How to prepare an AI evidence checklist
Short answer
Select the relevant AI use cases, start with unchecked items and record where supporting evidence can be found. Add an owner and date, then export Markdown/JSON or print. Checking an item records your assessment rather than independent verification.
- Select the use cases you need to reviewChoose interaction, synthetic content, deepfakes, public-interest text or other relevant areas. Human oversight and runtime documentation can support the review as practical preparation. Determine provider/deployer responsibilities separately.
- Attach a concrete evidence referenceExamples include an implementation screenshot, a versioned notice, a test report, a production release reference or a documented escalation procedure. Record the document/reference; the tool does not upload evidence files or open internal URLs.
- Assign responsibility and mark statusAll items start unchecked. Add the responsible team/person and a date, then check an item when you judge the evidence available. A checked item is not a certification. Changing use cases retains your entries only within the current tab; exports include selected items.
- Export before leavingMarkdown works well for readable repository documentation; JSON retains structured status and metadata. Print uses a readable checklist. Closing or reloading clears the in-memory entries, so download the record you need to retain.
Evidence quality depends on context
A notice screenshot can support a placement review, but it does not demonstrate machine-readable marking. A marking test can support technical review, but it does not resolve operator roles or every legal duty. Keep the scope, rule/guidance version and reviewer rationale alongside the record. Review changes to the application before reusing old evidence.