Transparency obligations for providers and deployers of certain AI systems

1. Purpose and context

Article 50 is the only provision of Chapter IV of the AI Act, the chapter devoted to transparency obligations. It establishes a distinct, lighter-touch regulatory layer that applies irrespective of whether an AI system is high-risk. Recital 132 explains the rationale: “Certain AI systems intended to interact with natural persons or to generate content may pose specific risks of impersonation or deception irrespective of whether they qualify as high-risk or not.”¹ The response is targeted: not bans or conformity assessment, but disclosure — telling people when they are dealing with AI, or when content has been artificially generated or manipulated. Transparency is one of the seven non-binding ethical principles the Regulation draws on (the AI HLEG's 2019 guidelines), alongside human agency and oversight; technical robustness and safety; privacy and data governance; diversity, non-discrimination and fairness; societal and environmental well-being; and accountability.²

The provision addresses four distinct situations, and the duty-bearer differs between them: providers carry the duties in paragraphs (1) and (2); deployers carry those in paragraphs (3) and (4). The Commission Guidelines on Article 50 (the “Guidelines”) — whose content the Commission approved on 20 July 2026 by Communication C(2026) 5054 final, with formal adoption to follow once all language versions are available, a step still outstanding in mid-September 2026 — confirm that these obligations apply to any AI system used in the covered situations, not only to high-risk or general-purpose AI, and that they can cumulate on a single output: where an AI system generates images as part of a direct interaction with a natural person, its provider may have to comply with both 50(1) and 50(2), and where that system is also able to generate images constituting deep fakes or text published on matters of public interest, 50(4) is added on the deployer's side.³ The Guidelines describe the shared purpose of the four duties as being “to reduce the risks of impersonation, deception, misinformation, manipulation at scale, and fraud.”³

The underlying concern is empirical. Recital 133 records that synthetic content “becomes increasingly hard for humans to distinguish from human-generated and authentic content,” raising “new risks of misinformation and manipulation at scale, fraud, impersonation and consumer deception.”⁴ Peer-reviewed evidence corroborates this premise on both limbs of the provision. For generated content, a 2024 systematic review and meta-analysis of 56 studies (86,155 participants) found human deepfake-detection accuracy at 55.5%, not significantly above chance (Diel et al., 2024), and other work shows people both fail to detect deepfakes and overestimate their own ability to do so (Köbis et al., 2021).⁵ For interaction, six experiments with 4,600 participants found that people could not identify AI-generated self-presentations in professional, hospitality and dating contexts, and that the intuitive cues they rely on are systematically wrong and can be exploited to make machine text read as more human than human writing (Jakesch et al., 2022).⁶ For emotion recognition specifically, Recital 44 notes “serious concerns about the scientific basis” of such systems, citing “limited reliability, the lack of specificity and the limited generalisability”⁷ — which helps explain why Article 50(3) imposes a notification duty rather than relying on the system's outputs.

Article 50's obligations become applicable on 2 August 2026 (Article 113). Six days before that date, on 27 July 2026, the amending Regulation (EU) 2026/1744 (the Digital Omnibus on AI, OJ L 2026/1744 of 24 July 2026) entered into force; it leaves the substance of the four duties untouched but recasts Article 50(7) and adds a transitional rule for the Article 50(2) marking duty — see sections 3 and 7.⁸

2. Interpretation of key terms

Article 50 turns on a small set of defined terms in Article 3; the boundaries of each definition determine the reach of the duties.

• AI system (Art 3(1)): “a machine-based system that is designed to operate with varying levels of autonomy and that may exhibit adaptiveness after deployment, and that, for explicit or implicit objectives, infers, from the input it receives, how to generate outputs such as predictions, content, recommendations, or decisions that can influence physical or virtual environments.”⁹ The definition is the gateway to the whole Regulation; deterministic software that does not “infer” outputs falls outside it.

• Provider vs deployer (Art 3(3) and (4)): a provider is a natural or legal person, public authority, agency or other body that “develops an AI system or a general-purpose AI model or that has an AI system or a general-purpose AI model developed and places it on the market or puts the AI system into service under its own name or trademark, whether for payment or free of charge.”⁹ A deployer is, correspondingly, a person, authority, agency or body “using an AI system under its authority except where the AI system is used in the course of a personal non-professional activity.”⁹ Two features of the provider definition matter for Article 50: the trigger is placing on the market / putting into service under one’s own name or trademark; and it applies whether for payment or free of charge (free and open offerings are not exempt). Both terms sit within the wider category of operator (Art 3(8)), which also covers product manufacturers, authorised representatives, importers and distributors. The split matters: the marking duty (50(2)) sits with whoever develops the generative AI system and places it on the market or puts it into service under its own name or trademark; the disclosure duties for deep fakes, and for AI-generated text published in order to inform the public on matters of public interest (50(4)), sit with the person who actually uses the AI system to produce that content. Personal non-professional activity falls outside 50(3) and 50(4) on two grounds: the definition in Art 3(4) excludes from the notion of deployer the case “where the AI system is used in the course of a personal non-professional activity,” and Article 2(10) adds that the Regulation “does not apply to obligations of deployers who are natural persons using AI systems in the course of a purely personal non-professional activity” — the provision on which the Guidelines rely. An individual using a deepfake tool privately, outside any professional or commercial context, therefore carries no disclosure duty under Article 50 (though national criminal law may still apply — see section 6).^9,10 The Commission’s Q&A accompanying the Guidelines draws the line at regularity and economic purpose: a natural person who generates deepfakes and disseminates them on social media in a personal capacity is outside the Act, but an activity through which the person gains an economic benefit on a regular basis, or that otherwise forms part of a business, trade, occupational or freelance activity, is professional and makes that person a deployer; and where the deployer is a legal person, the employees acting under its instructions and control — digital animators, web designers, content creators, journalists — are not separate deployers of the system.¹¹ That carve-out does not travel upwards: the Guidelines make clear that the provider’s marking duty under 50(2) is unaffected by the private character of the end use, so a system used to make a deepfake for a Christmas card must still mark its output.³ Three further boundaries drawn by the Guidelines matter in practice. First, the same entity can be both provider and deployer — typically where it develops a system in-house and also uses it. Second, actors who merely disseminate somebody else’s AI-generated content without authority over its generation — hosting services, online platforms, broadcasters — are not deployers for the purposes of 50(4), though they are encouraged to preserve any marking and labelling they receive; nor does labelling content or using already-labelled content by itself make a party a deployer. Third, the duties reach third-country actors: providers are covered where the output is used in the Union — though “incidental, unforeseeable or unauthorised downstream use should not alone trigger the application” of the duties — while a third-country deployer is covered only where it itself intends the output to be disseminated or used in the Union, not where content reaches audiences in the Union through channels that are unforeseeable and outside its control.³

• Deep fake (Art 3(60)): “AI-generated or manipulated image, audio or video content that resembles existing persons, objects, places, entities or events and would falsely appear to a person to be authentic or truthful.”⁹ The boundary line is the “would falsely appear to a person to be authentic or truthful” element: wholly fantastical content is outside the definition, whereas realistic manipulation of a real person is squarely within it. The Guidelines unpack the definition into four cumulative conditions, all of which must be met: (i) the content bears an appreciable resemblance; (ii) to something that exists; (iii) falling within one of the listed classes — persons, objects, places, entities or events; and (iv) it would falsely appear to a person to be authentic or truthful. They add two refinements that matter for borderline cases: the false-appearance test is objective and does not depend on any intention to deceive, and photorealism raises the likelihood of falling within the definition without being decisive on its own. Content that is plainly unrealistic (dragons, people flying unaided) stays outside.³ Note that text is not part of the deep-fake definition; AI-generated text is dealt with separately under the second limb of 50(4).

• Emotion recognition system (Art 3(39)) and biometric categorisation system (Art 3(40)): respectively, an AI system “for the purpose of identifying or inferring emotions or intentions of natural persons on the basis of their biometric data,” and one “for the purpose of assigning natural persons to specific categories on the basis of their biometric data, unless it is ancillary to another commercial service and strictly necessary for objective technical reasons.”⁹ Both build on biometric data (Art 3(34)). The ancillary-and-strictly-necessary carve-out in (40) keeps, for example, a face-detection filter that merely positions a camera outside the categorisation regime. The interaction with Article 5 is important: emotion recognition in the workplace and in education, and certain sensitive biometric categorisation, are prohibited outright; Article 50(3) governs only the permitted residual uses, for which it requires notification rather than imposing a ban.

• General-purpose AI model vs general-purpose AI system — these are not synonyms and must not be conflated. A general-purpose AI model (Art 3(63)) is an AI model that “displays significant generality” and can be integrated into downstream systems; a general-purpose AI system (Art 3(66)) is an AI system “based on a general-purpose AI model” that can serve a variety of purposes.⁹ By way of illustration only, and without qualifying any particular product: a language model such as Opus 5 is a model, whereas the Claude application that makes such a model available to users is a system. Article 50(2) expressly applies to “AI systems, including general-purpose AI systems”¹² — it is a duty on systems, not on models. The distinct transparency and documentation duties for general-purpose AI models live in Articles 53–55 (Chapter V), not here. Recital 133 adds the practical bridge: marking techniques “can be implemented at the level of the AI system or at the level of the AI model, including general-purpose AI models generating content, thereby facilitating fulfilment of this obligation by the downstream provider of the AI system.”⁴

3. Paragraph-by-paragraph analysis

Art 50(1) — Notification of interaction with an AI system (provider duty). Providers must ensure that AI systems “intended to interact directly with natural persons” are “designed and developed” so that the persons concerned are informed they are interacting with an AI system, “unless this is obvious from the point of view of a natural person who is reasonably well-informed, observant and circumspect, taking into account the circumstances and the context of use.”¹² The classic case is a customer-service chatbot. The “obviousness” qualifier is framed as an objective benchmark — assessed from the standpoint of a person who is “reasonably well-informed, observant and circumspect” in the circumstances, not by the subjective impression of the particular user — and Recital 132 directs that the characteristics of vulnerable groups (age, disability) be taken into account where the system is intended to interact with them.¹

The Guidelines break the trigger into four cumulative elements: (i) an AI system; (ii) an intended interaction; (iii) which is direct; (iv) with a natural person. They use those elements to place three familiar situations outside the paragraph. Passive data capture is outside it because there is no interaction with the person at all: a biometric access-control system reads a face and opens a door, but conducts no exchange with the individual. One-off feedback collection is outside it because there is no exchange with the person: the Guidelines give spam filters as an example of systems that just collect one-time feedback and are therefore not considered to interact with natural persons. And interaction mediated by a human is outside it because it is not direct: where a customer-service agent consults an AI assistant and then answers in their own words, the AI's output never reaches the customer. The Guidelines also address AI agents, which are inside the paragraph: they must disclose both their artificial nature and the person on whose behalf they act, and where a provider cannot know in advance whether an agent will deal directly with a natural person, the agent should be designed to disclose itself in every situation where that is reasonably likely. The form of the notification is open — text, audio, visual or a combination — but the Guidelines list what does not discharge the duty on its own: burying the information in terms and conditions or behind a link, relying on machine-readable marks the user cannot perceive, ambiguous references to an “assistant,” generalised statements such as “Services on this website use AI,” and purely technical descriptions such as “this system uses LLMs.” Consistently with its character as an exception, obviousness is to be read restrictively, in two steps: identify the target or reasonably foreseeable audience, then ask what an average member of it would understand. AI coding assistants aimed only at professionals, internal tools used by trained staff, and non-player characters in a single-player game are given as cases where the interaction is obvious; a robotic companion pet imitating an animal, realistic avatars in immersive VR or AR, and customer-support chatbots whose output may read as human are cases where it is not.³

The law-enforcement exception is narrow and conditional. It removes the duty for “AI systems authorised by law to detect, prevent, investigate or prosecute criminal offences,” but only “subject to appropriate safeguards for the rights and freedoms of third parties,” and — critically — it does not apply where “those systems are available for the public to report a criminal offence.”¹² So the exception covers, for example, a covert investigative tool used under a legal authorisation with third-party safeguards; it does not cover a public-facing crime-reporting chatbot, where ordinary members of the public interact with the system and the transparency rationale is undiminished.

Art 50(2) — Machine-readable marking of synthetic content (provider duty). Providers of AI systems — “including general-purpose AI systems” — that generate “synthetic audio, image, video or text content” must ensure the outputs “are marked in a machine-readable format and detectable as artificially generated or manipulated.”¹² The duty is qualified by feasibility: solutions must be “effective, interoperable, robust and reliable as far as this is technically feasible, taking into account the specificities and limitations of various types of content, the costs of implementation and the generally acknowledged state of the art, as may be reflected in relevant technical standards.”¹² Recital 133 lists candidate techniques — “watermarks, metadata identifications, cryptographic methods for proving provenance and authenticity of content, logging methods, fingerprints or other techniques.”⁴ The Code of Practice on Transparency of AI-Generated Content, finalised on 10 June 2026, is the Union-level compliance vehicle for this duty (see section 5); it prescribes no particular technique, tracking the statutory feasibility standard instead.¹³ Its draft iterations discussed a multi-layered approach combining secured metadata (e.g. C2PA Content Credentials) with invisible watermarking (e.g. SynthID) — a pairing that responds to the fact that no EU harmonised standard for Article 50(2) marking is yet in force and that metadata can be stripped by recompression or re-uploads.¹⁴ The statutory hedge (“as far as this is technically feasible”) tracks the technical reality that watermarking is, in the words of the literature, “not a silver bullet” and can be evaded.¹⁵

Two exceptions apply: the duty does not bite “to the extent the AI systems perform an assistive function for standard editing or do not substantially alter the input data provided by the deployer or the semantics thereof,” nor where the use is “authorised by law to detect, prevent, investigate or prosecute criminal offences.”¹² The first carve-out keeps routine tools (spell-checkers, format conversion, minor colour or crop adjustments) outside the marking duty; it turns on whether the system substantially alters the input or its semantics.

The Guidelines add four points of practical importance. First, the duty has two cumulative components — machine-readable marking and the availability of detection — so a provider that marks its outputs without making them detectable has not complied. Second, on the technical side providers should prefer publicly available industry standards that are interoperable across providers, with proprietary or shared solutions treated as a stop-gap until harmonised standards exist; the four statutory qualities (effective, interoperable, robust, reliable) are assessed against the state of the art and technical feasibility, and implementation costs may be taken into account. Third, the modality classes are read broadly: 3D content and VR and AR environments are treated as “video,” and digital twins fall within Article 50(2) where they are AI-generated or manipulated and take the form of audio, image, video or text — their modality being determined case by case rather than classified as video by default. Outside the paragraph fall short functional sequences (alt-text, interface labels), source code and structured machine formats, and purely machine-to-machine exchanges — as do intermediate steps of an AI agent that never surface as perceptible content. Fourth, the Guidelines identify narrow situations in which the marking expectation is relaxed, and they are not all relaxed in the same way. For technically closed products whose output cannot leave the product environment — in-vehicle navigation is the example — what the Guidelines contemplate is a less robust marking solution, such as metadata, not the absence of marking. Full non-marking is contemplated in two other, distinct cases: industrial or business-to-business applications meeting cumulative conditions (purely technical output, a limited predetermined circle of professionals, no shareability outside the undertaking); and ephemeral real-time content in games or VR that is not stored, where marking is not technically feasible and users are informed by other means, such as a notice within the experience.³ Adherence to a code of practice assessed as adequate under Article 50(7) is treated as a route to demonstrating compliance — and the Code has been so assessed, by the Commission on 8 July 2026 and by the Board on 9 July 2026 (see section 5).^3,13,16

Providers of generative systems already on the market benefit from a transitional rule inserted by the Digital Omnibus: under the new Article 111(4), “providers of AI systems, including general-purpose AI systems, generating synthetic audio, image, video or text content, that have been placed on the market before 2 August 2026 shall take the necessary steps in order to comply with Article 50(2) by 2 December 2026.” The deadline is therefore 2 December 2026. The Guidelines add that content generated before 2 August 2026 need not be marked retroactively unless it is published after that date.^8,3

Art 50(3) — Notification by deployers of emotion recognition / biometric categorisation systems. Deployers must “inform the natural persons exposed thereto of the operation of the system” and must process personal data in accordance with the GDPR, the EUDPR and the Law Enforcement Directive “as applicable.”¹² The duty is one of notification only; it does not authorise the underlying processing, which still requires its own legal basis under data-protection law. The law-enforcement exception here is again conditional: it covers systems “permitted by law to detect, prevent or investigate criminal offences,” but only “subject to appropriate safeguards for the rights and freedoms of third parties, and in accordance with Union law.”¹² Two boundaries should be kept in view: first, this paragraph governs only permitted uses — workplace and educational emotion recognition, and certain biometric categorisation, are prohibited by Article 5 and never reach Article 50; second, the notification duty does not displace any prohibition or consent requirement under data-protection law.

The Guidelines sharpen three further points. Every emotion recognition system that is not prohibited under Article 5(1)(f) is high-risk, so the 50(3) duty operates cumulatively with the Chapter III requirements rather than instead of them; by contrast, 50(3) catches any biometric categorisation system, whether or not it is high-risk. The content of the duty is narrow: the deployer must inform the persons exposed that the system is operating, not why or on what logic — reasons may be owed under data-protection law, but not under this paragraph — and the addressees include children. The law-enforcement exception is read more permissively here than in paragraphs (1), (2) and (4): where those exceptions “require an explicit law authorising the non-transparent use of the AI systems,” paragraph (3) is satisfied where the use is “permitted under the legal rules governing the powers of law enforcement authorities” to use such systems without informing the persons concerned. The safeguards for third parties’ rights and freedoms and the requirement of conformity with Union law continue to apply.³

Art 50(4) — Disclosure of deep fakes and public-interest text (deployer duty). This paragraph has two limbs. For deep fakes — image, audio or video content — deployers “shall disclose that the content has been artificially generated or manipulated.”¹² That obligation “shall not apply where the use is authorised by law to detect, prevent, investigate or prosecute criminal offence.”¹² Recital 134 grounds this in freedom of expression and the freedom of the arts: where the content forms part of “an evidently artistic, creative, satirical, fictional or analogous work or programme,” the obligation is “limited to disclosure of the existence of such generated or manipulated content in an appropriate manner that does not hamper the display or enjoyment of the work.”^12,17 The disclosure duty itself is not removed for such works; what is narrowed is the manner in which it is discharged. A satirical deepfake video must still be accompanied by a disclosure that the content is artificial, but that disclosure may take a form that does not spoil the work — a permanent intrusive overlay, for example, is not required. The Guidelines read the attenuated regime narrowly, taking the word “evidently” seriously, and stress that relying on it does not release the deployer from ensuring that “the rights and freedoms of third parties (including e.g. their right to the protection of personal data or intellectual property rights) are adequately safeguarded and respected”; it “cannot be a justification for failing to respect the fundamental rights of individuals or rightsholders under Union law on intellectual property or Union data protection law.” Separately — and for deep fakes generally, not only for the attenuated regime — they record that deployers must comply with Union intellectual-property law where the deep fake incorporates protected subject matter, and, where existing persons are depicted, with “relevant legal regimes protecting personality rights over personal images or voices.” They also locate the audience for this paragraph differently from 50(1): what matters is the actual, reasonably foreseeable audience of the content, including any vulnerable groups within it, rather than a hypothetical average person.³

The second limb concerns AI-generated or manipulated text “published with the purpose of informing the public on matters of public interest.” Here too there must be disclosure of artificial origin, but the duty falls away where “the AI-generated content has undergone a process of human review or editorial control and where a natural or legal person holds editorial responsibility for the publication of the content.”¹² This is the editorial-responsibility exception: a newsroom that runs AI-drafted copy through human editorial control, with a person accountable for publication, is not required to flag the text as AI-assisted. The second limb, too, is subject to the same law-enforcement exception.¹² The Guidelines give the second limb three elements — the text must be published, in the sense of being accessible to an indeterminate readership; published with the purpose of informing the public; and concern a matter of public interest, a category they illustrate with politics, public administration, justice, public health, security, the environment and economic, scientific or cultural developments. The exception itself has two cumulative elements: a genuine substantive human review or editorial control, which a spell-check does not supply, and a natural or legal person bearing editorial responsibility whose identity and contact details are publicly ascertainable — a notion the Guidelines connect to Article 2(8) of the European Media Freedom Act, Regulation (EU) 2024/1083.³ The Commission’s Q&A puts the two elements in operational terms: human review is the deliberate examination of the substance of the content by one or more natural persons with relevant knowledge and professional judgement in the subject matter; editorial control is the control exercised in practice by a responsible editorial entity, such as an editor-in-chief, with authority to approve, alter or reject the substance of the text on substantive grounds, including fact-checking; and superficial, solely formal or procedural checks such as spell-checking or grammatical correction do not qualify as either.¹¹

Art 50(5) — Manner and timing. The information under paragraphs (1)–(4) must be provided “in a clear and distinguishable manner at the latest at the time of the first interaction or exposure” and must conform to applicable accessibility requirements.¹² This is a cross-cutting quality standard: disclosure that is buried or delayed does not discharge the duty. The Guidelines read the reference to the first interaction or exposure per person: each further natural person who encounters the same output has their own first exposure, so for content that remains accessible — a deepfake circulating on a social network — the disclosure has to travel with it rather than being made once at publication. They also tie the accessibility requirement to Directives (EU) 2016/2102 and (EU) 2019/882.³

Art 50(6) — Relationship to other obligations. Paragraphs (1)–(4) “shall not affect the requirements and obligations set out in Chapter III” (the high-risk regime) and are “without prejudice to other transparency obligations laid down in Union or national law.”¹² Article 50 is therefore additive: a high-risk system that also interacts with people or generates content must satisfy both Chapter III and Article 50.

Art 50(7) — Codes of practice. This is the one paragraph of Article 50 whose wording has already changed. In its original form it made the AI Office “encourage and facilitate the drawing up of codes of practice at Union level” for the detection and labelling of artificially generated or manipulated content, and allowed the Commission to approve them by implementing act under Article 56(6) or, if it deemed a code inadequate, to set common rules by implementing act under the examination procedure in Article 98(2).¹² Regulation (EU) 2026/1744 replaced that paragraph with effect from 27 July 2026 — before Article 50 itself became applicable — so the text that governs from 2 August 2026 reads: “The Commission shall encourage and facilitate the drawing up of codes of practice at Union level to facilitate the effective implementation of the obligations regarding the detection, marking and labelling of artificially generated or manipulated content. The Commission, taking utmost account of the opinion of the Board, shall assess whether adherence to those codes of practice is adequate to ensure compliance with the obligations laid down in paragraphs 2 and 4 of this Article, in accordance with the procedure laid down in Article 56(6). If it deems the code of practice to be inadequate, the Commission may adopt an implementing act specifying common rules for the implementation of those obligations in accordance with the examination procedure laid down in Article 98(2).” The final sentence carries the fallback power over from the original paragraph unchanged.⁸

The amendment makes three changes, and the amending regulation states the reason for the principal one. First, the competent actor is now the Commission rather than the AI Office, which aligns the operative text with Recital 135, which had always spoken of the Commission.¹⁸ Second, the subject-matter of the codes now expressly covers marking alongside detection and labelling: the original wording spoke of the obligations “regarding the detection and labelling of artificially generated or manipulated content,” the new one of those “regarding the detection, marking and labelling of artificially generated or manipulated content,” closing the gap between the original paragraph 7 and the machine-readable marking duty in paragraph 2 that such codes are meant to serve.¹⁹ Third, approval of a code by implementing act gives way to an assessment of adequacy, which the Commission makes taking utmost account of the opinion of the Board — the European Artificial Intelligence Board (Article 65) — and which, under the recast Article 56(6), it publishes. Recital 41 of the amending regulation explains why the implementing act was dropped: because the codes of practice under Articles 50(7) and 56(6) “have limited legal effect, and in particular do not grant a presumption of conformity,” it “is not strictly necessary for these codes to be approved by an implementing act.”⁸ That reasoning is worth noting for what it confirms about the codes themselves: adherence is evidence of compliance and a supervisory convenience, not a legal presumption of it. Recital 135 makes the same point from the other direction, framing the codes as facilitative: “Without prejudice to the mandatory nature and full applicability of the transparency obligations, the Commission may also encourage and facilitate the drawing up of codes of practice at Union level (…).”¹⁸

The mechanism has already been used once. The Commission assessed the Code of Practice on Transparency of AI-Generated Content in Opinion C(2026) 4839 final of 8 July 2026, and the Board adopted its own adequacy conclusion the following day — that is, before the recast paragraph entered into force on 27 July 2026. Both bodies therefore acted under the original text: the Commission reasoned that its power to approve a code by implementing act implied a corresponding obligation to assess the code’s adequacy first, in accordance with Article 56(6), and both stated that the assessment, carried out in parallel under the existing rules, would remain valid after the amendment as long as their conclusions were consistent. They were. Each found that the Code adequately covers the obligations in Article 50(2), (4) and (5) and facilitates their effective implementation, while recognising that adherence to it does not constitute conclusive evidence of compliance.¹⁶

4. Relationship to other provisions

• Chapter III (high-risk systems) — Articles 6–49. Article 50(6) makes the transparency duties additive to, and unaffected by, the high-risk regime. Chapter III spans Articles 6–49; its Section 2 (Articles 8–15) sets the substantive requirements for high-risk systems. A high-risk system that also interacts with natural persons or generates content must meet both sets of duties.¹² The two regimes have now come apart in time. Regulation (EU) 2026/1744 deferred the application of Chapter III, Sections 1 to 3 (with the exception of Article 6(5)) to 2 December 2027 for systems classified as high-risk under Article 6(2) and Annex III, and to 2 August 2028 for those classified under Article 6(1) and Annex I, while leaving Article 50 applicable from 2 August 2026.⁸ Emotion recognition systems are the clearest illustration: on the Guidelines’ reading every permitted emotion recognition system is high-risk, yet its deployer owes the Article 50(3) notification duty from August 2026 and the Chapter III requirements only from December 2027.³

• Article 5 (prohibited practices). Article 50(3) and Article 5 are complementary, not overlapping: Article 5 prohibits emotion recognition in the workplace and education and certain biometric categorisation; Article 50(3) imposes a notification duty on the permitted residual uses. The same conduct is never simultaneously prohibited and merely subject to disclosure.

• General-purpose AI (Articles 51–56, Chapter V). Article 50(2) applies to general-purpose AI systems; the separate obligations for general-purpose AI models — including their transparency and documentation duties — are in Articles 53–55. The two regimes can stack along the value chain: an upstream model provider may implement marking at model level, “thereby facilitating fulfilment of this obligation by the downstream provider of the AI system” (Recital 133) — the downstream provider being the provider of an AI system (including a general-purpose AI system) that integrates an AI model (Art 3(68)). But the Article 50(2) duty itself remains a system-level obligation that binds that downstream provider, not the model provider; the Guidelines put the same point as an encouragement to model providers to implement measures at model level, particularly under Article 55(1)(b) where the model carries systemic risk.^4,3 Article 50(7) cross-refers to the Article 56(6) procedure, which Regulation (EU) 2026/1744 also recast. The change is one of instrument: under the old text the Commission could give a code legal standing by approving it in an implementing act; under the new one it assesses whether adherence to the code is adequate to ensure compliance and publishes that assessment. The code is thus no longer approved, but appraised — and the appraisal is a public statement of the Commission’s view, not a legal act conferring conformity.⁸

• Sanctions — Article 99(4)(g). Infringement of the Article 50 transparency obligations falls under Article 99(4), which caps administrative fines at up to EUR 15 000 000 or, for undertakings, up to 3% of total worldwide annual turnover for the preceding financial year, whichever is higher.²⁰ (This is the middle tier; the Article 5 prohibitions carry the higher EUR 35 000 000 / 7% ceiling, and supply of incorrect information the lower EUR 7 500 000 / 1% ceiling.) Two calibrations soften that ceiling. For Union institutions, bodies, offices and agencies it is EUR 750 000. And for SMEs, including start-ups, each fine is capped at the lower of the two figures rather than the higher (Article 99(6)). The Guidelines state the position for small mid-cap enterprises as it stood when their content was approved on 20 July 2026 — that the interests of SMCs and their economic viability are to be taken into account — but the point has since moved: Regulation (EU) 2026/1744 inserted a new Article 99(6a), in force from 27 July 2026, under which “in the case of SMCs, each fine referred to in paragraphs 4 and 5 shall be up to the percentages or amount referred therein, whichever is lower,” and amended Article 99(1) to require Member States to “take into account the interests of SMEs, including start-ups, and SMCs, and their economic viability when imposing penalties.” From the day Article 50 becomes applicable, therefore, SMEs and SMCs are alike entitled to the lower cap.^3,8 Supervision is shared: national market surveillance authorities designated by the Member States are the default, the AI Office is exclusively competent for AI systems built on a general-purpose AI model where the same provider supplies both, and for AI systems that constitute or are integrated into a very large online platform or very large online search engine designated under the DSA (Article 75(1)), and the European Data Protection Supervisor supervises Union institutions and bodies.^3,21 Enforcement is otherwise national; see section 6.

• Digital Services Act (Regulation (EU) 2022/2065). Recital 136 ties Article 50 to the DSA: marking and disclosure are “particularly relevant” to the duty of providers of very large online platforms or very large online search engines to identify and mitigate systemic risks from artificially generated content, “in particular the risk of the actual or foreseeable negative effects on democratic processes, civic discourse and electoral processes, including through disinformation.”²² Labelling under the AI Act is “without prejudice to” the DSA notice-and-action rules and “should not influence the assessment and the decision on the illegality of the specific content.”²² The Guidelines add a comparison with Article 35(1)(k) DSA, which — as they put it — “lists, among the possible risk-mitigation measures that providers of VLOPs and VLOSEs may adopt,” measures ensuring that an item of information “that appreciably resembles existing persons, objects, places or other entities or events and falsely appears to a person to be authentic or truthful is distinguishable through prominent markings.” It is thus one option among the mitigations available to those providers rather than an unconditional duty, and it is aimed at content answering that deep-fake-like description rather than at manipulated media at large. That duty is materially broader than Article 50(4) (it covers any technology, not only AI) and personally different (it binds VLOP and VLOSE providers, not deployers), so the two obligations are complementary rather than co-extensive.³ The relationship has since acquired a concrete addressee: on 31 August 2026 the Commission designated ChatGPT as a very large online search engine under the DSA — the first generative chatbot placed under that regime — giving the service four months, to January 2027, to assess and mitigate its systemic risks, while its provider already bears the Article 50(1) and (2) duties for the same system; and under Article 75(1) the designation also moves supervision of the system’s AI Act obligations to the AI Office.^23,21

• Data protection. Article 50(3) expressly preserves the GDPR (Regulation (EU) 2016/679), the EUDPR (Regulation (EU) 2018/1725) and the Law Enforcement Directive (Directive (EU) 2016/680); notification under Article 50 does not supply a legal basis for processing.¹²

• Compliance ≠ lawfulness. Recital 137 cautions that satisfying the transparency obligations does not make the use of the system or its output lawful under EU or national law, and is without prejudice to other transparency duties.²⁴ The Guidelines press the point inside the Regulation itself: a system may be fully compliant with Article 50 and still be prohibited under Article 5 or subject to the high-risk regime under Article 6. They also confirm the outer edges of scope. Research and development is excluded by two provisions with different triggers: Article 2(6) exempts systems developed and put into service for the sole purpose of scientific research and development, so what ends the exemption is a change in the purpose for which the system is used; Article 2(8) exempts research, testing and development activity prior to placing on the market or putting into service, so what ends that exemption is the act of placing on the market or putting into service, whatever the purpose. Testing in real-world conditions is not covered by the second exclusion. And the exclusion of AI systems released under free and open-source licences in Article 2(12) carries its own proviso: the Regulation does not apply to them “unless they are placed on the market or put into service as high-risk AI systems or as an AI system that falls under Article 5 or 50.” Providers and deployers of open-source AI systems within the scope of Article 50 must therefore meet their transparency obligations; what stays outside Article 50, on the Guidelines’ reading, is only providers of open-source components — software, data, models, tools, services or processes for customisation and integration — that do not in themselves, in whole or in part, constitute an AI system.^10,3

5. Official guidance (with status)

• Commission Guidelines on the implementation of the transparency obligations for certain AI systems under Article 50 of the AI Act. The Guidelines exist in final form as the Annex to Communication C(2026) 5054 final of 20 July 2026, by which the Commission approved their content; they run to 155 numbered paragraphs. Their legal basis is Article 96(1)(d) of the AI Act, and they were prepared after the targeted consultation on the draft dated 7 May 2026 and announced on 8 May (closed 3 June 2026), and after input from the Member States and the AI Board. Two qualifications go with them. On status, the Communication states that the Guidelines “will be formally adopted by the Commission at a later date, when all language versions are available. It is only from that moment that these Guidelines will be applicable.” On authority, the Guidelines say of themselves: “These Guidelines are non-binding. Any authoritative interpretation of the AI Act may ultimately only be given by the Court of Justice of the European Union.” They describe their own content as “a first interpretation” that the Commission will revise in the light of practice and case-law. As at 15 September 2026 the Commission’s library still offered only the English Annex and the Communication, and the Board noted in its July assessment of the Code that the Guidelines were at that time available only in draft. Status: final text, content approved 20 July 2026 — formal adoption and applicability still pending the language versions as at mid-September 2026; non-binding.^3,16

• Code of Practice on Transparency of AI-Generated Content (final version published on 10 June 2026, after the closing plenary of a multi-stakeholder process facilitated by the AI Office, in which the drafting was done by independent experts it had appointed and ran through three drafts between September 2025 and June 2026). It is a voluntary tool for providers and deployers of generative AI systems to demonstrate compliance with Article 50(2) and (4): Section 1 addresses provider marking and detection under 50(2); Section 2 addresses deployer labelling of deepfakes and public-interest text under 50(4). It prescribes no particular marking technique, tracking the statutory feasibility standard. The Guidelines describe the Code as “the only Union-wide recognised practical framework” for providers and deployers of generative AI systems to demonstrate compliance with those obligations, “regardless of their place of establishment, operation, and competent market surveillance authorities,” and adherence to an adequate code as “a straightforward, predictable, and legally certain way of demonstrating compliance.”³ The reduced administrative burden signatories are said to gain is how the Commission presents the instrument rather than a legal effect it confers.¹³ Compliance may equally be demonstrated “through adequate alternative means”; non-signatories then have to satisfy the market-surveillance authorities of the adequacy of those measures individually. The Guidelines describe the supervisory consequence concretely: for a signatory, oversight concentrates on whether it is in fact adhering to the Code, while a non-signatory can expect more detailed requests for information about how it achieves the same result.³ Signature is open to providers and deployers, and — although the Article 50(2) duty binds only providers of AI systems — Section 1 may also be signed by providers of marking and detection solutions and by providers of AI models. By the end of July 2026 about 190 organisations had signed, among them Aleph Alpha, Anthropic, Black Forest Labs, Cohere, Google, Meta, Microsoft, Mistral, OpenAI and Synthesia for Section 1 and Bulgari, Fastweb, Getty Images, Iberdrola, Lenovo and Lufthansa for Section 2; about half of the signatories are small and recently founded companies, and the Commission’s running list stood at 95 Section 1 and 192 Section 2 signatories at its 17 September 2026 update. The Code remains open for signature, and from September 2026 signatories are to work in two task forces on sharing industry practice, feedback on implementation and advancing the state of the art. Status: final — assessed as adequate by the Commission (Opinion C(2026) 4839 final, 8 July 2026) and by the Board (9 July 2026); voluntary, not binding.^13,16

The two instruments are distinct and complementary: the Guidelines interpret the scope of the duties laid down by the Regulation, whereas the Code is a vehicle for discharging them — the Code does not replace the AI Act or the Guidelines. Both are soft-law: they assist interpretation but cannot expand or contract the duties laid down by the Regulation, which apply in full from 2 August 2026 irrespective of when the Guidelines are formally adopted. The recast of the code-of-practice mechanics in Articles 50(7) and 56(6) is now in force and gives the Code a defined place in that architecture: the Commission assesses whether adherence to it is adequate to ensure compliance with Article 50(2) and (4), and publishes that assessment. That assessment has now been made (see section 3 on paragraph 7): the Commission and the Board each concluded that the Code adequately covers Article 50(2), (4) and (5) and facilitates their effective implementation, the Board adding that market surveillance authorities should treat it as the only EU-wide practical tool assessed as adequate as of July 2026, without prejudice to their power to investigate whether a provider or deployer has actually implemented its commitments. What follows from a positive assessment is evidential, not automatic: a provider that adheres to the Code has a recognised way of showing that it has done what the paragraph requires, but it does not thereby become compliant as a matter of law — as Recital 41 of the amending regulation puts it, the codes “do not grant a presumption of conformity,” and both assessing bodies said the same in their own words, recognising that adherence to the Code does not constitute conclusive evidence of compliance with the obligations.^3,13,8,16

6. National specifics

Czech Republic. The Czech adaptation bill (návrh zákona o umělé inteligenci, prepared by the Ministry of Industry and Trade, MPO) remains a draft, now past its inter-ministerial comment procedure (launched 25 September 2025; comments due 23 October 2025). Following the resolution of comments, the bill was entered into the Government Office's register on 26 June 2026 (VeKLEP status “zařazeno do evidence”); as at mid-September 2026 no Government resolution and no parliamentary print had been recorded, so it is not law and was not in force when Article 50 became applicable on 2 August 2026. It is purely an adaptation statute: it does not create substantive transparency or deepfake rules beyond the Regulation, but designates authorities and offences.^25,26

• Supervisory authority for Article 50: the bill designates three market-surveillance authorities under Article 70(1) — the Czech Telecommunication Office (Český telekomunikační úřad, ČTÚ), the Czech National Bank for entities it supervises, and the Office for Personal Data Protection for the high-risk systems in Article 74(8). ČTÚ is the residual authority for everything else and the single point of contact under Article 70(2) (§ 3).²⁵ In the meantime ČTÚ has presented itself publicly as one of the Czech market surveillance authorities for these duties: in a press release of 31 July 2026 it recalled that Article 50 binds providers and deployers alike, recommended a three-step approach — determine one’s own role, identify the specific obligations, set up the processes and measures to meet them — and stated that supervision lies with the market surveillance authorities of the Member State, which in the Czech Republic include ČTÚ, with the AI Office enforcing Article 50 in certain cases in their stead. That is a statement of the authority’s understanding of its role, not a legal designation, which only the adaptation act can supply.²⁷
• Offence and fine: the bill makes breach of the Article 50 obligations an administrative offence (přestupek) twice over — once for the provider (§ 27(1)(b)) and once for the deployer (§ 27(6)(c)) — and sets the fine itself rather than merely cross-referring: up to EUR 15 000 000 or, for an undertaking, 3% of total worldwide annual turnover, whichever is higher (§ 27(7)(b)), mirroring Article 99(4). Where the amount is derived from Article 99(3)–(7), it is converted into Czech koruna at the ČNB exchange rate for the day before the decision and rounded up (§ 31(4)); the limitation period is 5 years, and liability lapses at the latest 8 years after the offence (§ 31(2)).²⁵ For less serious breaches the authority may first warn and set a remediation period of at least 15 days; if the breach is remedied, no offence proceedings are opened (§ 25). Which authority hears the offence follows § 30(2): the Office for Personal Data Protection or the Czech National Bank within their remits, ČTÚ otherwise.²⁵
• Status caveat: section numbers and wording may still change in the remaining legislative process — government approval, parliamentary reading, promulgation; everything here should be cited as a draft.²⁵
• A separate criminal layer is already in force. Independently of the (still draft) adaptation bill, the Czech Republic has enacted a criminal-law response to harmful deepfakes through Act No. 270/2025 Sb. (an amendment to the Criminal Code), in force 1 January 2026. It adds a new § 191a, “Zneužití identity k výrobě pornografie a její šíření” (misuse of identity for the production and dissemination of pornography), targeting non-consensual AI-generated sexual imagery (“deep nudes”), with imprisonment of up to 2 years in the basic offence and 6 months–3 years / 1–5 years in the qualified and aggravated forms; and it extends § 181(2) (infringement of another’s rights) to non-pornographic deepfakes — reaching, in the terms of the Czech provision, anyone who with intent to cause serious harm to another’s rights produces a work that without authorisation depicts, captures or otherwise uses the likeness of another or an expression of their personal nature, and that appears to be genuine although the perpetrator knows it is not, or who makes such a work publicly accessible, brokers it, puts it into circulation, sells it or otherwise procures it for another. As in Italy, this criminal track is distinct from, and additional to, the administrative transparency duties of Article 50.²⁸

Italy. Italy went further and earlier. Law No. 132/2025 (Legge 23 September 2025, n. 132), in force 10 October 2025, is the first comprehensive national AI framework in an EU Member State; its Article 2 adopts the AI Act's definitions.²⁹ Most relevant to Article 50, it adds a criminal layer that the AI Act itself does not contain: Article 26(1)(c) inserts a new Article 612-quater into the Italian Criminal Code (“Unlawful dissemination of content generated or altered with artificial intelligence systems”), punishing anyone who causes unjust harm to a person by transferring, publishing or otherwise disseminating, without that person's consent, images, video or voice recordings falsified or altered by means of AI systems and apt to mislead as to their genuineness, with imprisonment of one to five years. The offence is prosecuted on the victim's complaint, and ex officio only where it is connected to another offence prosecuted ex officio or committed against a person incapable by reason of age or infirmity or against a public authority on account of its functions. Article 26(1)(a) adds a general aggravating circumstance for offences committed through the use of AI systems (Article 61(11-decies)) — not automatically, but where those systems, by their nature or manner of use, constituted an insidious means, hindered public or private defence, or aggravated the consequences of the offence.²⁹ This illustrates a structural point: Article 50(4)'s administrative labelling duty on deployers coexists with, and is distinct from, national criminal prohibitions on harmful deepfake dissemination.²⁹ The institutional side followed on 4 August 2026, when the Council of Ministers approved two legislative decrees under the delegation in Article 24 of Law 132/2025: the second designates the Agenzia per la cybersicurezza nazionale (ACN) as the principal market surveillance authority and single point of contact — and thus, once in force, the supervisor of Article 50 in Italy — and the Agenzia per l’Italia Digitale (AgID) as notifying authority, leaving specific powers with the financial regulators and the data protection authority, and the text was aligned with Regulation (EU) 2026/1744. Publication in the Gazzetta Ufficiale had not been recorded by mid-September 2026, so the decrees are cited here as approved, not yet in force.³⁰

Germany. The Act implementing Regulation (EU) 2024/1689, whose Article 1 is the KI-Marktüberwachungs-und-Innovationsförderungs-Gesetz (KI-MIG), was passed by the Bundestag on 11 June 2026 and cleared the Bundesrat on 10 July 2026; dated 22 July and promulgated on 28 July 2026 in the Federal Law Gazette (BGBl. I 2026 No. 223), it entered into force on 29 July 2026 — four days before Article 50 became applicable. It makes the Bundesnetzagentur the market surveillance authority for the Regulation wherever the Act provides nothing else (§ 2(1)) and the single point of contact under Article 70(2) (§ 6(1)), with BaFin supervising systems directly connected with regulated financial activity and Land authorities supervising the systems of Land public bodies; infringements under Article 99(3)–(5) are pursued under the German administrative-offences regime (§ 16), and no fines are imposed on public authorities (§ 17(2)). For Article 50 the practical point is that in Germany the supervisory and penalty machinery existed on the day the duties began to apply.³¹

7. Practical impact and open questions

• The compliance calendar for Article 50 now runs across three dates, not one. The Digital Omnibus on AI completed its passage while this provision was still in its pre-application period: adopted by Parliament at first reading on 16 June 2026 (P10_TA(2026)0198), approved by the Council without changes on 29 June 2026 (PE-CONS 30/26), signed on 8 July and published as Regulation (EU) 2026/1744 in OJ L 2026/1744 of 24 July 2026, entering into force on 27 July 2026 — the third day after publication, and less than a week before Article 50 itself starts to apply. The practical calendar is therefore: 27 July 2026, the recast Article 50(7) and the new Article 111(4) enter into force; 2 August 2026, Article 50 becomes applicable, with the duties in paragraphs (1), (3) and (4) biting in full and no transition; 2 December 2026, the end of the four-month transitional period in Article 111(4) for providers of generative AI systems placed on the market before 2 August 2026, who until then need not have completed the machine-readable marking required by Article 50(2). Recital 38 of the amending regulation explains the transition as time to adapt “without disrupting the market” — the Commission had originally proposed a longer runway, to 2 February 2027. The substance of the four transparency duties is untouched. The package's other headline change relevant to synthetic media — new prohibitions on non-consensual intimate-imagery and child-sexual-abuse applications, applicable from 2 December 2026 — operates through Article 5, not Article 50.^8,32,33 For signatories of the Code a fourth date exists, though one that binds only as a voluntary commitment, not by statute: 2 February 2027, by which Measure 3.4 of the Code requires a minimum level of interoperability for watermark-detection mechanisms — a staged commitment that both the Commission and the Board have said they will monitor: the Commission reserves the right to consider appropriate measures if progress is lacking, and the Board to recommend a targeted review of the Code on this point or the adoption of common rules under Article 50(7). The Commission has also said the AI Office will consider facilitating formal updates to the Code at least every two years.¹⁶ outlook

• The 50(2) and 50(4) duties are separate and do not substitute for one another. They are two different obligations, on two different actors, serving two different audiences. Article 50(2) requires the provider to embed a machine-readable mark in generated output — a signal aimed at detection systems and downstream tooling. Article 50(4) requires the deployer of a deepfake to make a disclosure the human audience can recognise. Discharging one does not discharge the other: a provider’s invisible watermark satisfies 50(2) but is not, by itself, the audience-facing deepfake disclosure that 50(4) demands of the deployer. A deployer who publishes watermarked output therefore cannot assume it is already compliant with 50(4); it must still add a perceptible disclosure of the artificial origin. analysis

• Whether disclosure achieves its aim is empirically contested. The premise of Article 50 is that informing people reduces deception. The behavioural evidence is mixed. In two preregistered experiments (n≈4,976), labelling headlines “AI-generated” lowered both perceived accuracy and willingness to share them — and did so whether the headline was true or false and whether it was written by a human or by a machine; the effect was about three times smaller than that of a “false” label and appeared to be driven by readers taking “AI-generated” to mean fully automated, without human supervision (Altay & Gilardi, 2024).³⁴ That is a documented “AI-aversion” which can penalise harmless or beneficial AI content. Labelling AI authorship is nevertheless believed: in a further study 92% of participants accepted the authorship label they were shown, yet the labels made no significant difference to attitude change, accuracy judgments or sharing — disclosure of AI authorship increased transparency without blunting the content’s persuasive effect (Gallegos et al., 2025).³⁴ The same ambivalence appears on the interaction side: disclosing that a service agent is a chatbot has been found to reduce purchases, because customers judge the disclosed bot less knowledgeable and less empathetic (Luo et al., 2019), while a later customer-service experiment found no significant effect of disclosure at all (van der Goot et al., 2024).⁶ At the same time, humans cannot reliably detect deepfakes unaided (Diel et al., 2024; Köbis et al., 2021)⁵ or recognise machine-written self-presentations (Jakesch et al., 2022)⁶, which is precisely why a disclosure duty exists. The tension suggests that how disclosure is designed (Article 50(5)'s “clear and distinguishable” standard¹²) matters as much as whether it is made. These are background observations, not legal propositions. context

• The design of Article 50(2) reflects what detection can and cannot do. The paragraph asks providers to ensure that the outputs of their AI system “are marked in a machine-readable format and detectable” — marking is thus a property of the output from the moment it is generated, which on the Guidelines’ reading the system must have built in at the latest when it is placed on the market or put into service — rather than leaving verification to downstream tools that guess whether a given artefact is synthetic; the literature on those tools explains why.^12,3 Evaluations of post-hoc detectors for AI-generated text report that they are “neither accurate nor reliable,” that paraphrasing and obfuscation degrade them sharply, and that they perform markedly worse against newer models and mixed human–AI drafting (Weber-Wulff et al., 2023; Walters, 2023; Hadra et al., 2026).³⁵ Marking at source avoids that guessing game, but is not immune: recursive paraphrasing reduces detection rates even for watermarked text, and spoofing can make human writing register as machine-written (Sadasivan et al., 2023)³⁵, while regeneration attacks can strip pixel-level invisible watermarks from images (Zhao et al., 2024)¹⁵. This is the empirical setting for the statutory hedge “as far as this is technically feasible” and for the Guidelines' insistence on interoperable, publicly available standards rather than a prescribed technique. These are contextual observations, not legal propositions. context

• The evidence on disclosure has sharpened around three points that map onto the statute’s design. First, in journalism — the field of the second limb of 50(4) — a systematic review of 47 studies found no consistent “AI penalty” from provenance cues and mostly null or conditional effects of disclosure, with scepticism concentrated where a disclosure implied full automation without accountability or oversight information (Licenji & Hoxha, 2026); a meta-analysis of 31 studies still finds a small but significant credibility penalty for AI-labelled news (Kim et al., 2026), and experiments show that rich, process-level disclosure naming editorial verification and sources mitigates it (Haq, 2025) and that an explained policy on AI use can raise trust (Waddell, 2026). That pattern sits comfortably with the statutory choice to lift the duty where human review or editorial control and editorial responsibility exist.³⁶ Second, for deep fakes a label reliably changes what viewers believe about origin but not necessarily what the content does to them: in three preregistered experiments most participants continued to rely on a deepfake “confession” even after being told, and believing, that it was fake (Clark et al., 2026); label design rather than mere presence determines trust in the label (Gamage et al., 2025); and generic warnings about deepfakes lead people to disbelieve genuine footage too (Ternovski et al., 2021). This is why the specificity and placement of the disclosure — the Code’s icon and placement specifications, and Article 50(5)’s “clear and distinguishable” standard¹² — matter as much as its existence.³⁷ Third, provenance metadata as a stand-alone layer often does not survive distribution: social networks systematically transcode uploaded files, which breaks C2PA (Ossowski et al., 2026); Twitter in particular strips C2PA on upload (Zewde et al., 2026, preprint); and the current C2PA specifications do not meet their own security goals (Golaszewski et al., 2026, preprint). That is the technical background to the Code’s default of at least two marking layers, its optional forensic-detection measure for cases where metadata has been stripped and watermarks removed, and the Board’s undertaking to monitor label retention on platforms. These are contextual observations, not legal propositions.^38,16 context

• The interpretive gaps have narrowed, and the remaining ones have moved. A year ago the open questions were the statutory thresholds themselves: what it means to “substantially alter the input data (…) or the semantics thereof” for the 50(2) editing carve-out; where “evidently artistic (…) satirical” content begins for 50(4); what counts as a matter of “public interest” and as adequate “human review or editorial control”; and what disclosure is “appropriate” so as not to “hamper the display or enjoyment of the work.”¹² The July 2026 Guidelines answer each of these at the level of a first interpretation, and the Code supplies a common technical footing for signatories. What is left is of a different kind. Two examples. The Guidelines' narrowed scenarios for 50(2) — closed product environments, industrial and business-to-business applications, ephemeral real-time content — have no express counterpart in the text of the paragraph, which qualifies the duty only by technical feasibility and by the assistive-editing carve-out; whether they are a reading of that feasibility standard or an addition to it is a question the Guidelines cannot settle for themselves, since they “are non-binding” and the Court of Justice has the last word.³ And the treatment of dissemination-only intermediaries leaves an addressee gap: where a platform is not a deployer and the original deployer’s labelling does not survive re-upload, nothing in Article 50 compels the intermediary to restore it — the Guidelines can only encourage preservation, and the DSA duty that does bind platforms reaches only very large platforms and search engines.³ The Board has flagged the same gap from the supervisory side, undertaking to monitor to what extent labels are in practice retained during further processing, platform distribution and third-party applications, and the Code addresses it through commitments of its signatories rather than duties on intermediaries: providers undertake to work with the value chain so that marks are not removed, and deployers to cooperate on a best-effort basis with the services through which they disseminate content so that the label travels with it.¹⁶ analysis

• Enforcement architecture is still settling, and a criminal layer is emerging alongside it. Three things are happening at once. First, the Union-level allocation is settled: national market surveillance authorities as the default, the AI Office where the system is built on a general-purpose AI model from the same provider or constitutes or is integrated into a designated very large online platform or search engine, and the EDPS for Union institutions.^3,21 Second, the national layer beneath it is uneven rather than uniformly late. Germany had the KI-MIG in force on 29 July 2026, with the Bundesnetzagentur in place as market surveillance authority before the duties began to apply; Italy’s Council of Ministers approved the implementing decrees designating ACN on 4 August 2026, two days after; and the Czech adaptation bill — which is what would designate the ČTÚ and create the domestic offence for breach of Article 50 — was, as at mid-September 2026, still only a draft in the Government Office’s register (since 26 June 2026), so in the Czech Republic the duties have applied for six weeks without the national supervisory and penalty machinery, with ČTÚ acting on its own announced understanding of its future role. Third, several Member States are adding national criminal offences for harmful deepfakes that sit on top of the AI Act's purely administrative regime: Italy through Law 132/2025 (new Art 612-quater of the Criminal Code), and the Czech Republic through Act No. 270/2025 Sb. (new § 191a and extended § 181(2)), in force since 1 January 2026 — the criminal track, unlike the administrative one, is already operative. For an organisation operating cross-border the result is a composite picture: one EU-wide administrative duty under Article 50, layered over national supervisory regimes that come into force on different dates and, in a growing number of states, national criminal prohibitions on the most harmful deepfake conduct.^{25,26,28,29,31,30,27} outlook