[{"data":1,"prerenderedAt":10},["ShallowReactive",2],{"article:en:article-50-of-the-ai-act-the-four-transparency-duties-since-2-august-2026-the-2-december-2026-transition-the-code-of-practice-and-the-eu-icon":3},{"locale":4,"slug":5,"title":6,"description":7,"published":8,"body":9},"en","article-50-of-the-ai-act-the-four-transparency-duties-since-2-august-2026-the-2-december-2026-transition-the-code-of-practice-and-the-eu-icon","Article 50 of the AI Act for a company that ships or uses generative AI: the four transparency duties in force since 2 August 2026, the 2 December 2026 transition, the code of practice and the EU icon","Article 50 is the AI Act obligation that reaches a company whether or not its system is high-risk: tell people they are talking to an AI, mark generated content so machines can detect it, disclose deep fakes and AI-written text on matters of public interest, inform people exposed to emotion recognition. It has applied since 2 August 2026, the Digital Omnibus left it unchanged and gave providers of generative systems already on the market until 2 December 2026 for the marking duty. The four paragraphs read in order, who is provider and who is deployer for each, the Commission's code of practice of 10 June 2026 with its two-layer marking and its AI icon, the fine, and the record an ISO 42001 system keeps.","2026-09-12","\nMost of the AI Act is about high-risk systems, and most companies will never provide one. Article 50 is different: it reaches any provider whose system talks to people or generates content, and any deployer that publishes what a system generated, from the general date of application, 2 August 2026, whatever the system's risk class. Regulation (EU) 2026\u002F1744, the Digital Omnibus on AI, left the article as adopted and added one transition in Article 111(4): providers of generative systems already on the market on 2 August 2026 comply with the marking duty by 2 December 2026. This article reads the four duties in order, says who owes which, summarises the Commission's Code of Practice on Transparency of AI-generated Content of 10 June 2026 and the EU icon that comes with it, and names the record behind each duty, from the two Regulations on CELLAR and the Commission's pages on 12 September 2026. It is not legal advice.\n\n## The four duties, and who owes them\n\n**50(1), providers: say it is an AI.** A provider ensures that a system intended to interact directly with natural persons is designed so that those persons are informed they are interacting with an AI system, unless that is obvious to a reasonably well-informed, observant and circumspect person in the circumstances. A customer-service chatbot, a voice assistant, an AI that answers the support inbox: the provider builds the disclosure in. Systems authorised by law for criminal-law purposes are excepted unless the public uses them to report an offence.\n\n**50(2), providers: mark the output.** A provider of a system, general-purpose systems included, that generates synthetic audio, image, video or text ensures the outputs are marked in a machine-readable format and detectable as artificially generated or manipulated, with technical solutions that are effective, interoperable, robust and reliable as far as technically feasible, taking the type of content, the cost and the state of the art into account. It does not apply to the extent the system performs an assistive function for standard editing or does not substantially alter the deployer's input or its semantics. This is the duty the code of practice spends most of its pages on.\n\n**50(3), deployers: emotion recognition and biometric categorisation.** A deployer of an emotion recognition system or a biometric categorisation system informs the natural persons exposed to it of its operation and processes their personal data under the GDPR and its sister acts. Article 5 already prohibits emotion recognition at the workplace and in education except for medical or safety reasons; what is left is disclosed.\n\n**50(4), deployers: disclose deep fakes and public-interest text.** A deployer of a system that generates or manipulates image, audio or video content constituting a deep fake, defined in Article 3(60) as content resembling existing persons, objects, places, entities or events that would falsely appear authentic or truthful, discloses that the content has been artificially generated or manipulated. Where the content is part of an evidently artistic, creative, satirical, fictional or analogous work, the duty is limited to disclosing the existence of generated content in a way that does not hamper the display or enjoyment of the work. A deployer that publishes AI-generated or manipulated text with the purpose of informing the public on matters of public interest discloses that too, unless the text has undergone human review or editorial control and a natural or legal person holds editorial responsibility for it.\n\n**50(5), the manner.** The information under the four paragraphs is given to the natural persons concerned in a clear and distinguishable manner, at the latest at the time of the first interaction or exposure, and conforms to the applicable accessibility requirements. 50(6) keeps Chapter III and other transparency laws untouched, and 50(7) is the legal basis for the code of practice.\n\nTwo readings follow from the four paragraphs. First, the same company is often both: a SaaS that ships a chatbot which drafts documents is the provider under 50(1) and 50(2), and when its marketing team publishes an AI-written explainer on a matter of public interest without editorial review, the deployer under 50(4). Second, internal use is largely outside: a draft nobody outside the company is exposed to is neither a deep fake shown to natural persons nor a publication, though the provider of the tool still marks the output under 50(2).\n\n## The code of practice, and the icon\n\nArticle 50(7) asked the AI Office to facilitate codes of practice for the detection and labelling duties. The Commission published the final Code of Practice on Transparency of AI-generated Content on 10 June 2026, drawn up by independent chairs and working groups in a multi-stakeholder process, and reports that about 190 companies and organisations had signed it by the end of July 2026. The Commission and the AI Board have found it an adequate voluntary tool to demonstrate compliance; a provider or deployer that complies by other means has to show a market surveillance authority, individually, that its means are adequate. The code has two sections.\n\nSection 1, for providers under 50(2) and 50(5), is built on a multi-layered marking: as long as no single technique meets all four requirements of Article 50(2) by itself, signatories mark outputs with at least two layers, digitally signed and time-stamped metadata where the format can carry it, and an imperceptible watermark, applied to free-form text longer than 200 tokens as well, with a single layer accepted for free-form text and for closed, embedded products whose output cannot leave the product. Signatories run a detection mechanism for their own markings, make best efforts against removal, meet effectiveness, reliability, robustness and interoperability, and keep a compliance process with testing, training and cooperation with market surveillance authorities. Upstream model providers are encouraged to watermark at the model level so that downstream system providers inherit it.\n\nSection 2, for deployers under 50(4) and 50(5), is built on a label: an icon whose main element is the capitalised acronym \"AI\", supplemented where possible by a second layer saying \"generated\" or \"modified\", placed where nothing overlays it (the top right corner of an image or video), shown at the beginning of a video and again after interruptions, near the headline of a published text, and as a short audible disclaimer at the start of audio-only content, with accessibility for people who cannot see or hear the label. The EU icon is free to use, in four variations (black, white, and each at 50% transparency) and three forms (basic, fully AI-generated, partially AI-modified); using it is optional, the labelling duty is not, and the icon alone does not establish compliance. Deployers also commit to an internal compliance process, awareness measures, and the two special regimes for artistic works and editorially controlled text.\n\nThe code is complemented by the Commission's guidelines on the scope of Article 50, which define the interactive systems, synthetic content, deep fakes and public-interest text the article means, with examples of what is in and out.\n\n## The dates\n\nArticle 50 applies from 2 August 2026, the general date of application in Article 113. The Omnibus added Article 111(4): providers of AI systems, general-purpose systems included, generating synthetic audio, image, video or text that were placed on the market before 2 August 2026 take the necessary steps to comply with Article 50(2) by 2 December 2026, a transition of four months for the marking duty on systems already out there. The other three duties have no transition: a chatbot in production on 2 August 2026 already had to say it is an AI, and a deep fake published that day already had to be labelled.\n\n## The fine\n\nArticle 99(4)(g) puts the transparency obligations for providers and deployers under Article 50 in the list carrying administrative fines of up to 15 000 000 euros or, for an undertaking, 3% of total worldwide annual turnover, whichever is higher; for SMEs, and since the amendment for small mid-caps, whichever is lower. It is the same ceiling as the provider and deployer duties for high-risk systems, and it applies to a company with no high-risk system at all.\n\n## The record\n\nArticle 50 is a set of design and publishing decisions, and each of them is something an AI management system already documents. The table is StandardOS's reading of where ISO\u002FIEC 42001 keeps each; the Regulation names no standard.\n\n| Article 50 | What is kept | ISO 42001 |\n|---|---|---|\n| 50(1) | The interaction disclosure as designed, and the reasoning where it is judged obvious | A.6.2.2, A.8.2 |\n| 50(2) | The marking layers per output type, the detection mechanism, the test results | A.6.2.4, A.6.2.7, A.6.2.6 |\n| 50(3) | The notice to persons exposed, and the data protection basis | A.8.2, A.5.3 |\n| 50(4) | The labelling rule for published content, and who holds editorial responsibility | A.9.2, A.9.3, A.2.2 |\n| 50(5) | The accessibility check of every disclosure | A.8.2, A.6.2.4 |\n| 111(4) | The register of generative systems on the market before the general date, with the date each was marked | A.4.2, A.6.2.5 |\n\n## A checklist for the week\n\nList every system you provide that talks to people or generates content, and every system you deploy that publishes what it generated; for each, mark which paragraph reaches it and whether you are the provider or the deployer. Check that each interactive system says it is an AI at the first interaction. Ask your generative provider, or yourself if you are it, which two marking layers the outputs carry and how they are detected; if the system was on the market before 2 August 2026, the deadline is 2 December 2026. Put the label on every deep fake and every AI-written public-interest text that goes out without editorial responsibility, or put a named editor behind the text. Decide whether to sign the code; if not, write down why your means are equivalent. File the lot with the system's documentation, dated.\n",1789383984049]