How to Fact-Check an AI Draft With a Claim Table

An AI draft can sound ready to share while adding a detail that never appeared in your source. The mistake may be small: an invented participation rule, a promise of supplies, or a time copied incorrectly. Before sharing a factual draft, split it into claims and trace each one back to the original material.

Yesterday's lesson handled conflicting notes before drafting. Today's step happens after a draft exists. The practical tool is a claim table: one row for each checkable statement, the source passage beside it, and a decision about whether the passage supports it.

A fictional notice and an illustrative draft

The following seed swap, notice, and draft are invented for teaching. They are not a real event, and the draft is a constructed example, not a response observed from a separate AI run.

Original fictional notice: Harbor Library Seed Swap. May 9, 2–4 p.m., Community Room. Bring labeled seeds if you have them. You may take seeds without bringing any. Supplies are limited.

Illustrative draft: Join the Harbor Library Seed Swap on May 9 from 2 to 4 p.m. in the Community Room. You must bring seeds to take part. Starter kits will be provided.

The first sentence carries several claims: event name, date, hours, and room. The second adds a requirement the notice actually rejects. The third adds a promise the notice does not make. A fluent tone does not tell us which claims are supported.

Check each claim against the notice

“Contradicted” and “not in source” are different. The participation rule is contradicted by an explicit line. The starter-kit sentence has no supporting line; the notice alone cannot tell us whether kits will exist. In either case, the draft should not go out unchanged.

Here is a source-faithful revision of the fictional copy: “Join the Harbor Library Seed Swap on May 9 from 2 to 4 p.m. in the Community Room. Bring labeled seeds if you have them; you may take seeds without bringing any. Supplies are limited.” This is a revision of the teaching example, not a verified public announcement. A real publisher would still confirm that the notice is current and approved.

A prompt you can reuse

Paste a non-sensitive source and a draft into your AI tool, then ask:

Make a claim table for this draft. Put each factual claim in its own row. For each row, quote the exact source words that support or contradict it, and mark Supported, Contradicted, or Not in source. If the source says nothing about a claim, write “no supporting quote” instead of inventing one. Suggest the smallest correction, but do not rewrite the full draft yet. I will check your table against the original source before sharing anything.

SOURCE: [paste the current, approved source]

DRAFT: [paste the draft to audit]

The AI's table is a review aid, not a certificate. Open the original source yourself. Check the quoted words in context, including conditions such as “if you have them” and limits such as “supplies are limited.” Inspect every name, date, time, quantity, link, and promise in the draft. If the AI misses a claim, add a row. If it produces a quote you cannot find, mark that quote unverified and investigate it. Ask the source owner about anything the source leaves unresolved.

This method only checks whether a draft matches the material you supplied. A perfectly matched draft can still repeat an outdated or inaccurate source. Check who owns the source, when it was last approved, and whether newer information supersedes it before publishing. Do not paste private records into an AI service unless you are authorized to use that service for those records.

One step to try: Take one harmless draft and its approved source. Build a three-column list—claim, exact source phrase, status—and manually verify every row before sharing the draft.

Sources and limits

• OpenAI, “Does ChatGPT tell the truth?”, checked October 25, 2026. Explains that fluent answers can contain wrong facts or fabricated quotes and recommends checking important claims and visiting sources directly.
• Anthropic, “Reduce hallucinations”, checked October 25, 2026. Recommends making uncertainty explicit, using direct source quotes, and verifying claims with citations. It does not guarantee a model's table will be correct.
• NIST AI 600-1, Generative AI Profile, MP-2.3-003, July 2024, checked October 25, 2026. Recommends fact-checking techniques for generated information, especially when sources are multiple or unknown. This is risk-management guidance, not proof that this lesson improves accuracy.

Disclosure: AI assisted research and drafting. Short Briefing constructed and checked the fictional notice, illustrative draft, claim table, and final wording. Any companion video uses synthetic narration and original project graphics. No real event or independent model performance was tested.