Module 2 · Lesson
Evaluating AI-generated output
Use the same rubric every time. A document is ready only when it passes every row.
| Dimension | The question | How to check |
|---|---|---|
| Accuracy | Is every fact correct? | Highlight each date, number, name, entitlement and legal statement. Trace each one to a source document. Anything untraceable is removed or confirmed. |
| Completeness | Is anything important missing? | Compare against the brief or the template headings. |
| Clarity | Would the intended reader understand it on one reading? | Read aloud; check reading level; have someone outside HR read it. |
| Tone and brand | Does it sound like us? | Compare with the voice card and an approved sample. |
| Fairness and inclusion | Could any wording exclude or stereotype? | Apply the inclusive language checklist. |
| Compliance | Does it create a promise, omit a required statement, or conflict with law or policy? | Look for “guarantee”, “always”, “will receive”. Confirm required statements are present and verbatim. Escalate anything legal. |
| Confidentiality | Does it reveal anything it should not? | Check for personal details and internal-only information. |
WATCH OUTWhere hallucinations hide
Be most suspicious of the most specific-looking content: exact numbers of days, percentages, eligibility periods, names of laws, quotations, names of benefit providers, and web addresses. These are exactly the details a model will supply confidently when it has no source. Asking a second AI tool whether the first is right is not verification. Only the source document is.
Jot down key ideas in your own words. Saves automatically.