Who Is Fact-Checking Your Manuscript?

Before approving an edited sentence, ask one more question.

Who checked whether it was true?

Agree the fact-checking scope in writing before the editorial pass. Give each claim a named checker, keep the evidence beside the exact wording, and return every unresolved question to a named owner for a publication decision.

If you are working alone, put your own name in those fields. The next reader is not a fact-checker you forgot to hire.

Use the following workflow to make the checking scope, evidence, and unresolved decisions visible.

An Edit Is Not a Verification Agreement

Begin with the work you are actually commissioning.

Turner Proofreading advises that checking every fact should not be assumed unless fact-checking is explicitly part of the author agreement. That is a practitioner's reason to have the conversation, not a universal rule about what every editor does.

“Will you check the facts?” sounds clear until you have to act on the answer.

Ask instead:

Which claims will you check, against which materials, and what will you return when you cannot establish an answer?

Name the manuscript version and chapters covered. Will the editor compare quotations with transcripts? Recalculate percentages? Check dates against original records? Inspect captions and tables? Search for missing sources?

Agree the time available, the exclusions, and who handles what falls outside the scope.

Here is an invented scope note for the example we will follow:

Manuscript: version 4, chapter 3, including its table and caption. Jo, the editor, will compare numerical claims and quotations with the source files Alex, the author, supplies, and record discrepancies in the claim ledger. New reporting and searches for missing evidence remain Alex's responsibility. Jo will return unanswered questions to Alex before proof approval. Alex will resolve them or remove or hold the affected text; revised claims go back to Jo for comparison with the evidence.

Adapt this as an operational note, not contract language to copy unquestioningly. Agree fees, deadlines, access to source material, and any additional work directly.

What matters is that the division of labor is visible.

“Edited” should not quietly come to mean “externally verified.”

Agreement Inside the Draft Is Only the First Check

Keep two jobs separate.

Internal consistency: Do the manuscript's own statements agree? Does the chapter use the same number as the table? Does the introduction describe the same group as the conclusion?

External verification: What evidence supports the statement? Does the source establish this number, for this group, over this period, with this meaning?

Imagine a chapter, table, and caption all saying that 60% of workshop participants improved.

Perfect agreement.

But you still do not know whether the underlying record describes all participants.

For numerical claims, Nonfictioness recommends returning to the original source of the data rather than relying on articles repeating it. The author also describes printing tables or graphs alongside their sources and ticking off individual figures.

Start with that comparison. Then read the sentence around the number.

What population does it name? What outcome does it describe? Is it reporting what people said, or asserting what happened?

Check the digits. Then check what the sentence has made them mean.

One Percentage, Two Different Claims

Everything in this example is invented for teaching: the workshop, source file, figures, and people. It is not a real study.

Suppose Alex's chapter contains these sentences:

Sixty percent of workshop participants improved their writing. The workshop caused that improvement.

The supplied fictional source is workshop-feedback-v1, page 2, Table A and its note:

“Of 200 participants invited to complete the feedback form, 100 responded. Of those respondents, 60 selected ‘I feel more confident about my writing.’ This form did not measure writing quality or establish what caused changes in confidence.”

Before rewriting, put the claims beside the evidence:

Exact manuscript claim	What the invented source supports	Editorial decision	Revised claim
“Sixty percent of workshop participants improved their writing.”	60 of 100 respondents reported greater confidence. The denominator is respondents, not all 200 participants; the answer concerns confidence, not measured writing quality.	Narrow the population and describe the reported outcome accurately.	“Of the 100 participants who answered the feedback form, 60% reported feeling more confident about their writing.”
“The workshop caused that improvement.”	The supplied note explicitly does not establish causation.	Seek separate substantiation; hold this sentence out of the publication text while the question is open. Remove it if adequate evidence is not obtained.	No replacement causal claim.

The revised first claim says less.

It should.

In this invented chapter, Alex wanted the workshop to be the turning point. The stronger sentence made that argument easy. The source supports a smaller statement about reported confidence.

Now Alex has a writing decision, not just a correction to accept. Can the chapter stand with that limitation? Does it need more reporting? Or does the argument itself need to change?

Keep that decision with the author. Do not ask an editor—or an AI—to make the stronger sentence sound defensible while leaving its unsupported meaning intact.

But it should not become “30% improved,” either. In this invented record, 60 is 30% of the 200 invited participants. The source still does not measure improvement or tell us how the nonrespondents felt.

And Alex cannot rescue the second sentence by writing “The workshop apparently caused the improvement.”

An adverb cannot supply the evidence the source is missing.

You are not trying to make the original assertion harder to challenge. You are deciding what the material permits you to say.

Give Each Claim a Record

Use one ledger entry per checkable assertion. If a sentence contains claims that need different evidence, split them, as we did above.

Copy this template into a document or use the fields as spreadsheet columns:

Claim ID:
Manuscript version and location:
Exact manuscript claim:
Source identity: creator, title, date/version, URL or file reference
Passage/data location: page, table, row, timestamp, or paragraph
Source passage/data and what it supports:
Assigned checker: name
Status and unresolved question:
Next-action owner, action, and deadline:
Editorial decision: substantiate / narrow / attribute / remove / hold
Revised wording and manuscript location:
Recheck: name, date, result

Keep the source's identity separate from your reading of it. A document title belongs in one field. What you think it proves belongs in another.

A status should describe the evidence, not how reassured everyone feels.

Use not checked, supported as written, partly supported, contradicted, or unresolved. Record “revision awaiting recheck” separately when necessary.

Do not overwrite the original claim when you revise it. Keep the comparison, then record the new wording and the human recheck.

“Supported as written” belongs only to the exact version someone inspected. It does not travel automatically with the next revision.

An Unanswered Comment Needs an Owner

Turner Proofreading describes searching with keywords from the manuscript sentence and advises leaving the author a comment when no clear answer emerges, rather than letting a search consume the editing session.

Make that comment something the author can act on. Give the invented causal claim the ID C3-02. Jo's handoff could read:

Alex — C3-02 remains unresolved. I checked workshop-feedback-v1, page 2, Table A and its note. It records reported confidence, not measured improvement or causation. Please supply separate supporting evidence before proof approval if you want to retain a causal claim. The sentence is held out of the publication text pending that decision.

Alex now knows what Jo checked, what the material did not establish, and what needs to happen next.

Use the same editorial choices throughout the ledger:

• Substantiate: obtain evidence that supports the actual assertion, then inspect it.
• Narrow: reduce the claim to the population, period, outcome, or relationship the evidence supports.
• Attribute accurately: state who reported or alleged something when that is what your evidence establishes. Do not turn their statement into your independent finding.
• Remove: take out an assertion you cannot responsibly support.
• Hold: keep the affected passage out of publication while the required work remains open.

A serious unsupported allegation needs particular care. Attribution or an “unverified” label is not publication clearance. Keep it held or remove it pending adequate substantiation and any necessary specialist review.

Leave the uncertainty visible in the editorial record.

That is not permission to publish through it.

If AI Helps, Give It the Clerical Work

Use the same evidence requirement whether the manuscript was drafted with AI or without it. A sentence does not deserve a lighter check because you remember writing it. A tool's contribution does not give it greater authority, either.

If you choose AI support, test one bounded job: organizing candidate claims and copying passages from source material you supply. This follows the division of labor explored in using AI on a book without giving up creative control.

You might ask:

Using only the manuscript excerpt and source extracts below,
propose claim-ledger entries.

Copy each candidate manuscript claim exactly.
Copy any potentially relevant supplied source passage exactly,
with its supplied file name and location.
If no passage is supplied, write “no supplied passage.”

Do not invent sources, quotations, data, or locations.
Do not certify accuracy or mark claims verified.
Leave checker and decision fields for human completion.
Mark every entry “not checked.”

Then inspect every entry yourself.

Check the manuscript wording, source identity, copied passage, location, and proposed connection. Read the surrounding source context. Add claims the tool missed. Reject mismatched passages rather than repairing them from memory.

What comes back is a candidate record, not a completed check.

Comparing the sources and deciding what can be published remain human work.

Before You Hand It Back

Send the manuscript, ledger, and accessible source files together. Before the handoff, check:

• The version, checking scope, exclusions, and responsible people are clear.
• Revised claims have been rechecked, including their appearances in tables and captions.
• Unresolved claims have an owner and deadline; held or removed claims are absent from publication text.
• The person giving final publication approval can see the evidence and the remaining decisions.

For Alex, the important result is not a row marked complete. It is knowing whether to keep the smaller, supported sentence and what that choice does to the chapter.

Use the record to make that judgment visible. Where the source cannot carry the sentence, be willing to change more than the wording.

Sources

• Nonfictioness — Fact-checking your work
• Turner Proofreading — How to Fact-Check in Nonfiction Editing Without Going Down a Rabbit Hole
• ClaraMuse — AI and Your Book: How to Use AI Without Giving Up Creative Control