AI-Assisted Writing Governance

AI Output Provenance Log

Record how an AI-assisted draft was produced, including tools, source materials, prompt versions, generated outputs, human decisions, checks, and approvals.

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

Download Markdown

Accessible HTML preview

Blank template

The downloaded file contains the same fields in editable Markdown.

Project record

Project ID: [Enter stable identifier]
Asset title and version: [Enter title/version]
Business or editorial owner: [Enter name]
Intended audience and use: [Describe]
Risk or review tier: [Enter internal classification]
Record location and access group: [Enter controlled location]

AI interaction entry

Duplicate for each materially different request.

Interaction ID: [Example AI-01]
Date and time zone: [YYYY-MM-DD, time zone]
Operator: [Enter name or approved role]
Service and feature: [Enter displayed names]
Model/version shown: [Enter value or Not exposed]
Material settings: [Temperature, search, files, or Not exposed]
Prompt version or exact prompt: [Link or include approved prompt]
Source IDs supplied: [Link to context inventory]
Sensitive-data check: [Passed / Escalated / Not applicable]
Exact output location and checksum/version: [Enter reference]
System warnings or errors: [Record or None]

Contribution and decisions

Output portion considered: [Describe section or suggestion]
Decision: [Accepted / Edited / Rejected / Reference only]
Reason: [Evidence, quality, voice, rights, or other reason]
Material human changes: [Summarize authorship and editorial work]
Final asset location: [Enter versioned reference]
Decision owner and date: [Name, YYYY-MM-DD]
Other interactions that influenced this section: [List IDs]

Verification and approval

Subject expert review: [Name, scope, decision]
Editorial approval: [Name, date, final version]
Disclosure decision: [Text/link, not needed, or escalation]

Every retained material claim was checked against appropriate evidence.
Quotations, names, figures, dates, and links were opened and verified.
Confidential, personal, licensed, and restricted material was handled under policy.
The final work was reviewed for unsupported stereotypes, harmful instructions, and invented authority.
Similarity or rights concerns were resolved before publication.

Record handling

Retention period or trigger: [Enter policy-based value]
Deletion owner: [Enter name or team]
Access exceptions: [Record or None]
Legal or policy hold: [Yes / No / Unknown]
Final record status: [Open / Complete / Superseded / Deleted]

How to use this template

Assign a project ID and create one entry for each materially different AI interaction.
Record the known tool details, approved context, prompt, settings, and exact output location.
Link accepted material to the final asset and describe significant human changes or rejections.
Document factual, rights, privacy, safety, and subject-expert checks with responsible people.
Close the record with approval, disclosure, access, retention, and deletion decisions.

Preserve enough context to reconstruct the contribution

A useful provenance record allows a later reviewer to answer what the system received, what it returned, and what people decided. Assign a stable project and interaction ID. Record the service, model or feature name shown at the time, date, settings that materially affected output, prompt version, and source references. Save the exact output or a controlled record location rather than describing it as “AI draft.” If the service does not expose a precise model version, record only what is known and mark the field unavailable. Provenance should reduce ambiguity, not manufacture technical certainty that the interface never provided.

Distinguish generation from human editorial decisions

Do not credit a final page to a prompt simply because generated text appeared early in the workflow. Note which suggestions were accepted, substantially rewritten, rejected, or independently created. Record who selected sources, checked claims, resolved conflicts, changed the argument, and approved release. A compact decision summary can be more useful than a huge transcript: identify the portions that influenced the final asset and link to controlled evidence where policy requires it. When multiple interactions occur, create separate entries so reviewers can follow the sequence instead of blending outputs from different contexts.

Align retention and disclosure with actual risk

The log may itself contain proprietary drafts, personal data, security-sensitive context, or licensed source material. Apply access controls, retention periods, and deletion rules appropriate to those inputs. Do not expose an internal prompt transcript merely to prove transparency. Public disclosure usually needs a concise, reader-relevant explanation of meaningful assistance and human review, not the full audit record. Escalate regulated, contractual, employment, or high-consequence use to the appropriate policy owner. This template supports accountability but does not define the legal record a particular organization must keep.

See the fields in context

Fictional example: onboarding email outline

PaperKite is an invented service. No real model output, customer data, or company record is represented here.

Interaction: AI-02 used an approved fictional product brief to propose three onboarding-email outlines; the displayed model version was not available.
Decision: An editor rejected two outlines and retained the sequence “goal, first step, help,” rewriting every sentence.
Verification: A fictional product owner checked feature names and links; no performance claims were retained.
Disclosure: The editorial owner recorded that no public disclosure was needed under the imaginary policy because AI only supported internal outline exploration.
Retention: The prompt and output remain in a restricted project folder for 90 days, then the record owner reviews deletion.

Frequently asked questions

Must every small AI interaction receive a full record?

Use a level of detail proportionate to policy, consequence, and reuse. A low-risk brainstorming note may need less than public, regulated, or customer-specific content.

What if the tool does not show its model version?

Record the service, visible feature, date, and “not exposed.” Do not infer a hidden model name from marketing pages or memory.

Should prompts and outputs be published publicly?

Not automatically. They may contain protected information and often overwhelm readers. Use a concise disclosure when relevant while preserving controlled internal evidence.

Can provenance prove that final content is accurate?

No. It shows how material moved through a workflow. Accuracy still depends on source evaluation, verification, specialist review, and correct application of edits.