Today’s WorkerThink deeply. Work with AI.

Field Guide

Source-check an AI draft before you publish it

Research-based guidance; original hypothetical example

An AI-assisted draft can sound researched while hiding where its claims came from. A linked article may support only part of a sentence. A statistic may describe a different year, population, product, or comparison. A confident conclusion may be the assistant’s inference rather than the source’s.

Treat the draft as a set of claims to examine, not as a source record. This guide gives you a claim ledger for checking important statements before publication. It is research-based guidance with an original hypothetical example. AI assisted with drafting this article; a separate editor checks its cited factual claims before publication. Today’s Worker has not yet validated the method through repeated use, and completing the ledger does not guarantee that a draft is correct.

Freeze the version you are checking

Save the exact draft before you begin. If its wording changes during review, mark the changed claims for another check. Otherwise your notes may support a sentence that no longer exists.

Work sentence by sentence, but split bundled statements into smaller claims. “The program cuts bills in half and gives every homeowner a $2,000 rebate” contains at least four questions: what is reduced, by how much, compared with what, and who qualifies for which incentive. One citation at the end does not make those claims rise and fall together.

Separate factual claims from recommendations and inferences. A source can report a measurement without supporting your advice about what a reader should buy. Label the reasoning that connects evidence to a recommendation so a reviewer can inspect both parts.

Spend more checking effort where an error costs more

Not every sentence needs the same treatment. Prioritize claims that could change a reader’s health, safety, legal, financial, purchasing, or business decision. Give extra attention to numbers, quotations, product capabilities, comparisons, dates, eligibility rules, and statements that may change quickly.

NIST’s generative-AI risk profile notes that these systems can produce confidently stated false content and that risk management should be tailored to context and severity. That supports a consequence-based review. It does not supply a universal error rate or prove that every generated sentence is unreliable.

For commercial copy, the standard may also be consequential. The US Federal Trade Commission’s advertising substantiation policy says objective express and implied claims should have a reasonable basis before they are disseminated, with the appropriate support depending on factors that include the claim and the consequences of error. This guide is an editorial workflow, not legal advice, and the FTC policy should not be presented as a universal rule for every jurisdiction.

Open the source and identify it precisely

Do not stop at a search result, an AI summary, or a citation that merely looks plausible. Open the source. Record the publisher, page or paper title, date or version when available, the relevant section, and the date you accessed it. If you cannot reach the source, record that limitation instead of treating the citation as verified.

For scholarly material, identifiers and databases help establish which work you are looking at. Crossref’s REST API can retrieve registered metadata for a DOI. PubMed records expose publication types and, in covered cases, relationships involving corrections, retractions, and other status notices. Those records can help with identity and status; they do not establish that a study is true, high quality, or applicable to your reader.

Check whether a more current primary source supersedes the one in the draft. If sources materially conflict, record the conflict. Do not silently select the result you prefer.

Match the claim to what the source actually says

Compare the draft with the source on six dimensions:

  1. Wording: Is the draft stronger or more certain than the source?
  2. Population: Who or what was studied, measured, or made eligible?
  3. Product or intervention: Is the draft discussing the same thing?
  4. Place: Does a rule or result apply in the reader’s jurisdiction or setting?
  5. Time: Is the source current for the date the draft names or implies?
  6. Comparison: What is the baseline, alternative, or denominator?

Preserve qualifiers such as “up to,” “may,” “among participants,” “electricity use,” or “for qualified property.” Removing a qualifier can turn a supported statement into an unsupported one.

Google’s guidance for helpful content asks publishers to consider whether material presents information with clear sourcing, avoids easily verified factual errors, represents authorship accurately, and discloses automation when readers would reasonably expect it. These are useful editorial questions. They are not an audit standard or a promise of search rankings.

Give each claim a disposition

Finish the review with one of five labels:

“Unable to verify” is a useful result. It prevents a missing check from becoming a quiet pass. Decide whether to remove the claim, find better evidence, narrow the wording, or obtain a qualified reviewer.

Copy this claim ledger

Draft version: File, URL, or exact hash being reviewed.

Claim ID and exact wording: Copy the smallest checkable statement.

Claim type: Fact, number, quotation, recommendation, inference, or opinion.

Consequence and change rate: What could go wrong, and how quickly could the fact change?

Source identity: Publisher, title, date or version, URL or identifier, and access date.

Supporting passage or data: Record the section, table, or short passage location without copying more than needed.

Scope match: Wording, population, product, place, time, and comparison.

What the source does not support: Record the nearest tempting overstatement or unresolved limit.

Disposition: Supported as written, qualified, unsupported, outdated, or unable to verify.

Revision: Exact change required, or “none.”

Reviewer: Who checked it, and what expertise or limitation should be visible?

A ledger is not a substitute for judgment. It makes the judgment traceable and gives another reviewer a concrete place to disagree.

Work through a fictional heat-pump claim

Imagine an AI-assisted article contains this invented sentence:

Every homeowner who installs a heat pump in 2026 will cut the household energy bill by 50% and receive a $2,000 federal rebate.

The sentence sounds specific, but the cited government pages do not support it as written.

  1. “Every homeowner” — unsupported. The Department of Energy page describes heat pumps in general terms; it does not promise the same result for every home. The IRS credit has qualification and tax-liability conditions.
  2. “Cut the household energy bill by 50%” — qualified. The DOE page says modern air-source heat pumps can reduce electricity use by 50% compared with furnaces and baseboard heaters. That is not the same as a promise to cut every household’s total energy bill in half, and the page does not establish the result for every home or heating system.
  3. “Receive a $2,000 federal rebate in 2026” — outdated and unsupported. The checked IRS page describes a nonrefundable tax credit of up to $2,000 per year for qualified heat pumps, not a rebate. It says the credit can be claimed for improvements made through December 31, 2025, while a later paragraph says qualifying property must be placed in service before December 31, 2025. That internal cutoff inconsistency does not support a 2026 benefit and should not be silently resolved.

A narrower draft could accurately describe the DOE comparison and the historical credit terms, with their dates and conditions. If the article needs current 2026 incentive advice, the editor must find an authoritative current source rather than extending the older page.

Save the record with the published version

Keep the claim ledger beside the reviewed draft and record which version it covers. If a high-consequence claim changes, recheck it. If a source is volatile, set a review date or remove language that implies permanent currency.

Record authorship and automation honestly. An editor who checks an AI-assisted draft is not certifying every possible implication. State the evidence limits, especially when specialist expertise was not part of the review.

Use acceptance criteria before delegation to define what the output must include. Use this claim ledger afterward to adjudicate evidence in the completed draft. If the same sourcing failure recurs across tasks, turn the correction into a reusable instruction with Fix It Once and test whether the named mistake returns. One completed check is evidence about one draft, not proof that the method works everywhere.

Sources and further reading