Labeling AI-generated and AI-assisted outputs

Start with the audience and possible confusion

Describe the artifact: internal draft, customer email, public article, image, audio, video, analysis, recommendation, or controlled record. Record who receives it, what they may reasonably infer, whether it resembles a real person or event, whether it informs the public, and what consequence follows from misunderstanding its origin.

Separate AI-generated content, AI-assisted editing, translation, retrieval, calculation, and ordinary automation. Also separate visible disclosure from technical provenance. A label can inform a person, while metadata or other mechanisms may support later tracing. Neither establishes that the content is accurate.

The NIST Generative AI Profile discusses content provenance, information integrity, transparency, and human-AI configuration as risk-management concerns. It is voluntary guidance and does not define a binding label for every context.

Sources: [1]

Build the contextual disclosure matrix

For each output class, record the intended use, audience, channel, AI function, degree of human review, editorial owner, possible confusion, affected group, jurisdiction, authoritative source, disclosure requirement or policy, wording, placement, persistence, accessibility, technical provenance, exception, approver, and re-review trigger.

Test the disclosure with the actual artifact. Confirm that it remains visible after cropping, reposting, exporting, translation, or assistive-technology use. Avoid wording that overstates what human review established. “Reviewed” should identify the scope of review rather than imply factual, legal, or safety assurance.

Distinguish disclosure from consent and correction. A person may need a route to challenge or correct an output even when its origin is clearly labeled. Record that route, its owner, and the evidence needed to investigate the concern.

The official EU Artificial Intelligence Act text contains transparency provisions for certain generated or manipulated content. Application depends on the facts, actor, content, timing, and jurisdiction; the matrix should route those questions to qualified review rather than convert a summary into legal advice.

Sources: [2]

Worked artifact: a synthetic public explainer video

Imagine a synthetic public-service team producing a short explainer video. A staff member writes the script from an approved policy, an AI tool creates narration and a presenter-like avatar, and an editor assembles the final version. The video could reasonably appear to feature a real spokesperson.

The matrix records synthetic voice and image generation, names the editorial owner, links the controlling policy, and requires a visible and spoken disclosure at the beginning plus persistent text in the description. Technical provenance metadata is retained where the publishing channel supports it.

During testing, the description survives reposting but the opening caption is cropped in a vertical format. The team adds an in-frame disclosure and checks screen-reader text. Review also finds that one policy sentence has been simplified beyond its source, so the script is corrected separately from the origin label.

Approval applies only to the recorded formats and channels. A later edit, translation, changed avatar, or new jurisdiction reopens the matrix. The example is synthetic and does not establish legal sufficiency or control effectiveness.

Verify labels in distribution

Sample released outputs across devices and downstream channels. Confirm wording, placement, persistence, accessibility, provenance, editorial ownership, and source accuracy. Record cases where platforms strip metadata or users remove visible context, then decide whether to restrict the format or add another disclosure layer.

Two Prune can observe how participants recognize provenance questions, verify content, and communicate uncertainty in controlled simulations. It does not issue legal conclusions, clear customer publications, certify disclosure controls, or make employment decisions.