belong with AI, so look at it again in 10.7 AI Task Map.

Generative tools write fluent text whether or not it is true. The NIST profile for generative AI names this risk confabulation: confident statements, sources or figures that are wrong or invented (NIST, 2024). The same profile also flags the human side, people trusting automated output more than it deserves. Research shows why that trust is costly. In a field experiment with consultants, people who used AI on a task chosen to lie just beyond its reach were less likely to reach a correct answer than people working without it, while on tasks within its reach they did better (Dell'Acqua et al., 2026). You cannot tell from the text which kind of task you had.

The ladder on this page is practice, not research. It spends effort where an error would cost most: a quick read for drafts, tracing every claim and figure when they matter, and an independent reviewer for decisions and anything external. Its limit is that it catches errors you look for. Section 3 trains your eye for the ones that slip past.

Pick the highest level that applies. If it is between two levels, choose the higher one.

Strong answer: "Recommendation memo on the new rota, with two studies and a legal reference, going to the management group for a decision. Level 4: it feeds a decision that affects staff."

The output, and who receives it:

Level (1 to 4), and why:

Illustrative example: the role, the organization and all numbers are made up.

An HR specialist in a Norwegian municipality uses an approved AI tool to draft a recommendation on a new rota for home care staff, citing two studies and a section of the working hours rules.

Section 1: it goes to the management group for a decision that affects staff: level 4.

Section 2: the quick read passes. At level 3, one study is real but covers hospital nurses, so the claim is narrowed to "in hospital settings". Recomputing the overtime saving finds an arithmetic slip, and the figure drops. One claim about absence cannot be traced and is marked unverified.

Section 3: stale fact found: the legal reference points to a section that has since been amended. Smoothed uncertainty found and corrected.

Section 4: about 25 minutes of checking turns a confident, wrong memo into a correct and more modest one. The HR legal adviser reviews it before the meeting, and the cover note says what was checked.