Mandatory AI Content Labeling (AI Act Article 50): How to Show a Work Is Human-Made and Prior?

An illustrator delivers a poster series to a client. Three weeks later, a competing agency objects: the style "looks like Midjourney," and the contract required human-made work. The illustrator has her working files, but nothing dates their successive versions in a verifiable way. Since 2 August 2026, that kind of dispute plays out in a new context: AI-generated content must now identify itself as such across the European Union.

Article 50 of Regulation (EU) 2024/1689 (the AI Act) mandates the marking of synthetic content and the disclosure of deepfakes. The law forces AI to be identified. Symmetrically, and this is the blind spot in the debate, human creators now have a concrete interest in dating their originals before releasing them. This article covers the new obligations, then what a creator can put in place today to document the priority of their human-made versions.

What changed with Article 50 of the AI Act on 2 August 2026?

Since 2 August 2026, Article 50 of Regulation (EU) 2024/1689 applies throughout the European Union. Providers of generative AI systems must mark synthetic outputs in a machine-readable format; deployers must disclose deepfakes. Fines can reach 15 million euros or 3% of total worldwide annual turnover.

The full text is available on EUR-Lex (Regulation 2024/1689). Article 50 organizes transparency around four situations:

• Interacting with an AI (Article 50(1)): people must be informed that they are interacting with an AI system, unless this is obvious from the context.
• Synthetic content (Article 50(2)): providers of systems generating synthetic audio, image, video or text must ensure the outputs are marked in a machine-readable format and detectable as artificially generated or manipulated. Technical solutions must be effective, interoperable, robust and reliable as far as technically feasible.
• Emotion recognition and biometric categorisation (Article 50(3)): exposed persons must be informed.
• Deepfakes and public-interest text (Article 50(4)): deployers generating or manipulating deepfake content must disclose that it has been artificially generated or manipulated. AI-generated text published to inform the public on matters of public interest must also be disclosed, unless it has undergone human review and editorial control.

Article 50(7) additionally tasks the AI Office with facilitating codes of practice for the detection and labeling of artificially generated content.

Who must do what: providers vs deployers

The regulation distinguishes the provider (who develops or places the AI system on the market) from the deployer (who uses the system under its own authority). Obligations and timelines differ:

The grace period until 2 December 2026 for machine-readable marking by systems placed on the market before 2 August 2026 is confirmed by the European Commission's FAQ on Article 50 (updated 24 July 2026). The same FAQ restates the penalty ceiling under Article 99 of the regulation: up to 15 million euros or 3% of total worldwide annual turnover for the preceding financial year, with proportionality adjustments possible for SMEs.

Does AI labeling solve the reverse question, proving content is human-made?

No. Article 50 identifies what comes out of an AI; it does not certify what does not. Content without an AI mark is not legally presumed human: marks can be missing, stripped, or lost through file conversions. For a creator, the evidentiary question of human origin and priority remains entirely open.

That is the system's blind spot. Three concrete situations show it:

• The commissioned work. More and more clients insert "human-made" or "no generative AI" clauses. If challenged, the contractor must be able to document the creation process, not merely assert it.
• The chronology dispute. An AI model produces a visual close to your work. Who was first? Without verifiable dates on your files, the debate stalls. This scenario mirrors the circumstantial-evidence logic we analysed in the French Senate's presumption of AI use of protected works.
• The platform that filters. Marketplaces and stock libraries increasingly separate AI content from human content. A creator wrongly accused of publishing undisclosed AI needs material, dated, verifiable elements.

How can you document the priority of a human-made creation?

As a general matter, you cannot directly prove "a human made this." What you can do is document a creation process whose chronology and materiality make human origin credible. It is a body of circumstantial evidence, and the verifiable date on each piece is what binds it together.

The classic pieces of that body of evidence:

• Working files: RAW and EXIF for photos, layered PSD or vector source files for design, draft manuscripts with version history for texts, multitrack sessions for music, Git history for code.
• Dated intermediate versions: human creation leaves successive states over days or weeks. An image generated from a single prompt has no such temporal depth.
• A documented release chain: emails, submissions to a publisher, archived uploads.
• Electronic timestamping: a cryptographic fingerprint (SHA-256) of each version, anchored to an independently verifiable date.

1. 1
   Timestamp working files at the start of the project

   Sketches, RAW files, first draft: compute and timestamp the SHA-256 fingerprint of each initial file. That is the earliest time marker of your body of evidence.
2. 2
   Timestamp intermediate versions at regular intervals

   A series of fingerprints spread over time documents a progression. That chronological depth is very hard to fabricate after the fact — which is precisely what makes it credible.
3. 3
   Timestamp the final version before any release

   Before sending to the client, publishing, or uploading to a platform: one last fingerprint freezes the exact state of the deliverable and establishes its priority over the release.

What a timestamp proves, and what it does not

An electronic timestamp proves that a file existed, in a strictly identical form, at a given date. It establishes existence and integrity as of that date. It proves neither the author's identity, nor ownership of rights, nor the creation process. Its evidentiary weight remains for the court to assess.

Each notion needs to be kept distinct:

• Existence: yes. The file matching that fingerprint existed at the anchored date.
• Integrity: yes. Any later modification, even of a single byte, produces a different fingerprint.
• Priority: yes, relatively. A timestamp sets a "no later than" date; compared against another dated event, it establishes chronological order.
• Identity: no. A hash does not say who created or who submitted the file.
• Ownership: no. Copyright arises from the original creation, not from timestamping. Timestamping creates no rights; it only helps prove chronology.
• Human origin: no, not directly. It is the chain of timestamps across successive working files that makes a human process credible, not one isolated timestamp.

On admissibility: Article 41(1) of the eIDAS Regulation (910/2014) prohibits denying an electronic time stamp legal effect and admissibility as evidence in legal proceedings solely on the grounds that it does not meet the requirements of a qualified electronic time stamp. Non-qualified does not mean invalid: a qualified time stamp enjoys a presumption of accuracy, while a non-qualified one must demonstrate its reliability, which, for a Bitcoin anchor, independent verification of the cryptographic chain makes possible.

Concrete use cases since 2 August 2026

The design studio under a "no generative AI" clause. The studio timestamps source files (layers, vectors) at three milestones: approved brief, intermediate mock-up, delivery. If challenged, it produces a verifiable chronology of native working files: serious circumstantial evidence of a human creation process, to be weighed with the other exhibits.

The photographer facing an "AI image" accusation. She timestamps her RAW files as soon as the memory card is offloaded. A timestamped RAW with its EXIF metadata, predating any publication, carries more weight than an isolated final JPEG.

The publisher using AI-assisted text. Article 50(4) exempts AI text that has undergone human review and editorial control from disclosure. Timestamping successive versions (raw output, edited version, published version) documents precisely that human editorial intervention, useful if a regulator or a reader asks.

The author before submitting to a publisher. A manuscript timestamped chapter by chapter during writing, then in its final version before the first submission: if a similar text later appears, AI-generated or not, the chronology is on record.

Where does LegalStamp fit in?

LegalStamp is a non-qualified electronic timestamping service under eIDAS, based on anchoring into the Bitcoin blockchain via the open OpenTimestamps protocol. Concretely:

• The SHA-256 fingerprint of your file is computed in your browser; the file never leaves your device; only the fingerprint is transmitted.
• The fingerprint is anchored into Bitcoin via OpenTimestamps. Anchoring is deferred: the attestation becomes available after a block confirmation, never instantly.
• You receive an .ots receipt, independently verifiable with OpenTimestamps tooling, even if LegalStamp were to disappear.

The limits are stated plainly: no presumption of accuracy (reserved for qualified timestamps), no proof of identity or ownership, no certification of human origin. What LegalStamp offers is the ability to timestamp systematically (working files, intermediate versions, deliverables) at a cost compatible with a regular practice, and thus to build the verifiable chronology most creators lack on the day a challenge arrives. How it works is detailed on our explainer page.

Conclusion

Article 50 of the AI Act establishes a world where synthetic content must identify itself, on pain of substantial fines. But it says nothing about the reverse situation: a challenged human creator gets no official "human-made" label. Their best resource remains evidence built in advance: a chain of working files and versions dated in a verifiable way, assembled before any release.

Blockchain timestamping is only one brick in that body of evidence: it freezes dates, not identities or rights. But it is the one chronological brick that nothing can reconstruct after the fact. In the post-Article 50 ecosystem, it is worth laying at the right moment: before you publish, not after the dispute.

Disclaimer: this article is provided for informational and educational purposes only. It does not constitute legal advice. For a specific situation (dispute, compliance, litigation), have your evidence strategy reviewed by a qualified legal professional.