An AI draft can sound finished while hiding unsupported facts, invented citations, stale details, or recommendations presented as settled truth. A dependable audit does not ask ChatGPT whether its own answer is accurate. It turns the draft into a list of testable claims, checks each claim against an approved source, and sends unresolved items to a human owner.
OpenAI warns that ChatGPT can produce incorrect facts and fabricated references, even when the writing sounds confident. Its guidance recommends treating the output as a first draft and verifying important information. The workflow below makes that advice operational without turning every editing job into a research project.
Decide how much verification the draft needs
Match the review depth to the consequences of an error. A private brainstorming note may need only a quick sense check. Public instructions, pricing, product claims, legal language, medical guidance, financial guidance, safety steps, and statements about identifiable people deserve stronger controls or a qualified reviewer.
Define the intended audience, publication channel, current date, and approved source types. Also name what ChatGPT must not decide. For example, it may flag an unsupported refund promise, but the policy owner must supply the correct wording.
- Low consequence: tone, structure, and nonfactual creative suggestions.
- Moderate consequence: public educational content, comparisons, and process guidance.
- High consequence: claims that affect money, health, safety, rights, employment, or compliance.
Build a small approved source pack
Give the audit a source boundary. Collect current primary documentation, official policies, original research, product release notes, or internal records that are permitted for the task. Label each item S-001, S-002, and so on. Record the publisher, title, URL or file name, and publication or review date when available.
Do not assume a link is valid because it appears in the draft. Open it, confirm the publisher, and verify that the page actually supports the claim. Search citations can be incomplete, outdated, or incorrect, so a cited result is a lead rather than proof.
Create a source inventory for this draft.
For each supplied source, record:
- source ID
- publisher or owner
- title
- publication or last-reviewed date, if shown
- topics it can support
- limitations, scope, or version constraints
Do not evaluate the draft yet. Do not introduce sources that I did not supply.Turn prose into atomic claims
A sentence may contain several claims. “The feature is free, works on every plan, and saves two hours a week” contains three different propositions, each needing separate support. Split compound sentences before checking them.
Extract dates, quantities, quotations, named entities, causal statements, product behavior, eligibility rules, deadlines, prices, comparisons, and implied guarantees. Recommendations such as “start with a small pilot” should be labeled as advice, not verified fact.
Extract every checkable claim from the draft as an atomic statement.
For each claim provide:
1. claim ID
2. exact draft sentence
3. the smallest testable claim
4. claim type: fact, quotation, number, product behavior, policy, causal claim, comparison, or recommendation
5. consequence if wrong: low, moderate, or high
Split compound sentences. Label recommendations as recommendations. Do not verify anything yet.Create a claim-to-source ledger
Now compare each atomic claim with the approved source pack. Require a precise location such as a section heading, page number, table, or paragraph. A source that discusses the same topic is not enough. It must support the wording actually used.
Use four outcomes: supported, partly supported, contradicted, or not found. “Partly supported” is important because drafts often overstate a narrower source. Preserve the source’s conditions, plan limits, date range, population, and uncertainty.
Compare every atomic claim with only the approved source inventory.
Return a claim ledger with:
- claim ID
- verdict: supported, partly supported, contradicted, or not found
- source ID and exact location
- short explanation
- missing condition, uncertainty, date, population, or version
- recommended action: keep, narrow, remove, or send to an owner
Never create a citation or infer support from a related topic. Use [SOURCE NEEDED] when evidence is absent.Manually verify high-risk and suspicious claims
Open the cited source and read beyond the highlighted line. Check whether the draft reverses a comparison, mistakes a proposal for a final rule, applies a result to the wrong group, or uses an old product label. Verify quotations word for word and numbers against the original table or dataset.
Check dates and versions independently. Product capabilities, prices, regulations, office holders, and schedules change. If the source has no visible date or owner, treat that as a limitation. For high-consequence material, involve someone qualified to approve the final statement.
- Does the source say this directly?
- Is the source current for the draft’s date and scope?
- Were qualifiers removed during summarization?
- Does the citation identify the exact supporting location?
- Would a reasonable reader interpret the sentence more broadly than the source allows?
Run contradiction and omission checks
A sentence can be individually sourced and still conflict with another part of the draft. Search for inconsistent dates, definitions, steps, eligibility rules, and conclusions. Then look for missing exceptions or counterevidence that would materially change the reader’s understanding.
Audit the draft and completed claim ledger for conflicts and material omissions.
Flag:
- claims that contradict another passage
- one concept described with different definitions
- advice presented as confirmed fact
- missing exceptions or limitations found in the approved sources
- conclusions that are stronger than the evidence
- stale dates, version labels, or interface names
Cite the claim IDs and source IDs. Do not resolve policy or expert judgment on your own.Revise without laundering uncertainty
Use the ledger to edit the draft. Keep supported claims, narrow partially supported ones, remove claims with no value, and route consequential gaps to an owner. Do not replace “not found” with vague language that still implies certainty.
Distinguish confirmed facts from recommendations with direct wording. “The documentation states…” is different from “A practical approach is…”. If estimates or scenarios are useful, label their assumptions and avoid presenting them as observed results.
Revise the draft using the approved claim ledger.
Rules:
- preserve the original purpose and voice
- keep supported claims within their documented scope
- label recommendations as recommendations
- retain necessary uncertainty and conditions
- replace unresolved consequential claims with [OWNER REVIEW]
- do not add facts, sources, quotes, numbers, owners, or deadlines
After the revision, list every material change and the claim ID that caused it.Complete a final human pass
- Every consequential claim has an approved source and precise location.
- Quotes, numbers, dates, and names were checked in the original source.
- Recommendations are clearly separated from confirmed facts.
- Contradictions and material omissions were reviewed.
- Links resolve and point to the intended publisher.
- Unresolved high-risk claims went to a qualified owner.
- The final draft contains no hidden placeholders or invented citations.
This workflow does not make an AI draft automatically trustworthy. It makes the evidence visible, the gaps explicit, and the final human decision easier to defend.
Want more practical AI workflows with built-in checks? Join ZippByte Weekly for concise prompts and verification habits you can adapt to your own work.