AI-generated drafts can sound confident while combining accurate details with unsupported claims, outdated information, or citations that do not actually support the surrounding sentence. This AI fact-checking checklist gives publishers a repeatable way to identify factual claims, inspect source quality, verify citations, label uncertainty, and maintain a correction log before publication.
Overview
Fact-checking ChatGPT output, Claude output, Gemini output, or any other LLM draft should be treated as an editorial process, not as a single prompt. A model can help identify claims, suggest search terms, summarize provided sources, and organize notes. It should not be treated as the final authority for whether a claim is true.
The most reliable workflow separates four tasks:
- Claim extraction: identify statements that a reader could reasonably verify or dispute.
- Source checking: locate evidence from an appropriate, accessible source.
- Citation matching: confirm that each citation supports the exact claim it follows.
- Editorial judgment: decide whether to publish, qualify, replace, or remove the claim.
This separation matters because a draft may contain a real citation attached to an exaggerated interpretation. It may also cite a reputable page that is relevant to the topic but does not establish the specific number, date, comparison, or conclusion in the sentence. Verification should therefore happen at the claim level.
Use the checklist below for articles, newsletters, scripts, product explainers, research summaries, and AI-assisted updates. For a broader workflow covering prompts and source checks, see the AI Content Fact-Checking Workflow.
What to track
1. Build a claim inventory
Read the draft once for meaning, then read it again specifically for claims. Record any statement involving a fact, measurement, event, person, organization, date, legal or health implication, technical behavior, historical account, quotation, or comparison. Opinions and clearly labeled analysis may not need the same treatment, but the basis for an opinion should still be clear.
A practical claim-audit worksheet can use these columns:
| Field | What to record |
|---|---|
| Claim ID | A short identifier such as C-01 or C-02 |
| Draft wording | The exact sentence or passage under review |
| Claim type | Fact, number, quote, prediction, comparison, interpretation, or instruction |
| Risk level | Low, medium, or high based on potential reader impact |
| Required evidence | The kind of source needed to support the statement |
| Source URL and date checked | Where the evidence was found and when it was reviewed |
| Result | Supported, partly supported, unsupported, unclear, or not applicable |
| Editorial action | Keep, qualify, rewrite, replace, or remove |
2. Classify claims by risk
Not every sentence needs the same level of scrutiny. A general explanation of a familiar concept may require less checking than a claim that could influence a medical, financial, legal, safety, or purchasing decision. Give extra attention to precise numbers, superlatives, causal statements, current product behavior, and claims about what an organization or policy permits.
Risk classification helps an editor allocate time without allowing low-risk wording to distract from consequential claims. If the risk is difficult to judge, treat the claim as higher risk until it has been reviewed.
3. Compare source quality
Source quality is contextual. The best source depends on the claim. A primary document may be appropriate for a product specification, a direct transcript for a quotation, an official record for a dated event, and a reputable secondary source for context or synthesis. A search result, summary page, social post, or AI answer may help locate evidence, but it is not automatically evidence itself.
| Source type | Typical use | Review question |
|---|---|---|
| Primary source | Original records, documentation, filings, transcripts, or research materials | Does it directly establish the claim? |
| Reputable secondary source | Context, reporting, explanation, or comparison | Does it accurately represent the underlying evidence? |
| Reference or institutional page | Definitions, background, and stable descriptions | Is the page maintained and suitable for this topic? |
| Search snippet or aggregator | Discovery only | Can the original page be opened and checked? |
| AI-generated answer | Question generation and leads for research | What independent source confirms it? |
4. Check citation-to-claim fit
For every citation, ask what the source actually supports. Check the title, publication or update date, author or responsible organization, relevant passage, scope, and any stated limitations. Then compare the source with the entire sentence, not just one phrase.
Watch for common mismatches: a source describing a possibility presented as a guarantee; a general statistic used to describe a specific population; an old page used to support a current claim; or a source supporting correlation while the draft asserts causation. If a citation supports only part of a sentence, split the sentence or narrow the wording.
5. Record uncertainty instead of hiding it
Useful uncertainty labels include verified, partly verified, unclear, contested, time-sensitive, and not independently verified. The label should guide the editorial action. A time-sensitive claim may be published with a clear date and source; an unsupported high-risk claim may need to be removed.
Cadence and checkpoints
Use checkpoints throughout the editorial workflow rather than waiting until the final proof. A simple process can run as follows:
- After the first AI draft: preserve the original output, extract claims, and mark any citations supplied by the model as leads to inspect.
- During research: replace weak references with sources that directly support the wording. Save the relevant passage in the audit notes.
- Before editing for style: resolve high-risk claims first. This prevents polished language from making uncertain information appear more authoritative.
- At fact-check review: test numbers, dates, names, quotations, links, and scope. Ask a second editor to review high-impact claims when practical.
- At publication: confirm that links open, citations are placed next to the claims they support, and the article distinguishes facts from interpretation.
- After publication: record corrections, reader reports, source changes, and recurring failure patterns.
A quarterly review is a useful default for evergreen AI-assisted content, while monthly checks may be appropriate for pages that discuss changing products, regulations, platform features, prices, or active events. The schedule should reflect how quickly the underlying information can change and how serious an error would be.
For repeatable AI workflows, retain the prompt, model or tool used, input sources, output version, reviewer, and decision date. This creates an audit trail without assuming that a model's output is reliable simply because the same prompt worked previously. Prompt regression testing can help identify when a workflow begins producing different structures or weaker evidence handling; see the LLM Prompt Regression Testing guide.
How to interpret changes
When a source, citation, or claim changes, do not automatically update only the affected sentence. Recheck nearby claims that depend on the same evidence. A revised definition can affect headings, examples, summaries, metadata, and internal links. A changed product behavior can affect instructions, screenshots, and conclusions.
Use these decision rules:
- Supported and current: keep the claim, retain the citation, and note the review date.
- Supported but time-sensitive: add a date, scope, or qualification and schedule a new check.
- Partly supported: narrow the wording to what the source establishes or find additional evidence.
- Conflicting sources: describe the disagreement, prioritize the most appropriate evidence, and avoid presenting a disputed point as settled.
- Unsupported: remove the claim or rewrite it as an explicitly labeled question or possibility only if that framing is useful.
- Broken or inaccessible citation: locate an equivalent authoritative source, update the link, or remove the citation-dependent statement.
Maintain a correction log with the article URL, claim ID, original wording, reported or discovered issue, evidence reviewed, correction made, reviewer, and date. Patterns in the log are valuable: repeated errors with dates may indicate weak research prompts, while repeated citation mismatches may indicate that the draft instructions ask for sources without requiring claim-level evidence.
When to revisit
Schedule the next review when the article's subject changes, not only when a calendar reminder appears. Revisit an AI-assisted article after a quarterly interval by default, or sooner when a source is updated, a cited page disappears, a reader reports an error, a platform changes its documentation, or new evidence challenges a central conclusion.
Before each review, duplicate the previous claim inventory and compare it with the live article. Check whether any citations now point to different content, whether dates and numbers remain accurate, whether uncertainty labels still fit, and whether the article makes claims broader than its evidence. Recheck every high-risk claim even if no obvious change has been reported.
To make the process practical, finish each audit with three actions: assign an owner, set the next review date, and record the trigger that would cause an earlier review. Keep the worksheet with the draft or content record, not only in a temporary chat. For related production controls, the prompt engineering evaluation checklist can help connect editorial review with repeatable prompt testing. A disciplined claim inventory, source check, citation match, and correction log will not eliminate every mistake, but it makes errors easier to detect, explain, and correct before and after publication.