AI-Assisted Writing Governance

AI-Generated Claim Audit

Find factual claims introduced by an AI system, trace them to reliable evidence, and keep, qualify, rewrite, or remove each unsupported assertion.

Free editable Markdown · Editors, fact-checkers, and AI content teams · Updated July 30, 2026

Download Markdown

Accessible HTML preview

Blank template

The downloaded file contains the same fields in editable Markdown.

Audit context

Project and asset ID: [Enter]
Generated output version: [Enter]
Prompt/provenance record: [Enter]
Approved source IDs: [List]
Intended reader and consequence: [Describe]
Audit owner and date: [Name, YYYY-MM-DD]
Final candidate version: [Enter]

Claim row

Duplicate for every atomic claim.

Claim ID and location: [Enter]
Exact generated wording: [Copy]
Claim type: [Name / Date / Number / Quote / Product / Cause / Comparison / Other]
Present in approved context: [Yes / No / Expanded]
Evidence required: [Describe]
Source checked: [Reference and access date]
What the source supports: [Describe precisely]
Citation status: [Valid / Mismatched / Fabricated / Outdated / None]
Decision: [Keep / Attribute / Qualify / Rewrite / Remove / Escalate]
Final wording: [Enter]
Verifier: [Name/date]

Pattern and omission checks

Recurring invention pattern: [Describe or None]

Citations, authors, titles, journals, URLs, and identifiers were opened.
Numbers retain units, denominators, definitions, and periods.
Product steps were checked in the stated current version.
Causal and comparative language matches the evidence design.
Necessary limitations, exceptions, and counterevidence are included.
Fictional examples are labeled and contain no invented real-world authority.
Headline, summary, metadata, captions, and actions were checked.

Completion and workflow action

Unresolved claims: [List or None]
Specialist review: [Question, owner, decision]
Final source comparison: [Complete / Incomplete]
Publication decision: [Approve / Return / Reject]
Workflow response: [Prompt change / Source restriction / Smaller task / Training / Stop]
Owner and reassessment date: [Enter]

How to use this template

Preserve the prompt, approved sources, exact generated output, and final-review version.
Extract every atomic checkable claim and identify what the system added or expanded.
Trace citations and statements to primary, current, relevant evidence independently.
Apply a keep, qualify, rewrite, remove, or escalation decision to every row.
Recheck the complete final asset and use recurring failures to redesign or stop the workflow.

Compare the generated draft with supplied sources

Preserve the exact output and prompt record, then mark every name, date, number, quotation, event, product behavior, definition, comparison, causal statement, and citation. Identify which claims were present in approved context and which the system introduced or expanded. Generated text often blends a supported premise with an unsupported detail, such as adding a year, percentage, cause, or guarantee. Split compound sentences into atomic claims so one valid phrase does not shelter another. Note missing facts too: an omitted limitation or exception can make otherwise accurate wording misleading.

Trace claims independently

Open cited links and identifiers; check that sources exist, have the stated author and title, and support the exact sentence. Search for the closest primary evidence rather than trusting a plausible bibliography. Record when a citation is fabricated, mismatched, circular, obsolete, or inaccessible. For product behavior, reproduce the step in the current version. For calculations, retain units, inputs, denominator, and method. When no reliable source appears, do not “verify” the claim through repeated generated responses or the number of sites that copied the same unsupported text.

Resolve every row in the final version

Choose keep, attribute, qualify, rewrite, remove, or escalate. Add uncertainty and scope where the evidence requires them. If a generated example is fictional, label it clearly and ensure it does not accidentally resemble a real person, case, or result. Check the headline, summary, metadata, visual labels, and call to action after edits because unsupported language can survive outside the body. Record who verified the final candidate and which claims remain unresolved. If claim invention is frequent, redesign the prompt, reduce the task, strengthen the source packet, or stop using the workflow.

See the fields in context

Fictional example: invented research citation

The researcher, paper, journal, and statistic below are invented to demonstrate removal of unsupported material.

Generated claim: “A 2024 study by Dr. Mira Venn found a 72% improvement.”
Audit result: No such fictional citation exists in the approved sources, and the identifier in the output leads nowhere.
Decision: Remove the sentence rather than replace it with another generated citation.
Gap: The draft still needs real evidence for its broader performance claim.
Workflow change: Future requests must map proposed claims only to supplied source IDs.

Frequently asked questions

Are AI-generated claims always false?

No. They may be correct, partly correct, outdated, or invented. Their origin does not establish truth, so verify them with appropriate evidence.

Can I ask the same system to provide its sources?

You may collect leads, but independently open and verify every source. A confident citation list can itself contain errors.

What counts as an introduced claim?

Any checkable detail not supported by the approved context, including added scope, certainty, cause, quotation, statistic, date, or named authority.

When should the AI workflow be stopped?

Stop when risks exceed controls, restricted data is involved, reviewers cannot verify outputs, or repeated invention makes the process unreliable or inefficient.