← Back

Guide · Published November 14, 2026 · Reviewed November 14, 2026 · Ingenious AI team · Not legal advice

← All guides

AI Content Marking and Deepfake Disclosure: Article 50's Content Duties Explained

The short answer: Two of the EU AI Act's Article 50 duties concern content rather than decisions. Providers must mark AI-generated or manipulated content in a machine-readable format (Article 50(2)) — and systems placed on the market before August 2, 2026 have until December 2, 2026 for this duty. Deployers must disclose deepfake content, and AI-generated text published to inform the public on matters of public interest (Article 50(4)). Both have applied since August 2, 2026. For a business whose outputs include decision documents and public content, these duties live in the publishing workflow — and the marking one carries a test most teams forget: the mark must survive your export path.

Check your duties and dates — free Build your first notice — free

What is content marking? It is the machine-readable signal attached to AI-generated or manipulated content so that downstream systems — not just human readers — can detect its origin. Disclosure, its sibling duty, is the human-facing statement: that a deepfake is a deepfake, or that public-interest text was AI-generated. The kit's checker groups both under Article 50, whose full set is introduced in EU AI Act Article 50 Transparency Duties.

The marking duty: for machines to read, not just people

A label a person can read but a system cannot does not exhaust Article 50(2): the marking is specified as machine-readable. In practice that means metadata or an equivalent technical signal applied at generation, carried with the content. Inventory where your organisation generates content with AI — documents, summaries, images, letters, web copy — and for each output type, identify where the mark is applied and in what form.

Decision documents deserve a specific decision of their own. The kit's notices and explanations are template-generated documents with a stated generator (their footer says they were generated with the kit) — a human-facing provenance statement by design. Whether a given output class also needs a machine-readable mark is a classification to make deliberately per output type, with the reasoning recorded, rather than a default in either direction.

The grace period: December 2, 2026, for earlier systems only

The marking duty has the only internal deadline split in Article 50 as the checker states it: the duty applies from August 2, 2026, but systems placed on the market before that date have until December 2, 2026 for this obligation. Two planning errors follow predictably: treating the grace period as covering all of Article 50 (it covers the marking duty), and treating it as covering systems bought after August 2026 (it is keyed to when the system was placed on the market).

Establish, per generating system, its placed-on-market position from procurement records, and diary the December date for the ones it covers. The grace period is short; a marking retrofit discovered in November is a project, while the same retrofit planned in September is a task.

Worked example: Alpenrose Learning Group

Example — fictional business, real duty structure. Alpenrose is the fictional education provider from the Article 50 guide, returning to complete its content duties. Its marking and disclosure pass found:

Deepfake disclosure: the deployer's duty

Article 50(4)'s deepfake disclosure sits with the deployer — the business publishing or using the content — not with the tool that generated it. Any workflow that produces or distributes synthetic media of real people or events needs the disclosure step built in before publication, owned by the publishing role, not appended by legal review afterwards (which is where schedules go to die).

For most businesses in the ADMT space, exposure is modest but real: marketing content, testimonial-style media, synthetic imagery on product pages. The control is a publishing checklist item — is any of this content a deepfake or synthetic depiction? is the disclosure attached? — asked at the point of publication, with the answer logged for public-facing pieces.

Public-interest AI text: the second half of 50(4)

The same paragraph covers AI-generated text published to inform the public on matters of public interest: deployers must disclose it as AI-generated. News-style content, public information pages, and advisory content generated by AI fall in this frame when published to inform the public on public-interest matters. Internal drafts and purely commercial copy are different cases; the line is the informing-the-public purpose.

A content operation should classify its AI-assisted publishing into buckets — public-interest informing text (disclose), other public content (mark per 50(2) analysis), internal content (no Article 50 duty) — and record the classification rules once, so per-piece decisions are applications of a policy rather than fresh judgments under deadline.

The export-survival test, and the record

Marks fail silently at format boundaries: applied in the generating tool, stripped by the PDF export, the CMS, the email system, or the image pipeline. Test the full path per output type — generate, export through every channel the content actually travels, and verify the mark at the far end. A marking duty evidenced only at generation is a duty half-discharged, and the failure is invisible until someone checks the published artefact.

Keep the content-side record with the same discipline as the decision record: systems inventoried with their marking method and placed-on-market dates, classification rules for disclosure, publication logs for disclosed pieces, and the export-test results with dates. It is a smaller record than the ADMT chain, but it has the same reader a year from now. The wider Article 50 set, including the interaction and emotion-recognition duties, is at Article 50 Transparency Duties, and the emotion-recognition duty has its own guide next in this series.

  1. Inventory every content-generating system with its marking method and placed-on-market date; diary December 2, 2026 for pre-August systems
  2. Apply machine-readable marking at generation — a human-only label does not exhaust Article 50(2)
  3. Test mark survival through every real export channel (PDF, CMS, email, image pipeline) and file the dated results
  4. Add the deepfake disclosure question to the publishing checklist, owned by the publisher, answered before release
  5. Classify AI public-interest text under a written policy and disclose it as AI-generated when published

Sources, as carried in the kit's own data: the duty summaries in app/rules.py and the checker data in app/checker_data.py, verified against primary or authoritative sources on October 28, 2026. Run the free duties checker for your jurisdiction, decision type, and the tool's role, and compare the kit's plans for the record, builder, and review queue. More from Ingenious AI is on the hub.

Build your first notice — free

Related guides

See all ADMT guides for the full set.

Frequently asked questions

What does Article 50(2) require?

Providers must mark AI-generated or manipulated content in a machine-readable format. Systems placed on the market before August 2, 2026 have until December 2, 2026 for this duty.

What does Article 50(4) require?

Deployers must disclose deepfake content, and AI-generated text published to inform the public on matters of public interest.

Is a visible 'AI-generated' label enough for marking?

The duty specifies machine-readable marking — a technical signal systems can detect. A human-facing label may serve disclosure, but the marking duty is about machine readability, and the mark must survive your export channels.

Who owes the deepfake disclosure — the tool or the publisher?

The deployer: the business publishing or using the content. Like the ADMT notice duties, the obligation sits with the business whose name the content carries.

When did these duties start?

August 2, 2026, with the single grace period noted: the marking duty allows systems placed on the market before that date until December 2, 2026.

Build your first notice — free