AI-Assisted Writing Governance

AI Draft Human Review Record

Document the substantive human checks, corrections, additions, source decisions, disclosures, and publication judgment applied to an AI-assisted draft.

Free editable Markdown · Editors, compliance teams, and publishers · Updated July 30, 2026

Download Markdown

Accessible HTML preview

Blank template

The downloaded file contains the same fields in editable Markdown.

Draft and reviewer

Project and asset ID: [Enter]
AI draft version and location: [Enter]
Provenance and prompt record: [Enter]
Approved brief and source packet: [Enter]
Intended reader and outcome: [Describe]
Reviewer name, role, and expertise: [Enter]
Review date and final candidate version: [Enter]
Risk tier and required specialists: [Enter]

Review findings

Important omission: [Describe or None]
Expert escalation: [Question, owner, decision]

The draft directly serves the approved reader task.
Every material factual claim is supported by opened, current evidence.
Names, quotations, figures, dates, links, and product behavior were checked.
Required limitations, uncertainty, counterevidence, and context are present.
No invented authority, source, personal experience, or guarantee remains.
Voice, terminology, representation, and accessibility fit the audience.
Rights, privacy, sensitive information, and disclosure requirements were reviewed.

Material decision log

Duplicate for each significant change.

Draft section or claim: [Enter]
Decision: [Accept / Rewrite / Add / Remove / Escalate]
Reason: [Evidence / Accuracy / Scope / Voice / Rights / Safety / Other]
Source or policy used: [Reference]
Human contribution: [Describe the substantive change]
Decision owner: [Name/date]
Verified in final version: [Yes / No]

Approval

Unresolved issues: [List or None]
Disclosure decision and text: [Enter]
Attribution and source links: [Confirmed / Changes needed]
Final comparison completed: [Yes / No]
Decision: [Approve / Approve with named condition / Return / Reject]
Approver and version: [Name/version/date]
Retention and access: [Enter]
Reopen trigger: [Material edit, new source, product change, incident]

How to use this template

Link the exact AI draft, provenance, approved brief, source set, and review version.
Evaluate usefulness, evidence, omissions, claims, voice, rights, safety, and audience fit.
Record every material human addition, correction, rejection, and escalation with a reason.
Reconcile the final candidate with sources and required disclosure or attribution.
Approve, return, or reject a named version and preserve the accountable review record.

Establish a trustworthy comparison set

Identify the generated draft version, its prompt and provenance record, the approved brief, and the source packet. Confirm that the source packet itself is current and permitted. Ask the reviewer to read the key evidence before editing, because fluent prose can anchor judgment and make invented details feel familiar. Separate passages copied or closely derived from sources, system-generated suggestions, and material written independently by people. If the provenance is missing for a consequential asset, treat that as a review limitation rather than assuming the likely workflow.

Record substantive judgment, not cosmetic activity

Human review means more than correcting spelling. Check whether the draft answers the intended question, represents sources accurately, includes necessary limitations, avoids unsupported claims, and uses examples and language appropriate to its audience. Log material additions, deletions, rewrites, and rejected sections with reasons. Verify citations by opening them; reproduce important figures; check product behavior; and involve subject, privacy, legal, accessibility, or safety specialists when triggered. A page can be heavily edited yet remain poorly supported, so record the properties tested and the evidence used.

Approve only the reviewed final version

After revisions, compare the complete final candidate with the decision log. Ensure disclosures, attributions, source links, and limitations survived later copy or design changes. Record unresolved issues and choose to return, qualify, remove, or escalate instead of relying on the label “human reviewed.” Name the reviewer, role, date, version, and approval scope. Keep the review record under appropriate access and retention controls. If the draft changes materially, reopen affected checks; approval does not attach permanently to the asset title.

See the fields in context

Fictional example: imaginary feature announcement

NoteHarbor and its feature are invented. This record does not describe a real AI draft or product.

Finding: The fictional AI draft says the feature “never loses formatting,” but the approved source only documents three tested file types.
Decision: Remove the guarantee and add the verified supported-format scope.
Human addition: A product editor adds a tested recovery step missing from the draft.
Disclosure: The imaginary publisher uses its approved statement for meaningful AI-assisted drafting and human verification.
Approval: The editor approves version 5 after reopening all changed claim checks.

Frequently asked questions

How much human editing is enough?

There is no reliable percentage. The reviewer must test the properties required by the content's consequence and be able to reject or rebuild the draft.

Can one reviewer cover every check?

Sometimes for low-risk work, but consequential product, scientific, legal, privacy, accessibility, or safety claims may require different expertise.

Should every small wording edit be logged?

Record changes that affect meaning, evidence, risk, voice, originality, or approval. Minor mechanical corrections can be summarized.

What if the prompt and original output are unavailable?

Record the limitation, increase source-based scrutiny, and decide whether policy permits approval without that provenance.