AI Labels Should Explain the Human Contribution

An “AI-generated” label can tell readers that a machine played a part. It cannot tell them who supplied the original idea, chose the evidence, rejected weak outputs, corrected errors or accepted responsibility for the finished work. When a single badge replaces that fuller story, transparency becomes technically accurate but editorially incomplete.

The better question is not simply, “Was AI used?” It is, “What did each participant contribute, and who stands behind the result?” That shift matters to writers, artists, photographers, researchers, publishers and audiences trying to judge a work fairly.

The binary label problem

Creative work rarely moves in a straight line. A person may begin with field notes, an interview, a sketch, a melody or a business problem. An AI system may help organise, draft, translate or visualise part of that material. The person may then compare versions, remove inventions, restore context and make the final decision.

A binary label compresses that sequence into one fact: AI appeared somewhere in the process. That can be useful, but it can also make very different workflows look identical. A one-line prompt and a carefully directed, heavily edited project may receive the same description even though the human contribution is substantially different.

The opposite mistake is equally serious. A creator should not use the existence of a human idea to minimise extensive machine generation. Fair attribution must preserve both truths: human judgement deserves recognition, and significant automation deserves disclosure.

Disclosure, provenance and responsibility are different

Three ideas are often bundled together, but they answer different questions.

• Disclosure tells the audience that AI generated or materially altered content.
• Provenance records information about the origin and history of a digital asset.
• Responsibility identifies the person or organisation accountable for publishing the result.

The distinction already appears in policy and technical standards. Article 50 of the EU AI Act sets transparency obligations for certain synthetic and manipulated content, while also recognising human review, editorial control and editorial responsibility in its treatment of some public-interest text.

Meanwhile, the C2PA guiding principles describe provenance as information that producers and custodians can assert about creation and subsequent actions. Importantly, C2PA says a provenance system should verify associated assertions without deciding whether those facts are “good” or “bad”. Technology can preserve a history; people still have to interpret its meaning.

Use a contribution record, not a vague badge

A useful label should be short enough to read and specific enough to inform. For substantial work, pair a visible summary with an expandable contribution record covering five points.

1. Origin

State where the work began. Was the source a human photograph, original reporting, a recorded interview, a hand-drawn concept, supplied data or a prompt created for the project? Origin helps an audience distinguish transformation from invention.

2. Human contribution

Name the meaningful work: research, lived observation, composition, art direction, source selection, interviewing, editing, fact-checking or final approval. Avoid inflated verbs. “Directed and edited” is more credible than a sweeping claim of sole creation when a system produced most of the first draft.

3. Machine contribution

Describe the function rather than merely naming the tool. Useful phrases include “generated the initial visual”, “suggested alternative structures”, “translated a human-written draft” or “removed background noise”. The audience learns more from the action than from a product name.

4. Review and responsibility

Say who checked the result and who is accountable for publication. Human review should not be treated as a ceremonial last click. It should include checks appropriate to the risk: factual verification for reporting, consent for identifiable people, rights checks for commercial assets and safety review for consequential advice.

5. Supporting provenance

Where the workflow supports it, attach a durable history. Content Credentials can carry digitally signed information about an asset’s origin and editing history. This evidence can strengthen a disclosure, although its absence does not by itself prove deception and its presence does not guarantee that a creative claim is wise or ethical.

Why percentages create false precision

It is tempting to call a work “30% AI” or “80% human”. Unless a platform defines and audits the calculation, those numbers are usually impressions disguised as measurements. Time spent is not the same as creative importance. Word count does not capture who found the story. Pixel count does not capture who chose the subject, obtained consent or decided what the image should communicate.

Contribution categories are more honest than invented percentages. They let audiences see the decisions that shaped the work without pretending that authorship can be reduced to a universal formula.

Three examples of fairer attribution

A reported article: “Reporting, interviews and source verification by Amina Cole. An AI assistant suggested an outline and copy edits. The author checked every quotation and accepts editorial responsibility.”

A conceptual campaign image: “Concept and art direction by the studio; initial image generated with AI; composition, colour and brand details edited by the design team. The pictured scene is illustrative, not a photograph of an event.”

An archival photograph: “Photograph by Luis Mensah. AI-assisted restoration removed scratches and reconstructed a damaged corner. The historical subject and principal composition were not generated.”

Each statement tells the reader something actionable. It also protects the person whose original material could otherwise disappear behind a generic machine label.

A practical disclosure test for publishers

1. Materiality: Would a reasonable reader judge the work differently if they knew how AI was used?
2. Specificity: Does the note explain the function AI performed?
3. Credit: Does it preserve the names and roles of people who supplied original material or judgement?
4. Context: Could someone mistake an illustration, reconstruction or synthetic voice for documentary evidence?
5. Accountability: Is a person or organisation clearly responsible for the final publication?
6. Durability: Can the disclosure travel with the asset through downloads, edits or reposting?

The last question is where technical provenance can help. The Content Credentials adoption guidance explains how signed updates can add information about changes to a content history. But publishers should still keep plain-language disclosures visible, because readers should not need specialist software to understand the central facts.

Protect consent as well as credit

Attribution is not a substitute for permission. A person can be accurately credited and still have their work, face or voice used beyond the scope they agreed to. A responsible workflow records the source, the permitted use, any restrictions and the person authorised to approve changes.

This is especially important when AI makes transformation cheap. The ability to imitate a style, extend a photograph or synthesise a voice does not answer whether doing so is fair. Consent, cultural context and contractual rights remain human questions.

The label should start a conversation, not end one

Good transparency does not force audiences to choose between celebrating technology and defending human creativity. It gives them enough context to recognise both. A concise label can signal machine involvement; a contribution record can show the original sources, the human decisions, the automated steps and the accountable publisher.

That is a more durable standard than a badge alone. It credits the people whose imagination and judgement shaped the work, discloses the system’s role without euphemism, and gives readers a clearer basis for trust.