Practical guide
Output labeling policy for AI-enabled work
An output labeling policy should tell people which AI-assisted or synthetic outputs require a disclosure, what the disclosure must communicate, where it must persist, who owns it, and how exceptions are tested. A visible badge alone is not a policy. The operating design must connect audience needs, content provenance, distribution channels, accessibility, evidence retention, and incident response.
An output labeling policy should tell people which AI-assisted or synthetic outputs require a disclosure, what the disclosure must communicate, where it must persist, who owns it, and how exceptions are tested. A visible badge alone is not a policy. The operating design must connect audience needs, content provenance, distribution channels, accessibility, evidence retention, and incident response.
Source and policy boundary
NIST AI 100-4 reviews provenance tracking, watermarking, detection, embedded labels, and user-interface disclosures, while emphasizing context, modality, implementation, and user-experience limits. The C2PA Content Credentials specification defines a technical architecture for cryptographically bound provenance information and continues to evolve; version 2.4 was published in April 2026. These sources support policy design choices. They do not create a universal disclosure rule, establish legal compliance, or show that every user will interpret a label correctly. NIST: Reducing Risks Posed by Synthetic Content C2PA: Content Credentials specification 2.4
Sources: [1] [2]
Design the policy with TRACE
### T — Trigger and scope
Define the policy objective before choosing label text. Objectives may include helping a recipient understand material AI involvement, preserving origin and edit history, distinguishing an official communication, or routing a high-consequence output for review.
Create a decision table based on content type, intended audience, channel, consequence, degree of AI involvement, and contractual or jurisdiction-specific requirements. Avoid rules such as “label everything made with AI” when minor editing and fully synthetic media create different reader questions.
Include text, images, audio, video, data products, and composite documents where relevant. State exclusions narrowly and record why they do not undermine the objective.
### R — Representation and reader meaning
Specify what a label asserts. “AI generated,” “AI assisted,” “synthetic media,” “content credentials available,” and “reviewed by a person” are not interchangeable. Each communicates a different fact.
Use plain language and place the disclosure where the audience will encounter it before relying on the content. Test visual prominence, contrast, screen-reader order, audio treatment, localization, and comprehension. Do not imply that provenance proves truth or that human review removes every error.
### A — Attachment and persistence
Choose one or more mechanisms appropriate to the channel:
Define how the disclosure survives export, resizing, cropping, transcoding, screenshots, copying, syndication, and third-party platforms. Where a mechanism cannot persist, specify a fallback and record the loss.
### C — Control ownership and operation
Name the policy owner, workflow operator, content approver, technical control owner, accessibility reviewer, records owner, and incident route. Define which role may approve an exception and which role independently tests the control.
Integrate labeling into creation and publishing workflows. A control that depends on memory at the final click will fail differently from one generated from structured workflow data. Keep the source, tool, transformation, review status, label decision, issue time, and issued asset linked.
### E — Evidence, exceptions, and evolution
Test an independently constructed population of outputs rather than only the examples supplied by the operator. Reconcile all eligible outputs to labeled, exempted, failed, and unknown categories. Sample across channels and transformations.
Measure both technical persistence and user interpretation. Log exceptions with scope, reason, approver, compensating action, expiry, and retest. Review the policy when tools, formats, distribution partners, user research, incidents, or external requirements change.
- an overt in-content mark;
- a disclosure beside the content in the user interface;
- embedded provenance metadata or Content Credentials;
- a linked record containing origin, transformations, review, and publisher information.
Policy template
### 1. Purpose
This policy enables intended recipients to understand specified material AI involvement and, where applicable, inspect provenance information. It does not assert that labeled content is accurate or that unlabeled content had no AI involvement.
### 2. Applicability
The policy applies to named business units, content types, channels, and audiences. The applicability register records the rule, owner, effective date, and rationale for each class.
### 3. Required disclosures
For each applicable class, define exact approved wording, location, timing, accessibility treatment, metadata or credential requirement, and persistence fallback. Maintain examples for correct, incorrect, and ambiguous cases.
### 4. Workflow
The creator records AI involvement and source material. The workflow evaluates applicability. The approver checks content and disclosure. Publishing attaches the approved label and provenance record. Monitoring reconciles issued assets and detects loss or alteration.
### 5. Exceptions
Only the named authority may approve a time-bounded exception. The record must include affected assets, reason, residual exposure, compensating action, expiry, and notification requirement. Emergency use receives retrospective review within the stated period.
### 6. Testing and monitoring
Test eligible-population completeness, wording, placement, accessibility, persistence through transformations, provenance retrieval, audience comprehension, exception closure, and incident response. Report denominators and missing records.
### 7. Records and review
Retain the applicability decision, source lineage, transformation history, approval, issued version, exception, test result, incident, and corrective action according to the organization’s records schedule. Review on a fixed date and after material change.
Worked policy application: synthetic campaign-image workflow
A communications team creates social images from approved photography. Generative tools may extend backgrounds, replace objects, or create an entire image. The policy distinguishes minor color correction, material AI editing, and fully synthetic composition.
The trigger table requires a visible “AI-assisted image” disclosure for material edits and “synthetic image” for a fully generated composition. Both classes require a linked provenance record. Minor color correction is recorded internally but does not receive the same reader-facing disclosure under this organization’s stated objective.
The publishing workflow generates the label from the asset record, places it in the caption and asset page, adds alt text that communicates the disclosure, and attaches a Content Credential where the channel supports it. For a channel that strips metadata, the caption and linked record are the fallback. The policy does not describe the credential as proof that the depicted event occurred.
The first operating test reconciles 240 exported assets. Forty-eight are eligible for a reader-facing label. A stratified sample of 50 includes all 48 eligible assets plus two exempt controls. Review finds three template exports without the caption, two crops that remove an in-image mark, four files with stripped metadata, one low-contrast disclosure, and two cases where alt text omits the disclosure.
A comprehension test with intended internal reviewers also finds that some interpret “AI assisted” as meaning a person checked every factual detail. The wording is revised to “Material elements were created or changed with AI; see production record,” and the policy explicitly separates disclosure from content approval.
The control disposition is operating with material exceptions, not fully effective. The operator repairs the template, adds a crop-safe caption check, monitors metadata loss, corrects accessibility defects, and retests the affected routes. Exceptions expire in 14 days. Any externally issued eligible asset without a disclosure triggers incident review and distribution correction.
The evidence artifact contains the approved policy, applicability matrix, role map, wording library, accessibility specification, asset population, sample plan, transformation tests, comprehension notes, exception log, corrective actions, and review record.
Acceptance and hold conditions
Accept the policy when the objective, scope, assertions, mechanisms, owners, exceptions, tests, records, and review triggers are explicit and have been exercised on representative output routes.
Hold publication when the disclosure asserts more than the evidence supports, the recipient cannot encounter or understand it, the label disappears through a material channel, an exception has no authority or expiry, the eligible population cannot be reconstructed, or the policy is presented as a substitute for jurisdiction-specific review.
Two Prune provides realistic AI-enabled work assessments and evidence about observed participant performance. It can sample whether a participant applies disclosure rules, checks persistence, verifies sources, and handles exceptions. It does not determine legal duties, certify policy operation, or make employment decisions.
Sources
NIST: Reducing Risks Posed by Synthetic Content
- C2PA: Content Credentials specification 2.4
Sources: [1] [2]
Sources
- 1.NIST: Reducing Risks Posed by Synthetic Content · National Institute of Standards and Technology
- 2.C2PA: Content Credentials specification 2.4 · spec.c2pa.org