AI Output Review: A Claim-by-Claim Checklist

AI output review means checking a draft against the task and its sources before anyone relies on it. Check each claim, then look for missing facts and actions that need approval. Open the sources yourself. Decide whether to use the draft, fix it, or stop and ask for more evidence.

This checklist is for routine internal writing, such as status updates and summaries. It does not validate legal, medical, accounting, or hiring judgments. The worked example is made up. The process is a proposed way to review drafts, not a proven accuracy benchmark.

Separate the draft from the evidence

Keep the source and draft side by side. The draft cannot be its own proof. NIST’s Generative AI Profile , section 2.2, calls false content stated with confidence “confabulation.” This includes outputs that conflict with the input or earlier statements in the same context.

First, record the task and audience. List the sources the draft may use. If the job is to summarize approved meeting notes, remove claims not found there or mark the gap. A research task may allow outside sources, but you still need to open and check them.

Save the source version or date when it matters. If the brief changed after the draft was written, note which version you checked. Flag what needs to change. Do not quietly mix an old draft with new evidence and call it verified.

Review claims with an evidence ledger

Create one row for each claim that could change what a reader thinks or does. Check names and dates, amounts, task status, and promises. Split a sentence with several claims if the source supports only part of it.

The ledger is a working aid. You do not need a long audit report for every email; a few rows may be enough. Choose the level of detail based on what a mistake could cost and your team’s review rules.

If a draft cites a page, check more than whether the link opens. Read the cited section. Does it support the claim? Does it apply to this product and date? A real link to an unrelated page still fails the check.

Work through a flawed status update

Consider this fictional source note:

Project Cedar: the outline is complete. The design review is planned for Thursday, pending confirmation. The operations owner will request the missing image files. No delivery date has been agreed.

Now compare it with this deliberately flawed draft, written for the example:

Project Cedar is ready for delivery on Friday. The design review is confirmed for Thursday, and the missing images have been requested.

The short draft changes key facts. No model produced it; we wrote the errors to show how the ledger works.

A corrected update would read: “The Cedar outline is complete. The design review is planned for Thursday, pending confirmation. The operations owner will request the missing image files. No delivery date has been agreed.” The source is uncertain, so the update must keep that uncertainty.

Use the team prompt library example to turn these rules into reusable instructions. Keep the flawed draft as a review exercise. Label it as made up, so nobody mistakes it for evidence about a product.

Check omissions and requested actions

Check what the draft leaves out, too. A summary can contain only true claims yet omit a blocker that changes the next step. Look for missing caveats, tasks that depend on other work, clashing facts, and unassigned work.

Distinguish a request from an action. “Send the update” is an instruction. “The update was sent” is a claim about an event. If an AI tool can act in another system, check the record there. Do not rely on the tool’s claim that it finished. Anthropic’s evaluation guide likewise separates what an agent says from what changed in its environment.

Keep a draft internal until the person who owns that message accepts it. If several people or systems take part, use a workflow handoff record . Name who reviews the draft, who can act, and who handles unresolved work.

Decide whether to accept, revise, or stop

Record a clear decision. “Checked” does not tell the next person whether you found and fixed a false claim. These rules apply to the proposed writing process; they are not a certification standard.

Do not let small successes hide a serious error. A correct heading and tidy tone do not make up for a false deadline. Another model’s agreement is not proof either. Use its suggestions to find issues, then check the source yourself.

Label the final reviewed version so it cannot be confused with the first draft. If a later edit changes the facts or audience, run the relevant checks again. One review does not approve every future version.

Frequently asked questions

Does every sentence need a source link?

No. Focus on facts that matter to the task. An internal summary may cite an approved document and section rather than a public URL. Give the reviewer access to that source without exposing private files to others.

Can another AI tool review the draft?

It can suggest gaps or clashing facts to check. Treat its findings as suggestions, then return to the source. Two generated answers that agree can still both be wrong.

What if the source documents disagree?

Show the conflict and ask the person responsible which source to use. Do not let the draft pick one unless the task gives a clear, approved rule for that choice.

Sources and method

Sources were checked on October 23, 2026. NIST supports the account of false outputs; Anthropic supports the distinction between claimed and actual outcomes. LastVerified devised the ledger, Cedar example, and decision table. These aids have not been validated as a universal review method.