Article 50 of the EU AI Act (Regulation (EU) 2024/1689) sets the transparency duties for AI systems that interact with people or generate synthetic content. It applies from 2 August 2026, it was not postponed by the Digital Omnibus, and breaching it carries a fine of up to €15 million or 3% of total worldwide annual turnover, whichever is higher.
On 20 July 2026, less than two weeks before that date, the European Commission adopted the final version of its guidelines on implementing Article 50. Those guidelines are what turn the provision from an abstract rule into a checklist you can test a product against. This article works through them, focusing on where companies actually get it wrong.
What Article 50 Actually Requires
Article 50 imposes four separate duties, each falling on a different actor in the chain. That distinction comes first, because a single company can carry all four at once.
- Article 50(1), interaction disclosure, a provider duty: a system that interacts directly with people must inform them they are dealing with an AI, unless that is obvious.
- Article 50(2), synthetic content marking, a provider duty: outputs of a system generating audio, image, video or text must be marked in a machine-readable format as artificially generated or manipulated.
- Article 50(3), emotion recognition and biometric categorisation, a deployer duty: whoever operates such a system must inform the people exposed to it that it is running.
- Article 50(4), deepfakes and public-interest text, a deployer duty: whoever publishes a deepfake, or AI-generated text intended to inform the public on a matter of public interest, must label it with a disclosure a person can perceive.
The final guidelines add an example that shows why this matters: an image-generating chatbot can be subject to Articles 50(1), 50(2) and 50(4) simultaneously, on one and the same system.
Provider or Deployer? Almost Always Both
A provider develops an AI system, or has one developed, and places it on the EU market or puts it into service under its own name or trademark. A deployer uses an AI system under its authority in the course of a professional activity.
The common mistake among Israeli SaaS companies is assuming they are only deployers. A company that builds a feature on top of a third-party model and sells it to customers under its own brand is a provider of its own system, full stop. At the same time, when that company uses AI to produce marketing content it publishes, it is a deployer. Both sets of duties apply, and they are not the same duties.
Two further clarifications in the final guidelines are worth knowing. First, the Act reaches a non-EU provider whose system output is used in the Union, but incidental, unforeseeable or unauthorised downstream use in the EU does not automatically trigger the obligations. Second, open-source AI systems remain in scope, while open-source components that do not constitute a complete system (libraries, data, models, tools) fall outside it.
Chatbots: What Counts as Valid Disclosure Under Article 50(1)
The duty arises when four cumulative conditions are met: the system qualifies as AI, it is designed for a genuine two-way exchange rather than mere data collection, the interaction is direct, and the counterparty is a natural person. Systems operating solely in the background, machine to machine, fall outside it.
The timing is fixed: notification must be given from the start of the first interaction, in a clear and distinguishable manner. A small banner at the foot of the page, or a line that appears only after the user has already asked a question, does not meet the standard.
The "obvious" exception is where most companies slip. The guidelines say it must be construed restrictively, assessed against an average person who is reasonably well-informed, circumspect and observant. Labelling the chat window "AI Assistant" is not sufficient on its own. The assumption that a user "surely understands" is something you have to evidence, not assert.
On AI agents the final guidelines expanded substantially. An agent must disclose not only its artificial nature but also the identity of the person directing it. In multi-agent architectures, where one agent contacts another on a user’s behalf, the requirement carries through the chain. If you are building agents that contact suppliers, customers or candidates, design this into the protocol rather than bolting it on at the end.
Marking Synthetic Content Under Article 50(2): What Is In, What Is Out
The provider must embed machine-readable marking that is effective, reliable, robust and interoperable, so that content can be detected as artificially generated or manipulated. This is a technical obligation, not a drafting one, and it applies to the provider even if a user never sees the mark.
The final guidelines significantly widened the list of outputs that need no marking:
- a short sequence of numbers, symbols or letters
- source code
- AI-generated translations
- outputs intended exclusively for machine-to-machine communication
- intermediate-stage outputs in a closed-loop production workflow (the final output still requires marking)
There is also an exception where the system performs an assistive function for standard editing. Polishing or rephrasing text a human wrote does not trigger the marking duty. Generating the text does.
The B2B carve-out is routinely read too broadly. The guidelines require three cumulative conditions: the output is not intended for external sharing, deployment takes place in a controlled environment, and safeguards exist against foreseeable misuse (cloud isolation, role-based access). Public-facing and consumer-facing systems are explicitly excluded from the carve-out.
Two Dates for the Calendar
2 December 2026: systems placed on the market before 2 August 2026 get a limited grace period to meet the Article 50(2) marking duty. A system going live today does not.
2 February 2027: the date set in the Code of Practice for implementing an interoperable detection solution based on publicly available industry standards.
Content generated before 2 August 2026 does not have to be labelled retroactively. Note the distinction, though: for visual or audio deepfakes the operative date is generation, while for public-interest text it is the date of publication. Text generated in July and published in September must be labelled.
Deepfakes and Public-Interest Text Under Article 50(4)
A deepfake is defined by three cumulative criteria: resemblance to existing persons, objects, places, entities or events; a subject that exists or could plausibly have existed; and content that falsely appears authentic or truthful. The guidelines confirm that a photorealistic invented person qualifies. Film special effects are generally not deepfakes, but an AI-generated actor does require disclosure.
Here sits the most common technical failure: a deployer cannot rely on the machine-readable marking the provider embedded in the content. Article 50(2) marking is built for machines. Article 50(4) requires a label a human sees and understands. These are two separate layers, and implementing only the first does not satisfy the second.
For the human-perceivable label the Commission published an official icon set built around the capitalised acronym "AI". It comes in four colour variations plus two labelled versions, "Fully AI-Generated" and "Partially AI-Modified", in SVG and PNG. Use is free and requires no attribution, and the icons were user-tested across several Member States. Using them is optional; labelling is not.
The Human Review Exception Is Not Proofreading
Published text that has undergone human review or editorial control need not be labelled. The exception sounds broad, and the guidelines narrow it sharply. Human review requires deliberate examination of the substance of the content by a person with relevant knowledge and professional judgement. Editorial control requires an entity actually exercising authority to approve, alter or reject the substance, and holding ultimate legal responsibility for publication. Superficial, purely formal or procedural checks, such as spell-checking or grammar correction, do not qualify.
What that means for a marketing team producing posts with AI: running a quick language check before publishing does not get you into the exception. Appointing a responsible editor, with the approval process documented, does.
The Creative Works Exception Is Read Strictly
Content that is evidently artistic, creative, satirical or fictional needs a disclosure appropriate to the work, without hampering its display or enjoyment. The guidelines say this exception is to be interpreted strictly. Content that is exclusively informative or commercial in nature cannot benefit from it, and where informative and creative characteristics combine, the informative character always prevails. Advertising can be "evidently creative" only in specific situations, and every advertising example in the guidelines requires standard labelling.
One Disclosure at the Start Is Not Always Enough
The final guidelines generalised a principle that previously applied only to live broadcasts. Where people may not perceive content from the beginning, a single disclosure at the outset is not enough and should be repeated at later points. This matters particularly for audio playing in the background and for eyes-free contexts.
Is It Worth Signing the Code of Practice?
The Code of Practice on Transparency of AI-generated Content was published on 10 June 2026. On 8 and 9 July 2026 the Commission and the AI Board concluded that it is adequate to facilitate the practical implementation of Articles 50(2), 50(4) and 50(5). By the end of July 2026 roughly 190 organisations had signed.
The Code has two sections that can be signed separately, though individual commitments within a section cannot be cherry-picked. Section 1 covers providers, Section 2 covers deployers. An organisation acting as both should sign both.
On the technical side, the provider section offers three marking mechanisms: digitally-signed metadata with a timestamp (in practice, the C2PA standard), imperceptible watermarking, and registry-based fingerprinting as an optional route. Alongside these sit a commitment not to remove or alter existing markings when marked content becomes an input for further processing, including a corresponding prohibition in terms of use, and a commitment to make detection tools available for public verification.
What signing buys in practice: future enforcement against signatories will focus on monitoring adherence to the Code, authorities may treat adherence as a mitigating factor when setting penalties, and the burden of proving compliance by other means drops. Not signing is not a breach, but non-signatories should expect more requests for information from market surveillance authorities and will have to demonstrate compliance independently. The deadline for inclusion in the initial signatory list passed on 27 July 2026, but signing remains open after that.
Five Failure Points We Keep Seeing
- Assuming machine-readable marking also covers the labelling duty. Two separate obligations, on two different actors, for two different audiences.
- Leaning on the "obvious" exception for a chatbot. It is construed restrictively, and the burden sits with whoever relies on it.
- Reading the B2B carve-out broadly. Three cumulative conditions, and consumer-facing systems are explicitly excluded.
- Treating "human review" as proofreading. Substantive examination with documented editorial responsibility is required, not a spell-check.
- Assuming no exposure because you do not target Europe. Posting a deepfake on the globally accessible internet may trigger Article 50(4) if the content is used in the EU, even without deliberate targeting, unless geo-blocking is applied.
One more addition in the final text that is easy to miss: the guidelines expressly link the Article 4 AI literacy duty to the Article 50 obligations. Providers and deployers must ensure adequate AI literacy among staff involved in transparency compliance. In other words, you need training records, not just product records.
What to Do in the Next 30 Days
- Map every touchpoint. Every chatbot, agent, content-generation feature, and every internal use of AI to produce material that gets published externally.
- Assign a role and a paragraph to each one. Provider or deployer, and which of the four duties applies. Some touchpoints will attract more than one.
- Test the human-perceivable labelling layer separately from the technical layer. This is the check most companies skip.
- Put your human review process in writing. Who the responsible editor is, what gets examined, and where it is recorded.
- Check legacy systems against 2 December 2026. The grace period only covers systems placed on the market before 2 August 2026, and it closes in under four months.
- Decide on signing the Code of Practice and document the decision and its reasoning, whichever way you go.
How Lawyerit Works on This
Lawyerit guides Israeli technology companies through Article 50 implementation: mapping touchpoints, assigning the role for each system, drafting the disclosures and labels, building the human review procedure, and deciding on the Code of Practice. We work from the official documents, not from summaries of them.
For background on what was postponed and what was not, see our article on the deferral of the high-risk obligations to December 2027.
Get in touch for a no-cost introductory call.
Frequently asked questions
When does Article 50 of the EU AI Act apply?
From 2 August 2026. The Digital Omnibus postponed the high-risk obligations to December 2027 and August 2028 but did not postpone the Article 50 transparency duties. Systems placed on the market before 2 August 2026 get a limited grace period until 2 December 2026, and only for the Article 50(2) marking duty.
What is the fine for breaching Article 50?
Up to €15 million or 3% of total worldwide annual turnover, whichever is higher. That is lower than the €35 million or 7% ceiling that applies to the prohibited practices under Article 5. Proportionality applies for SMEs and small mid-caps. Enforcement is mainly by national market surveillance authorities.
Does Article 50 apply to an Israeli company with no EU presence?
Yes, if the system output is used in the European Union. The final guidelines narrow this slightly: incidental, unforeseeable or unauthorised downstream use does not automatically trigger the obligations. For deepfakes, posting on the globally accessible internet may fall in scope even without deliberately targeting a European audience.
Is labelling the chat window "AI Assistant" enough?
Not necessarily. The "obvious" exception is construed restrictively and assessed against an average person who is reasonably well-informed, circumspect and observant. Disclosure must be clear and distinguishable from the start of the first interaction. A product name alone does not count as disclosure.
If our provider already marks the content, are we exempt from labelling?
No. The machine-readable marking the provider embeds serves automated detection under Article 50(2). A deployer publishing a deepfake or public-interest text must additionally apply a human-perceivable label under Article 50(4), at the latest at the time of first exposure.
Do we have to sign the Code of Practice?
No. Signing is voluntary and not signing is not a breach. But signing provides a recognised framework for demonstrating compliance, is treated as a mitigating factor, and focuses supervision on monitoring adherence to the Code. Non-signatories should expect more requests for information from market surveillance authorities and must demonstrate compliance independently.
Which AI-generated content is exempt from marking?
A short sequence of numbers, symbols or letters, source code, AI-generated translations, outputs for machine-to-machine communication only, and intermediate-stage outputs in a closed-loop production workflow. There is also an exception for assistive standard editing, and a narrow B2B carve-out requiring three cumulative conditions that does not cover consumer-facing systems.
The above is general information only and does not constitute legal advice. For guidance tailored to your specific circumstances — talk to us.
Sources
- Regulation (EU) 2024/1689 — Article 50 (EUR-Lex)
- European Commission — Guidelines on transparency obligations for providers and deployers of AI systems (20 July 2026)
- European Commission — FAQ: Transparency obligations under Article 50 of the AI Act
- European Commission — Code of Practice on Transparency of AI-generated Content
- European Commission — FAQ: Signing the Code of Practice on Transparency of AI-generated Content
- European Commission — EU icons for labelling AI-generated content
- EU AI Act Service Desk — Article 50
Want to know what applies to your company?
Talk to us