Practical guide
Source-attribution policy for AI-enabled work
A source-attribution policy should state which claims and assets require evidence, what records must be retained, who reviews them, when release must stop, and who may approve a bounded exception. It should not treat a hyperlink, provenance credential, or AI disclosure as proof that a claim is true. The policy artifact below governs source-dependent work from intake through release and reuse.
A source-attribution policy should state which claims and assets require evidence, what records must be retained, who reviews them, when release must stop, and who may approve a bounded exception. It should not treat a hyperlink, provenance credential, or AI disclosure as proof that a claim is true. The policy artifact below governs source-dependent work from intake through release and reuse while leaving subject-matter approval with accountable owners.
Policy objective and scope
Use SCOPE: Scope material uses, Classify evidence, Own decisions, Perform release controls, Escalate exceptions and change. The policy sets mandatory rules. The attribution packet is the working record for a specific output, while operating-evidence testing later determines whether the control actually ran.
W3C PROV-O provides a vocabulary for representing entities, activities, agents, and derivations. The NIST Generative AI Profile identifies information-integrity risks and discusses provenance and source-verification practices. The C2PA explainer states that Content Credentials provide tamper-evident provenance but do not determine whether content is true. These primary sources support distinctions between origin, transformation, and claim support; they do not make this sample policy legally sufficient for any jurisdiction. W3C PROV-O NIST Generative AI Profile C2PA explainer
### S: Scope material uses
This policy applies to externally released or internally consequential AI-enabled work containing factual claims, quantitative statements, calculations, recommendations, quotations, licensed material, images, audio, video, datasets, or summaries derived from a source. It also applies when approved content is reused for a new audience, period, decision, language, or channel.
The accountable output owner defines materiality before drafting. At minimum, treat as material any source-dependent statement that could change a decision, create legal or operational exposure, misrepresent a person or organisation, or be difficult for a reader to verify. Teams may apply the same control more broadly.
This policy does not declare a source authoritative merely because it is official, primary, recent, or widely cited. Source suitability depends on the claim, population, period, method, and intended use.
### C: Classify evidence and generated contribution
Every material element must be classified as one of the following:
The working record must retain the source identity, publisher or custodian, title, version or date, stable location or repository ID, relevant passage or data cells, access date, transformation, final wording, reviewer, and disposition. Calculations must also retain units, periods, filters, formulas, and rounding. Generated text is not a factual source; record the human-selected evidence and material revisions.
Provenance and support must remain separate. A valid Content Credential may show origin and editing history without showing that a depicted event occurred. A live URL may identify a page without supporting the sentence beside it.
### O: Own decisions and segregation
The output owner defines purpose, materiality, and permitted use. The author assembles the attribution record. The subject-matter reviewer determines whether evidence supports the final wording. The data or asset owner confirms calculation lineage, licence, or reuse conditions where relevant. The release approver accepts residual limitations. One person may hold several roles only when the consequence and local controls permit it.
Technology teams may operate citation, provenance, or repository systems, but those systems do not own factual approval. AI tools may suggest sources, extract passages, compare versions, or flag missing records. They may not approve support, waive review, or silently substitute a source.
The policy owner maintains the procedure and training. They do not certify every output. Functional owners remain responsible for their domain judgments.
### P: Perform controls before release
Before release, the author and reviewer must:
Release must stop when a material claim is unsupported, the source population or period conflicts with the wording, a consequential calculation cannot be reproduced, a generated asset could be mistaken for a real event, required permission is absent, or the reviewer cannot inspect the evidence. A deadline does not change an unsupported claim into a supported one.
Use sampling only under an approved risk-based procedure. Sample size and selection should reflect consequence, novelty, quantitative content, generated contribution, source volatility, and prior exceptions. Sampling does not excuse an author from creating a complete attribution record for the covered material.
### E: Escalate exceptions, reuse, and change
An exception request must identify the affected element, reason, consequence, interim wording or control, audience, duration, owner, approver, and closure date. Only the designated authority may accept it. Some conditions are non-waivable under this policy: fabricated sources, knowingly unsupported consequential claims, missing rights for restricted material, or a synthetic asset presented as a real event.
Approved exceptions must remain visible to the release approver and downstream user. They expire at the stated date and require correction or removal. Emergency procedures may narrow the review path only when an approved process names the authority and deferred checks.
Reuse triggers a fresh review when the source, claim, audience, decision, period, language, format, or generated contribution changes. Automated source refreshes must not overwrite the retained evidence for an earlier release. The policy owner reviews this policy after material incidents, recurring exceptions, relevant system changes, or at the defined review date.
- direct support from an identified source;
- partial or contextual support requiring qualified wording;
- calculation with identified inputs and reproducible transformation;
- inference or professional judgment;
- assumption or scenario;
- generated or synthetic material;
- unsupported, inaccessible, or superseded and therefore held or removed.
- reconcile all material claims and assets to the attribution record;
- resolve the retained source and confirm its version and date;
- inspect the relevant passage, cells, or media rather than only the title or URL;
- reproduce sampled transformations and calculations;
- distinguish direct support, inference, assumption, and generated material in final wording;
- verify licence, confidentiality, and permitted reuse where applicable;
- confirm that the reviewed version matches the release version;
- record unresolved limitations, exceptions, and refresh triggers.
Sources: [1] [2] [3]
Minimum attribution record
| Field | Required entry | | --- | --- | | element ID | stable claim, calculation, or asset reference | | classification | direct, partial, calculation, inference, assumption, generated, or held | | source | identity, publisher, version/date, location | | support | relevant passage, cells, or asset relationship | | transformation | summary, extraction, formula, edit, or generation | | final wording | approved claim or visible label | | roles | author, reviewer, output owner, release approver | | disposition | approve, qualify, replace, remove, or exception | | lifecycle | review date, refresh trigger, exception expiry |
Worked policy application: synthetic quarterly briefing
A fictional strategy team prepares a quarterly briefing using an official statistical release, a consultancy summary, an internal calculation, and an AI-generated illustration. The AI assistant drafts a sentence claiming a 16% market increase and links the consultancy page.
The author classifies the official release as a candidate direct source, the consultancy page as contextual, the chart as a calculation, and the illustration as synthetic. Review finds that 16% compares one region with the prior quarter, while the draft describes the whole market year over year. The consultancy page repeats the number but does not repair the population mismatch.
The output owner asks for a deadline exception. The release approver denies it because the material quantitative claim is unsupported and the population conflict is consequential. The author replaces the sentence with a narrower statement supported by the official table, records the region and period, preserves the calculation inputs, and labels the illustration as synthetic. The briefing is released with a refresh trigger for the next official publication.
Later, a communications team proposes reusing the chart in a global campaign. The audience and claim have changed, so the prior approval does not travel automatically. The new owner must establish a suitable population or decline the reuse. This example shows the policy's decision path; it does not establish that the fictional control operated across other outputs.
Two Prune boundary
Two Prune can assess whether a participant classifies sources, preserves evidence, tests claim support, handles exceptions, and applies a release boundary within a realistic simulation. It does not certify provenance systems, audit organisational controls, determine legal compliance, establish factual truth across a domain, predict workplace outcomes, or make employment decisions.
Sources
W3C: PROV-O
- NIST AI 600-1: Generative AI Profile
- C2PA: Explainer
Sources: [1] [2] [3]
Sources
- 1.W3C PROV-O · World Wide Web Consortium
- 2.NIST Generative AI Profile · National Institute of Standards and Technology
- 3.C2PA explainer · spec.c2pa.org