(last updated: 28.10.2026)
Since 2 August 2026, the transparency obligations of the EU AI Act apply. The provision has attracted more public commentary than most, and a good deal of it has been inaccurate.
Part of the attention comes from what model providers have been doing. In July 2026, around 190 signatories adopted the Code of Practice on Transparency of AI-Generated Content, and the marking of AI output followed over the summer. Anthropic set out its own approach in an explanation published on 14 August 2026, describing how Claude embeds an imperceptible watermark in generated text and attaches signed C2PA provenance metadata to supported image files. The marking applies worldwide rather than only in the EU, because the company has no reliable way to scope it by region. Other providers had already moved on images, OpenAI having joined the C2PA coalition in May 2026.
Alongside this, a different claim has been circulating, which we have seen repeated on social media. It holds that any organisation using AI to produce content must now add an AI disclosure. That claim is wrong, and this short note explains why.
Which date applies to which obligation
Article 50 became applicable on 2 August 2026. That date holds for AI systems placed on the market or put into service on or after it, and it holds for all four paragraphs of the provision.
The AI Omnibus, published as Regulation (EU) 2026/1744 and in force since 27 July 2026, added one transitional rule. Generative AI systems that were already available on the market before 2 August 2026 have until 2 December 2026 to comply with the machine-readable marking obligation of Article 50(2). The disclosure duty for interactive systems under Article 50(1), the information duty for emotion recognition and biometric categorisation under Article 50(3), and the labelling duty for deep fakes and public-interest text under Article 50(4) all applied from 2 August without any transition.
Content that was generated and made available before 2 August 2026 does not have to be marked or labelled retroactively, although the Commission encourages it where technically feasible.
Two practical consequences follow. Organisations need to know when each of their systems went on the market, because the applicable deadline depends on it. And organisations relying on the December window for a supplier tool should check that the supplier is actually working towards it.
Providers and deployers do not carry the same duties
Article 50 sets out four obligations. Two of them fall on providers, meaning those who develop an AI system and place it on the EU market or puts it into service under their own name. Two fall on deployers, meaning those who use a system under their own authority in a professional capacity.
Article 50 sets out four obligations, split between two roles:
Two of them fall on provider, meaning the party that develops an AI system and places it on the EU market or puts it into service under its own name:
- Under Article 50(1), a system designed to interact directly with a person must be built so that the person knows they are dealing with a machine, unless that is obvious from the context.
- Under Article 50(2), a system that generates synthetic audio, image, video or text must mark its output in a machine-readable format and make it detectable, using solutions that are effective, interoperable, robust and reliable. What the provision requires is a machine-readable mark and not a human-readable one, meaning a signal that a detection tool can read rather than a notice appearing on the page for the reader. This is precisely what the watermarking announcements of recent weeks implement.
The other two fall on deployers, meaning those who use a system under their own authority in a professional capacity. Both consist of informing the persons concerned:
- Under Article 50(3), an organisation operating an emotion recognition or biometric categorisation system must inform the people exposed to it.
- Under Article 50(4), an organisation using AI to produce a deep fake, or to generate text published to inform the public on matters of public interest, must disclose the artificial origin of that content (but with notable exceptions we present below).
One point deserves attention for organisations that put their own brand on a third-party tool. A bank or a retailer that deploys a white-label chatbot under its own name becomes the provider of that system for the purposes of the AI Act, and inherits the provider obligations of Articles 50(1) and 50(2). The same applies to an organisation that takes a general-purpose system and deploys it for a high-risk purpose.
| Provision | Who is bound | What has to be done | Main exception |
|---|---|---|---|
| Art. 50(1) | Provider | Design an interactive system so the person knows they are talking to an AI | The artificial nature of the interaction is obvious to a reasonably well-informed user |
| Art. 50(2) | Provider | Mark synthetic audio, image, video and text in a machine-readable format and make it detectable | Assistive function for standard editing, or no substantial alteration of the input |
| Art. 50(3) | Deployer | Inform people exposed to an emotion recognition or biometric categorisation system | Use authorised by law to detect or prevent criminal offences |
| Art. 50(4) | Deployer | Disclose deep fakes, and AI-generated text published to inform the public on matters of public interest | Human review with editorial responsibility for text, attenuated regime for artistic or satirical works |
Article 50(5) applies across all four obligations and requires that the information be clear, distinguishable and provided at the latest at the time of the first interaction or exposure.
Fines reach EUR 15 million or 3% of worldwide annual turnover, whichever is higher.
What Article 50 does not require
Article 50 does not oblige every organisation to disclose that it used AI to draft a document. The idea has spread because two distinct things have been merged in the retelling.
The marking obligation of Article 50(2) is a technical requirement that rests on the provider of the generative system. The signal it produces is addressed to detection tools rather than to readers, which means it says nothing about whether the organisation using that system has to tell anyone anything.
The disclosure obligation of Article 50(4) does rest on the deployer, but its scope is narrow. For text it is confined to publications intended to inform the public on matters of public interest, meaning topics that merit public debate or scrutiny. A newspaper article on a municipal decision falls within it, as do an academic paper on the effects of a treatment and a corporate report published for investors, whereas an internal memorandum, a commercial proposal, an email to a counterparty or an advertisement making no public-interest claim all fall outside it. Even where a publication is within scope, the obligation lifts if the text has been through genuine human review or editorial control and a person or entity holds editorial responsibility for it, though superficial checks will not do, and a spellcheck or a cursory sign-off leaves the duty in place.
A firm that uses AI to prepare a client memorandum therefore owes nothing under Article 50, while a firm that publishes AI-generated commentary on a regulatory development without meaningful review is in a different position.
Finally, we stress that other rules may require disclosure where Article 50 does not, among them consumer protection law, professional and deontological standards and contractual commitments.
What we suggest doing now
- Map your systems, distinguishing those you provide from those you deploy, and record when each went on the market.
- Check that every customer-facing conversational tool discloses its artificial nature at the first exchange, in the interface rather than in the small print.
- Identify any use of emotion recognition or biometric categorisation, which triggers a duty even where the system is not classified as high risk.
- Review your publication workflow for content of public interest, and document the editorial review that relieves the labelling duty.
- Ask your suppliers, in writing, how they comply with Article 50(2) and by which date.
A note for Swiss organisations
Switzerland has no equivalent legislation (yet). The Federal Council signed the Council of Europe Framework Convention on AI in March 2025 and has instructed the Federal Department of Justice and Police to prepare a draft bill by the end of 2026, which means nothing will be in force before 2028 at the earliest.
That does not put Swiss organisations outside Article 50. The AI Act reaches providers and deployers established outside the Union wherever the output of the system is used within it. A Swiss company with European customers, a European subsidiary or a European audience should assume it is concerned and check which of the four obligations applies to it.
Update, 28 October 2026. Claude has recently updated its guidance on watermarking. It will now roll out the functionality to earlier models as well. Claude also provides a tool to detect Content Credentials, such as an indication that an image has been AI generated. Watermark detection is only available in private preview for certain companies, which may request access through a specific form.
Sources: Regulation (EU) 2024/1689 (AI Act), art. 50; Regulation (EU) 2026/1744 (AI Omnibus), in force 27 July 2026; European Commission, draft Guidelines on the implementation of the transparency obligations under Article 50, published for stakeholder consultation; Code of Practice on Transparency of AI-Generated Content, European AI Office, July 2026; Anthropic, How Claude's text watermark works, 14 August 2026; Anthropic, How Claude marks AI-generated content, Claude Support Center.
This note is general information and does not constitute legal advice. For an assessment of your own systems, please get in touch.