Art 50(2) for developers: mark AI output, earlier systems by 2 December 2026
What Article 50(2) of the EU AI Act requires providers to mark in AI output, who gets until 2 December 2026, and what the Code of Practice asks of developers.
Key Takeaways
- 1.Build two marking layers: signed metadata (provenance data with a digital signature) and a watermark (a signal hidden in the content). Free-form text (a raw sequence of characters with no container format, such as chat or web text) and certain physical products need only one layer (Code of Practice, Measure 1.1).
- 2.The Code of Practice is voluntary, and its Section 1 for providers names no marking standard, so C2PA is an example, not a requirement.
- 3.Free-form text longer than 200 tokens still needs a watermark, the one layer it can carry, while the Code excepts very short text (Measure 1.1.2; Code glossary).
- 4.Providers of systems placed on the market before 2 August 2026 must comply with Article 50(2) by 2 December 2026 (Art 111(4)). Systems first placed on the market on or after 2 August 2026 have no grace period (Art 111(4), Art 113).
Providers that place AI systems on the EU market must mark synthetic audio, image, video or text outputs in a machine-readable format. This is Article 50(2) of Regulation (EU) 2024/1689, the AI Act. The outputs must also be detectable as artificially generated or manipulated.
The duty applies from 2 August 2026 (Art 113). Systems placed on the market before that date have until 2 December 2026, a grace period (Art 111(4)). Placing on the market means the first making available of an AI system on the EU market (Art 3(9)). The duty applies to providers only and has exceptions for assistive editing (Art 50(2)).
The voluntary Code of Practice on Transparency of AI-Generated Content asks for two layers, signed metadata and a watermark. Its Section 1 names no marking standard, such as C2PA. C2PA is an open specification for signed provenance data.
Fines reach up to EUR 15 000 000 or, if the offender is an undertaking, 3 % of worldwide annual turnover, whichever is higher (Art 99(4)(g)). Small and medium-sized enterprises (SMEs) pay whichever is lower (Art 99(6)). Check which of your outputs are synthetic and whether you placed your system on the market before 2 August 2026.
Article 50(2) requires providers to mark outputs and make them detectable
Article 50(2) makes the provider responsible for the marking, and it covers general-purpose AI systems too. The Act's wider transparency duties are in Article 50: Transparency & Disclosure Requirements.
The provider's technical solutions must be "effective, interoperable, robust and reliable as far as this is technically feasible" (Art 50(2)). The text adds three factors: the limits of each content type, implementation costs and the generally acknowledged state of the art.
The duty does not apply to the extent the system performs an assistive function for standard editing. It also does not apply where the system does not substantially alter the input data or its semantics. A third exception covers cases where the law authorises detecting, preventing, investigating or prosecuting criminal offences (Art 50(2)).
Small mid-cap enterprises (SMCs) pay the same lower fine as SMEs (Art 99(6a)). SMC is a size class that Regulation (EU) 2026/1744 (the Digital Omnibus on AI, which amends the AI Act) added as Article 3(14b).
National market surveillance authorities, the national bodies that enforce product rules, enforce the duty. The AI Office enforces it for the systems it supervises (Art 75(1); Commission quick-facts page on transparency rules).
The grace period ends on 2 December 2026, and only for systems placed on the market earlier
Providers of systems placed on the market before 2 August 2026 have until 2 December 2026 to comply with Article 50(2). Article 111(4) names only those systems. Article 113 applies Article 50 from 2 August 2026, so systems first placed on the market on or after that date have no grace period.
Article 111(4) says those providers "shall take the necessary steps in order to comply with Article 50(2) by 2 December 2026". It covers providers only and Article 50(2) only.
Regulation (EU) 2026/1744 inserted Article 111(4). The Omnibus is dated 8 July 2026 and appeared in the Official Journal on 24 July 2026. It entered into force on 27 July 2026 (Art 4, Regulation (EU) 2026/1744).
Within Article 50, Regulation (EU) 2026/1744 changed only paragraph 7, on codes of practice. Paragraph 2 is unchanged, and so is the deployer-side disclosure in paragraph 4.
The Code of Practice sets a two-layer approach, but it is voluntary and its Section 1 names no standard
The voluntary Code of Practice on Transparency of AI-Generated Content asks providers for two layers (signed metadata and a watermark). Its Section 1 names no marking standard. Section 1 sets Commitments, each split into Measures, the Code's numbered steps. Independent experts drafted the Code, facilitated by the AI Office, the Commission's EU-level AI body. It reached its final version on 10 June 2026 (Commission code page). Section 1 covers providers (Art 50(2) and (5)). Section 2 covers deployers, the organisations that use AI systems (Code of Practice, Section 2).
Adherence "does not constitute conclusive evidence of compliance" (Code of Practice, Objectives). Signatories can rely on the Code's measures to demonstrate compliance (Commission code page). Non-signatories must show the market surveillance authority that their measures are adequate (Commission code page).
The Commission published its opinion on 9 July 2026. It found that the Code adequately covers Art 50(2), (4) and (5) (Commission opinion). The Commission Guidelines on the implementation of the transparency obligations for certain AI systems under Article 50 of Regulation (EU) 2024/1689 were first published on 20 July 2026. The English text dated 5 November 2026 is C(2026) 7702 final. The Guidelines are non-binding (Guidelines, para (5)). They say providers that do not adhere to the Code must show compliance through alternative, equivalently adequate means (Commission guidelines policy page).
| Code measure | What the Code says | What it means for your build |
|---|---|---|
| 1.1.1 Signed metadata | Record AI origin where the data format supports metadata. Sign and time-stamp it in a tamper-evident way. | Sign at generation and protect the signing keys. |
| 1.1.2 Watermark | Hidden and hard to separate from the content. The Code excepts very short text. Its glossary defines that as text shorter than 200 tokens, until better methods emerge. Free-form text longer than 200 tokens still needs a watermark, which may have lower reliability. | Add a watermark per modality, and watermark free-form text longer than 200 tokens. |
| 1.1.3 Fingerprinting or logging | Optional. Fingerprinting (matching outputs against stored signatures) or logging alone is not sufficient. | Treat it as an extra, never as the marking. |
| 1.2 Non-removal | Keep existing metadata where technically feasible. Ban intentional removal in your terms. Open-source systems may alert users in the documentation instead. | Pass metadata through your pipeline and edit your acceptable-use policy. |
| 2.1 Detection | Provide a "detection solution" (we call it a detection interface) per technique, as a public specification, a library or a cloud API. It is free of charge, except that providers with fewer than 1 000 000 monthly users and substantial operating costs may charge a reasonable fee above a volume threshold. Access to text watermark detection may be limited to verified expert users (Measure 2.1.2). | Ship a detection interface. |
| 3.4 Interoperability | Adopt established standards or best practices for metadata marking and detection when Art 50(2) applies. By 2 February 2027, implement one or more of four interoperability solutions for watermark detection: an industry-standard access method, a public signpost, a shared consortium detector or a comparable solution. | Plan a second release for 2 February 2027. |
| Commitment 3 | Measure reliability and test robustness against re-compression, screenshot, cropping, paraphrasing and similar operations. | Build these tests into your release pipeline. |
| 4.1 and 4.2 Process and testing | Keep a documented compliance process. Test before placing the system on the market or putting it into service, and regularly thereafter (Measure 4.2). | Write down who tests what and when. |
The Code states the multi-layer duty for audio, images, video and containerised text, such as PDF, Word or HTML files. Free-form text needs a single layer, the watermark, because free-form text cannot carry metadata (Measure 1.1).
A single layer is also considered sufficient where a generative AI system is embedded in physical products "in a technically controlled and closed environment mainly instructive in nature". This applies "to the extent effective technical measures are embedded in the product to prevent the output from leaving the environment of the product" (Measure 1.1). Measure 1.1 does not say which layer. Measure 3.3 exempts systems "exceptionally subject only to one layer of metadata marking" from its robustness requirement.
The Commission's Code and task-forces pages (updated October 2026) list no SDK, reference implementation or test suite, and the open specifications have known limits
The Commission's code page (last updated 19 October 2026) and task-forces page (last updated 28 October 2026) list no SDK, reference implementation or test suite. The Commission's AI standardisation page (last updated 19 October 2026) names ten areas for the CEN/CENELEC request and does not mention Article 50, marking or watermarking.
When the Code was published, it said interoperability standards and best practices were still to be developed, except for signed metadata (Measure 3.4). It also called for relevant watermarking standards (Code of Practice, recital (f)).
You can still build on existing open specifications.
- C2PA is an open specification from a project of the Joint Development Foundation (c2pa.org). A C2PA manifest is a set of provenance information with digitally signed assertions (C2PA explainer, spec.c2pa.org).
- The IPTC digital source type trainedAlgorithmicMedia is labelled "Created using Generative AI" (cv.iptc.org).
Metadata can be removed from a file (C2PA explainer). The Code concedes that "in most cases, in particular for content that can be disseminated online, no single marking technique suffices" (Code of Practice, recital (b)). When it was published, the Code judged forensic detection not mature enough (Measure 2.2). A watermark on free-form text may have lower reliability (Measure 1.1.2). Plan for removal and test for it.
Nine steps map the Act and the Code onto your build
- Decide which outputs are synthetic audio, image, video or text, and whether an exception applies (Art 50(2); Code, Commitment 1).
- Check whether you placed the system on the market before 2 August 2026 (Art 111(4)).
- Choose a signed metadata layer and a watermark layer per modality. Use the watermark alone for free-form text. A single layer is also sufficient for AI embedded in physical products "in a technically controlled and closed environment mainly instructive in nature" (Code, Measure 1.1). That holds to the extent effective technical measures stop the output leaving the product.
- Sign and time-stamp the metadata, and protect the signing keys (Code, Measure 1.1.1).
- Keep existing metadata through your own pipeline where technically feasible (Code, Measure 1.2).
- Add the non-removal clause to your acceptable-use policy or terms. For open-source systems, the Code lets you alert users in the documentation instead (Code, Measure 1.2).
- Provide a detection interface for each technique, free of charge unless the fee exception applies. Access to text watermark detection may be limited to verified expert users (Code, Measures 2.1 and 2.1.2).
- Test against re-compression, screenshot and cropping, and against paraphrasing for text (Code, Commitment 3).
- Document the process. Test before placing the system on the market or putting it into service, and regularly thereafter. Plan the 2 February 2027 interoperability step (Code, Measures 4.1, 4.2 and 3.4).
How Scanara helps: it looks for marking patterns in your code and structures the Art 50 work
Scanara supports discovery and documentation of your Article 50 work. It does not build or verify your marking.
- Code scan. Scanara's rule-based code scan looks for marking patterns in source code for Article 50(2). The keywords include watermark, provenance, content_credentials, c2pa and synthetic_content_marking. A hit shows where marking-related code may exist. A missing hit does not show that none exists. The scan detects standard patterns only and can miss non-standard code.
- Guided assessment. The transparency assessment guides you through Article 50.
- Disclosure templates. Scanara provides document templates for the Article 50 disclosures.
You build and ship the marking and remain responsible for whether it is effective. Scanara does not verify that a marking survives re-encoding. A passing scan does not mean your system is compliant.
Next step: scan your repository with Scanara, then open the transparency assessment to record which outputs you mark and how.
Three questions stay open, and authorities or courts decide them
Open question 1: does a re-release or update of a system placed on the market before 2 August 2026 still count as placed before that date? Article 111(4) does not say. Market surveillance authorities decide in practice, and the Court of Justice decides ultimately. Until then, treat any update that changes how your system generates or marks content as a risk to the grace period. That is our advice, not Article 111(4) text.
Open question 2: whether "put into service" also qualifies. The Commission Guidelines describe the transition as covering systems "placed on the market or put into service before 2 August 2026" (Commission Guidelines, C(2026) 7702 final, para 153). Article 111(4) says only "placed on the market". The Guidelines are a non-binding Commission view, and the Court of Justice ultimately decides which reading holds.
Open question 3: what "effective, interoperable, robust and reliable as far as this is technically feasible" requires. The answer depends on the state of the art and on relevant technical standards. Market surveillance authorities assess each case. The Code expects methods and benchmarks that the AI Office endorses (Code of Practice, Measure 4.2). Before endorsing them, the AI Office consults the AI Board, the body of Member State representatives for the AI Act (Code of Practice, Measure 4.2).
Related Reading
Sources & References
- Regulation (EU) 2024/1689 (AI Act, original text) (Art 3(9), 50, 98, 99(4), 99(6), 113)
- Regulation (EU) 2026/1744 (Digital Omnibus on AI) (Art 111(4), 99(6a), 3(14b), replaced Art 75(1), Art 113 as amended, replaced Art 50(7), dates, Art 4)
- Code of Practice on Transparency of AI-Generated Content (final, 10 June 2026) (Section 1 measures, glossary, recitals)
- European Commission, code of practice on AI-generated content (last updated 19 October 2026) (voluntary status, non-signatories, timeline, no SDK listed)
- European Commission, opinion assessing the Code of Practice (adequacy for Art 50(2), (4), (5))
- European Commission, Guidelines on transparency obligations for providers and deployers of AI systems (library page) (first published 20 July 2026)
- European Commission, Guidelines on the implementation of the transparency obligations for certain AI systems under Article 50 of Regulation (EU) 2024/1689 (English PDF, C(2026) 7702 final, 5 November 2026) (non-binding, para (5); transition, para 153)
- European Commission, guidelines on AI transparency obligations (policy page) (alternative, equivalently adequate means)
- European Commission, quick facts on transparency rules for AI systems (enforcement by national authorities and the AI Office)
- European Commission, task forces for the code of practice (last updated 28 October 2026) (no SDK, reference implementation or test suite listed)
- European Commission, AI Act standardisation (last updated 19 October 2026) (ten areas of the CEN/CENELEC request; no mention of Article 50 or marking)
- C2PA, about page (Joint Development Foundation project)
- C2PA Explainer (specification version 2.2) (manifests, removal limit)
- IPTC digital source type (label and definition)
Frequently Asked Questions
Can I limit access to my text watermark detection interface?
Yes, for text watermark detection: access may be limited to verified expert users, until more reliable detection exists (Measure 2.1.2).
Does the grace period cover a system first placed on the market on or after 2 August 2026?
No. Article 111(4) names only systems placed on the market before that date, and Article 113 applies Article 50 from 2 August 2026.
Is fingerprinting or logging enough on its own?
No. The Code makes it optional and states that relying on it alone is not sufficient (Measure 1.1.3).
Can I charge for my detection interface?
Only in one case. Providers with fewer than 1 000 000 monthly users and substantial operating costs may charge a reasonable fee above a volume threshold (Measure 2.1).
Can the Commission still set mandatory marking rules?
It may adopt an implementing act, a binding Commission rule adopted under Article 98(2), only if it deems the Code inadequate (Art 50(7)). Its opinion, published on 9 July 2026, found the Code adequate for Art 50(2), (4) and (5).
Scanara prepares evidence from your repositories
Scanara reads your connected GitHub repositories and runs rule-based scans for standard AI and machine-learning patterns. It can miss non-standard code. Separately, it generates draft documentation from templates and your assessment answers as evidence for your compliance work. You make the classification decisions.