Beginner AI Output Verification Worksheet

Use this worksheet when an AI answer includes facts, recommendations, instructions, or details that someone may rely on. It helps you turn “sounds right” into a review decision.

How to Use This Worksheet

Use it in three passes:

1. Capture the output snapshot so the review has a clear purpose.
2. Extract the claims that matter and choose the right verification depth.
3. Label confidence, revise unsupported claims, and decide whether the output is ready to use.

You do not need the full worksheet for every private brainstorm. Use the full flow when someone may act on the output or repeat it as fact.

Output Snapshot

• What did you ask for?
• What did the AI produce?
• Who might read or use the output?
• What could go wrong if it is inaccurate?
• Is this private draft, internal work, or something that may be shared?

Claim Extraction

Mark the claims that matter.

Look for:

• names
• dates
• numbers
• links
• quotes
• citations
• owners
• deadlines
• recommendations
• tool steps
• policy statements
• summaries of source material

Copy important claims into the table.

Verification Depth

Choose a depth for each claim.

Use these definitions:

• No check: private, low-risk brainstorming only.
• Light check: read for fit, remove obvious unsupported details, mark uncertainty.
• Source check: compare important claims to source material or a system of record.
• Expert check: route to a qualified owner before use.

Worked Example

Scenario: an AI answer says a draft project note should tell users that “the new request form opens on Monday, requires manager approval, and replaces the old email process.”

Output snapshot:

• Asked for: a short draft note explaining a process change.
• AI produced: a polished announcement with dates, approval steps, and a replacement process.
• Reader: internal users who need to know how to submit requests.
• Risk if inaccurate: users may follow the wrong process or miss a required approval.
• Intended use: internal draft, not ready to send.

Claim extraction:

Verification depth:

Decision: partially verified. The draft can be revised, but it should not be sent until the process owner confirms the replacement language.

Fast Path (Low-Stakes Output)

Use the short path only when the output is private, exploratory, and low consequence.

• Read once for obvious unsupported specifics.
• Remove exact numbers, names, links, or claims you did not provide.
• Add “draft” or “unverified” if the output might be reused later.
• Do not share it as fact until it goes through the full worksheet.

Red Flag Review

Slow down when the output includes:

• exact numbers or percentages without a source
• quotes you did not provide
• links, citations, or titles you have not opened
• recent information without a source date
• confident advice in high-stakes topics
• tool steps, commands, methods, or menu paths you have not tested
• assumptions presented as facts

Technical hallucinations can be especially tempting because they look precise. Precision is not the same as proof. Treat exact commands, menu paths, configuration names, dates, and citations as claims until checked.

Red flags found:

• [List red flags or write “none found.”]

Hallucination Types Quick Reference

Watch for these patterns from the verification lesson:

• Factual: a name, date, number, quote, or source detail is wrong or unsupported.
• Citation or link: a title, URL, quote, or source reference cannot be verified.
• Technical: a command, field, setting, workflow step, or API behavior sounds precise but has not been tested.
• Outdated: the answer may have been true once, but no current source or date is provided.
• Overconfident synthesis: the answer hides disagreement, uncertainty, or missing source material.

When you see one, lower the confidence label until the claim is checked.

Confidence Label

Choose one label for the output.

• Verified: important claims were checked.
• Partially verified: some important claims were checked, and gaps are listed.
• Unverified: the draft may be useful, but important claims were not checked.
• Escalated: the draft needs expert or owner review before use.

When in doubt, choose the lower confidence label. It is easier to raise confidence after review than to undo a confident unsupported claim after it spreads.

Selected label:

Reason:

Revision Actions

Before using the output, decide what to do with unsupported claims.

• Keep because it was verified.
• Rewrite as a question.
• Label as “needs verification.”
• Remove because it is unnecessary.
• Escalate to a qualified owner.

Actions taken:

• [List changes.]

Sign-Off

• Reviewer:
• Date:
• Intended use:
• Sources checked:
• Claims still uncertain:
• Escalation owner, if any:
• Final decision: use / revise / verify further / escalate / do not use

Reference Links

These public references are useful starting points for deeper study. They are linked for attribution and further reading; the worksheet above is synthesized as original LIW training guidance.

• NIST AI Risk Management Framework - useful for matching review effort to risk.
• People + AI Guidebook - useful for trust and user experience patterns around AI systems.