EU AI Act Article 50 for AI Video Creators: What to Label and When
Credits and specs follow TWno1's live billing settings. Last checked: 2026-10-05
EU AI Act Article 50 has applied since August 2, 2026, so disclosure duties for AI video are already live.
You may have heard that the Digital Omnibus pushed AI labeling to December. Mostly wrong. The Omnibus gave a grace period to one paragraph, 50(2), and only for generators already on the market before August 2. The disclosure duty that falls on people who publish AI content, 50(4), was not delayed at all.
This is a summary of public texts and European Commission pages, checked on 2026-10-05. It is not legal advice. Whether your work falls in scope, and who is responsible, is a question for a lawyer who knows EU law.
The dates that matter
| Date | Event | What it means for AI video |
|---|---|---|
| 2026-05-07 | Omnibus trilogue deal | Political agreement, not yet law |
| 2026-05-08 | Draft Article 50 guidelines published (comments until 2026-06-03) | Useful signal; the July final version governs |
| 2026-06-10 | Final Code of Practice on marking and labeling AI-generated content | Voluntary; the Commission and AI Board call it an adequate tool for 50(2) and 50(4) |
| 2026-07 | Final transparency guidelines published | We have not read the final wording |
| 2026-07-24 | Omnibus (Regulation (EU) 2026/1744) published in the Official Journal | First amendment to the AI Act |
| 2026-07-27 | Omnibus enters into force | Adds the Art. 111(4) transition |
| 2026-08-02 | Article 50 applies (Art. 113) | 50(1), 50(3) and 50(4) start here |
| 2026-12-02 | 50(2) deadline for older generators; two new Art. 5 bans apply | Systems placed on the market before Aug. 2 must have machine-readable marking |
Read the last row carefully. New Art. 111(4) covers providers of systems that generate synthetic audio, images, video or text and were placed on the market before August 2, 2026. They have until December 2 to meet 50(2). A system launched on or after August 2 gets no window.
The same December date also brings two new prohibitions in Art. 5: AI practices that generate non-consensual intimate material, and child sexual abuse material. Those are outright bans, not labeling rules.
Provider or deployer?
Article 50 splits the work between two roles. Pick the wrong one and you will prepare for the wrong duty.
| Provider | Deployer | |
|---|---|---|
| Paragraph | 50(2) | 50(4) |
| Duty | Mark outputs in a machine-readable format so they can be detected as artificially generated or manipulated | Disclose that the content is artificially generated or manipulated |
| Usually | The company that builds and offers the generator | The person or company that uses a generator and publishes the result |
| Grace period | Older systems: until 2026-12-02 | None; applies from 2026-08-02 |
| Carve-outs in the text | Assistive editing, law enforcement | Evidently artistic, creative, satirical or fictional works get a lighter duty |
A typical AI video creator sits in the right column. You use someone else's generator and publish what it makes. That is our reading of the text, not a ruling. If you train a model and offer generation to others, read the left column too.
One practical consequence: you do not need to learn watermarking to satisfy 50(2). That is the provider's job. But a provider's marking does not discharge your 50(4) duty either. They are separate.
The fiction carve-out is lighter, not empty
For evidently artistic, creative, satirical or fictional works, 50(4) limits the duty to disclosing that such content exists, in a way that does not hamper the display or enjoyment of the work.
That changes how you disclose. It does not let you skip it.
Two points from the Commission's May 2026 draft guidelines are easy to miss:
- A fictional character is not automatically out of scope. The draft says a realistic synthetic depiction of a fictitious but natural-looking person can still meet the Art. 3(60) definition.
- Intent to deceive is not required.
The final guidelines came out in July 2026. We only read summaries of the draft, so treat these two points as signals, not settled law. Our suggestion for photorealistic AI drama: assume disclosure applies, then use the light form. One card at the start or end, plus one line in the video description.
Machine-readable marks do not survive every edit
The Code of Practice published on 2026-06-10 asks for at least two layers of machine-readable marking, such as metadata plus a watermark. The Commission listed about 190 signatories by late July 2026. Signing is voluntary.
Why two layers? The C2PA specification (v2.2, May 2025) admits that a Content Credentials manifest can be separated from a file when metadata is removed or corrupted. That is why "soft binding" exists, meaning fingerprints and invisible watermarks. The spec also says soft bindings are not statistically unique and are weaker than cryptographic binding.
So here is what happens in practice. You download a generated clip, cut it in an editor, and export. The original metadata may be gone. Nobody did anything wrong, but you cannot rely on the tool's marking to cover your disclosure. Put your 50(4) disclosure where viewers can see it.
One more correction to a common belief: C2PA is not a legal requirement. The EU Code, China's labeling Measures and platform auto-labels are separate regimes. None of them makes a C2PA manifest mandatory in itself.
Four things to do this week
- Add the disclosure card in post, not in the prompt. Text inside generated video, like signs and captions, often warps. An "AI-generated" line rendered by the model can come out garbled. Add it in your editor, where the letters stay exactly as you typed them.
- Put a fixed line in every description. For example: "Characters and footage generated with AI. All events are fictional." It fits the "does not hamper enjoyment" standard and leaves a record.
- Keep the raw, unedited exports. Your edited master may have lost the original metadata. The raw file is your evidence of where the footage came from.
- Ask your tool provider two questions in writing. Was your system placed on the EU market before or after 2026-08-02? Which machine-readable marks do your output files carry? File the answer.
Limits: what this article does not answer
- Penalties. We do not cover fines or which authority enforces Article 50. We did not have a primary source we had read.
- Territorial scope. Whether publishing on an EU platform, or targeting EU viewers, puts you in scope is a legal question.
- Personal, non-professional use. We did not re-check whether the deployer definition excludes it, so we draw no conclusion.
- Final guideline wording. We read only summaries of the May 2026 draft.
- High-risk and general-purpose model rules. Those sit outside Article 50.
- TWno1 will not tell you whether Article 50 covers your work or sign a compliance opinion; its own rules are on the content policy, terms, privacy and pricing pages, and questions go through the contact page.
Other legal explainers: policy, copyright.
FAQ
When does EU AI Act Article 50 start to apply?
EU AI Act Article 50 applies from August 2, 2026, under Article 113. The only exception is paragraph 50(2), machine-readable marking: providers of generators placed on the market before 2026-08-02 have until 2026-12-02 under the new Article 111(4) added by the Digital Omnibus. The deployer disclosure duty in 50(4) was not delayed.
Did the Digital Omnibus delay AI video labeling to December 2026?
Only in part. The Digital Omnibus, Regulation (EU) 2026/1744 in force since 2026-07-27, gave a transition period to Article 50(2) alone, for providers of generators placed on the market before 2026-08-02, ending 2026-12-02. Chatbot disclosure, emotion-recognition notices and deployer disclosure of realistic synthetic content all apply from 2026-08-02.
Does a fictional AI short drama still need an Article 50 disclosure?
Yes, in a lighter form. Article 50(4) limits the duty for evidently artistic, creative, satirical or fictional works to disclosing that such content exists, without hampering display or enjoyment. The Commission's May 2026 draft guidelines also said a realistic fictitious person can still be covered and that intent to deceive is not required.
Is the EU Code of Practice on AI-generated content mandatory?
No. The final Code of Practice on marking and labeling AI-generated content was published on 2026-06-10 and is voluntary, with about 190 signatories by late July 2026. The Commission and AI Board describe it as an adequate tool for Article 50(2) and 50(4), and it asks for at least two layers of machine-readable marking, such as metadata plus a watermark.
Do AI content marks survive video editing and re-export?
Not reliably. The C2PA specification v2.2 acknowledges that a Content Credentials manifest can be separated from a file when metadata is removed or corrupted, which is why weaker soft bindings such as watermarks exist. Under EU AI Act Article 50(4), creators should place their disclosure where viewers can see it, such as an on-screen card or the video description.
References
- EU AI Act, Article 50 (artificialintelligenceact.eu) (checked 2026-10-05)
- EU AI Act, Article 111, including new paragraph 4 (checked 2026-10-05)
- EU AI Act, Article 113: entry into force and application (checked 2026-10-05)
- European Commission: regulatory framework for AI (checked 2026-10-05)
- European Commission: Code of Practice on AI-generated content (checked 2026-10-05)
- Jones Day: Commission publishes final Code of Practice (checked 2026-10-05)
- Cloud Security Alliance: Article 50 watermarking deadline note (checked 2026-10-05)
- C2PA technical specification v2.2 (checked 2026-10-05)