Skip to main content Scroll Top

Article 50 of the EU AI Act: Could AI Marking Become a Persistent Identifier?

[Ready To Publish] Article 50 of the EU AI Act Could AI Marking Become a Persistent Identifier_Evane Alexandre_Image
Provenance, attribution and confidentiality under Article 50
Abstract

Article 50 of Regulation (EU) 2024/1689 (the “AI Act”) requires providers of certain generative artificial intelligence (“AI”) systems to ensure that synthetic audio, image, video or text outputs are marked in a machine-readable format and detectable as artificially generated or manipulated.

These transparency obligations became applicable on 2 August 2026, although providers of relevant AI systems placed on the market before that date have until 2 December 2026 to comply with Article 50(2).

As implementation moves into practice, attention is turning not only to how providers will comply, but also to the infrastructure through which compliance will be achieved. One question is whether technologies introduced to establish that content is synthetic could also carry information about its provenance. Where that information is sufficiently durable and linkable to a system, transaction, account or individual, a mechanism intended to establish artificial origin may also create a route to attribution.

I. Article 50: the legal starting point

Article 50(2) requires providers of AI systems, including general-purpose AI systems, that generate synthetic audio, image, video or text content to ensure that their outputs are marked in a machine-readable format and detectable as artificially generated or manipulated.

The technical solutions must be effective, interoperable, robust and reliable as far as technically feasible, taking account of the nature of the content, implementation costs and the state of the art. The obligation is subject to specified exceptions, including where an AI system performs an assistive function for standard editing or does not substantially alter the deployer’s input or its semantics.

The AI Act does not prescribe a single technical method. Recital 133 refers, non-exhaustively, to watermarks, metadata identification, cryptographic methods for proving provenance and authenticity, logging methods and fingerprints.

The statutory objective is therefore machine-readable marking and detectability. Attribution to the person who generated the content is not part of that requirement. But the technologies used to achieve it may also record information about the origin and history of content and, where combined with information held elsewhere, enable it to be traced to a particular generation event, account or user.

II. From detectability to provenance and attribution

Detectability, provenance and attribution are related but distinct: (i) detectability establishes that AI was involved; (ii) provenance provides information about the origin or history of content; and (iii) attribution connects it to a particular system, organisation, account or person.

Technologies capable of supporting this kind of attribution already exist. In film and premium video distribution for instance, forensic watermarking can embed unique or session-specific identifiers into individual copies or streams. If content is later leaked, the watermark can be extracted and used to identify the distribution copy, account or viewing session from which it originated.

The relevance to Article 50 is not that AI marking will necessarily work in the same way, but that machine-readable marking technologies are already capable of combining content identification with source tracing.

The same distinction appears in provenance standards such as those developed by the Coalition for Content Provenance and Authenticity (“C2PA”). C2PA enables cryptographically verifiable information concerning the origin and history of digital content to be associated with an asset. Its architecture can support identity and attribution, but does not require them: identity assertions are optional and may be anonymous or pseudonymous.

The final Code of Practice on Transparency of AI-Generated Content similarly reflects a layered approach. It contemplates machine-readable marking through techniques including digitally signed metadata and imperceptible watermarking, while richer provenance information may sit alongside the basic indication that content was AI-generated or manipulated. The Code also cautions against including unnecessary privacy-sensitive or business-sensitive information.

The technical question is therefore not simply whether content is marked, but what information the marking and associated provenance infrastructure preserve. A system might record only that content was AI-generated. A richer implementation might record the generating service, model, time of generation or another reference associated with the generation event.

None of those elements necessarily identifies an individual. But where a reference corresponds to information retained elsewhere – for example, an account record held by the provider – it may provide a route to attribution. If that relationship remains capable of being resolved over time, the reference can operate as what this article describes as a persistent identifier: a durable machine-readable reference through which content can be correlated or linked back to its source.

III. When provenance becomes personal data

That possibility engages Regulation (EU) 2016/679 (the “GDPR”).

Personal data includes information relating to an identified or identifiable natural person. Identifiability is not limited to information that names a person directly; it may arise where information can reasonably be combined with other data to identify an individual.

A marker recording only that content is AI-generated will ordinarily fall well short of identifying its creator. The analysis changes where the marker or associated provenance data contains a unique reference capable of being linked to a generation event, account or user. A pseudonymous identifier may therefore remain personal data where additional information permits it to be connected to an identifiable person.

Where personal data is processed, the GDPR principles apply, including lawfulness, fairness and transparency, purpose limitation, data minimisation, storage limitation, and integrity and confidentiality.

Data minimisation and purpose limitation are particularly relevant. If the Article 50 objective can be achieved by establishing that content is artificial, the inclusion or retention of information capable of linking that content to a particular user raises further questions: what purpose does that additional linkability serve, who needs to be capable of resolving it, and for how long?

The AI Act does not displace that analysis. EU law on personal-data protection, privacy and confidentiality of communications continues to apply, and the EDPB’s 2026–2027 work programme includes joint guidance with the Commission on the interaction between the AI Act and the GDPR.

IV. Confidentiality, trust and secondary use

The consequences extend beyond data protection.

Businesses and professionals use generative AI systems in connection with confidential, commercially sensitive and unpublished material. Contractual protections typically address what a provider may do with that material within its own systems – for example, retention, training and disclosure.

A linkable provenance mechanism raises a different issue. If information associated with an output remains capable of being connected to the account from which it was generated, the confidentiality question extends beyond how the provider handles the user’s input. It also concerns whether the output itself carries, or can be correlated with, a durable record of its source.

That distinction matters because provenance information may acquire uses beyond transparency. A marker or associated record identifying a generation event could become relevant to an internal investigation, regulatory enquiry or litigation where the origin of disputed content is in issue. If a provider holds records capable of resolving an identifier to an account, those records may also become relevant to lawful demands for information or disclosure.

The forensic-watermarking analogy is instructive. The value of such a watermark lies precisely in the fact that information embedded in content can later be extracted when the provenance of a particular copy becomes important.

The same possibility exists with AI provenance. Once reliable information linking content to its generation history exists, its usefulness will not necessarily remain confined to the transparency purpose that led to its creation.

The implications therefore depend heavily on architecture: what information is embedded in the content, what information is maintained externally, who can access any mapping between them, and how long that relationship remains capable of being reconstructed.

V. Drawing the boundary

Article 50 will contribute to an infrastructure in which AI-generated content is increasingly machine-readable and capable of carrying information about its origin. The distinction between detectability, provenance and attribution will therefore matter. A marker may simply establish that AI was involved; it may also form part of a provenance system capable of linking content to a generation event, account or individual.

Forensic watermarking shows that such traceability is technically well established, while C2PA and the final Article 50 Code demonstrate that attribution is not an inevitable consequence of provenance. The boundary is therefore largely a matter of design.

As Article 50 is implemented at scale, the key question is not only how reliably AI-generated content can be detected, but what information the emerging marking infrastructure should preserve, who should be able to resolve it, and for what purposes.

Related Posts