Insights

Article 50 EU AI Act: Your Complete Guide to AI Transparency Obligations

Last updated · June 26, 2026

Of all the EU AI Act's requirements, Article 50 is the one most likely to land on a marketer's desk, a product manager's sprint board, or a newsroom's editorial policy before the end of this year. It doesn't require a conformity assessment or a CE mark. It doesn't demand a risk-management system or a technical dossier. What it demands is simpler - and, for many teams, more disruptive: tell people when they're dealing with AI.

That sounds straightforward. The draft guidelines published by the European Commission in May 2026 make clear it is anything but. Chatbot disclosures buried in terms and conditions won't cut it. A human "spell-check" of AI-generated news copy won't unlock the editorial exemption. And a photorealistic AI-generated person who never existed still counts as a deepfake.

Article 50's transparency obligations become applicable on 2 August 2026. If your organisation uses generative AI in customer-facing products, publishes AI-assisted content, or deploys any form of synthetic media, the clock is running.

The Four Obligations at a Glance

Transparency obligations under Article 50 are not limited to systems classified as "high-risk" - they apply to any AI system used in the four situations the Article covers. In practice, Article 50 is relevant to every business that uses generative AI to produce content.

An organisation with no high-risk AI may still have significant obligations under Article 50: for example, because it develops a customer-facing chatbot, deploys an AI tool that generates news content for publication, or relies on a system that produces deepfake imagery.

Here is what each of the four obligations actually requires.

Art. 50(1) - Chatbot and Interactive AI Disclosure

Who it applies to: Providers of AI systems designed to interact directly with people.

What it requires: Providers must design their systems so that users are informed they are interacting with AI, not a human. This information must be provided in a clear and distinguishable manner at the latest at the time of the first interaction, and must conform to applicable accessibility requirements.

The Commission's draft guidelines are explicit about what does not qualify. A reference in the terms and conditions or product documentation is not sufficient. Technical labels such as metadata or watermarks alone do not meet the requirement either - users typically do not notice them at the point of interaction. Vague terms such as "assistant" or technical descriptions like "this system uses LLMs" also fall short.

AI systems must disclose themselves clearly and upfront. Users need to know they are interacting with AI at the moment of contact - not through a reference buried in fine print. This includes agentic AI that acts autonomously on behalf of users.

Practical examples: A customer-service chatbot on your website must display an unambiguous notice - "You are chatting with an AI" - before or at the very start of the conversation. An AI-powered phone system must announce its nature at the opening of the call. An autonomous AI agent booking meetings on your behalf must identify itself as AI when it contacts third parties.

Article 50(1) exempts providers from the disclosure obligation where the AI nature of the interaction is "obvious from the point of view of a natural person who is reasonably well-informed, observant and circumspect." The Guidelines adopt the "average consumer" standard from EU consumer protection law as the benchmark for assessing obviousness, and provide a multi-factor test that considers the target audience, the potential for vulnerable groups (including children, elderly persons, and persons with disabilities) to be part of that audience, and the level of AI and digital literacy among the intended users.

Art. 50(2) - Machine-Readable Marking of AI-Generated Content

Who it applies to: Providers of AI systems that generate or manipulate synthetic audio, image, video, or text.

What it requires: Providers must implement machine-readable marking and ensure outputs are detectable as AI-generated or manipulated. Article 50(2) is the obligation with the broadest reach in practice. Any AI system that generates or manipulates synthetic image, video, audio or text content falls within scope unless one of the narrow exceptions applies. The guidelines confirm that scope covers AI image and video generators, text-to-speech and voice cloning tools, music and audio synthesis tools, AI text generators including chatbots producing substantively new content, AI tools that manipulate existing media in substantive ways, and agentic AI producing perceptible audio, image, video or text outputs.

This is a provider-side obligation - it sits with the company that builds or places the AI system on the market, not the company that deploys it. If you are a SaaS vendor whose product generates content, this is your obligation to embed.

The Omnibus transitional relief: On 7 May 2026, EU co-legislators agreed on a targeted grandfathering rule for Article 50(2): generative AI systems placed on the market before 2 August 2026 have until 2 December 2026 to comply. The relief is limited to Article 50(2); the other three transparency duties under Article 50(1), (3) and (4) continue to apply from 2 August 2026 without any transition. Only legacy systems benefit. Generative AI systems newly placed on the market after 2 August 2026 must comply from day one.

The AI Omnibus proposal's revised deadline of 2 December 2026 for Art. 50(2) marking obligations has been announced by the Council, though formal adoption remains pending.

Art. 50(3) - Emotion Recognition and Biometric Categorisation

Who it applies to: Deployers of emotion recognition and biometric categorisation systems.

What it requires: Deployers of emotion recognition and biometric categorization systems must inform exposed individuals of the system's operation. This is a deployer-side obligation. If your organisation uses AI to analyse facial expressions during video interviews, infer emotional states from voice patterns in a call centre, or categorise individuals by inferred characteristics, the people being analysed must be told.

This obligation is often overlooked by HR and marketing teams using "engagement analytics" or "sentiment detection" tools. If the tool uses AI to infer emotional or biometric attributes from individuals, Article 50(3) applies.

Art. 50(4) - Deepfake Labelling and AI-Generated Public-Interest Text

Who it applies to: Deployers of deepfakes and deployers publishing AI-generated text to inform the public on matters of public interest.

What it requires: Deployers of deepfakes and of AI-generated text published to inform the public on matters of public interest must disclose the artificial origin of the content.

This is the obligation with the most immediate relevance to marketing, communications, and newsroom teams - and it contains the most nuance.

The Expanded Definition of "Deepfake": What Counts, What Doesn't

The Commission's draft guidelines take a deliberately broad view of what constitutes a deepfake under Article 50(4).

Deepfake rules apply even without intent to deceive. Content that looks or sounds like a real person must be labeled - even if no deception was intended and even if no real individual is depicted.

That last clause is the one that surprises most creative teams. A photorealistic AI-generated person who is entirely fictitious - a composite face, a synthetic spokesperson, a "natural-looking" avatar - still triggers the labelling obligation if a reasonable viewer could mistake them for a real person.

The draft Guidelines clarify that clearly fantastical or physically impossible content - such as dragons or humans flying unaided - falls outside the deepfake definition. The dividing line is realism, not reality. A synthetic landscape, a fantasy creature, or an obviously stylised animation is out of scope. A photorealistic AI-generated human face, even a wholly invented one, is in scope.

For artistic, satirical, and creative content: Where deepfake content forms part of an evidently artistic, creative, satirical, fictional or analogous work or programme, the disclosure obligation is reduced: it is limited to disclosing the existence of the generated or manipulated content "in an appropriate manner that does not hamper the display or enjoyment of the work." This is a reduced obligation, not an exemption. The label must still exist - it just needs to be proportionate to the context.

Does Your Content Trigger the Deepfake Labelling Obligation?

The "Public Interest" Text Obligation - Broader Than You Think

The draft guidelines define "public interest" broadly: it encompasses public administration, fundamental rights, health, the environment, and consumer protection, as well as economic, political, scientific, or cultural developments of societal relevance.

Anyone who publishes AI-generated or AI-edited text with the purpose of informing the public on matters of public interest must disclose that the text has been artificially generated or manipulated. The draft guidelines define "public interest" broadly: it encompasses, among other things, public administration, fundamental rights, health, the environment, and consumer protection, as well as economic, political, scientific, or cultural developments of societal relevance.

Text is considered "published" if it is accessible to an indefinite and relatively large number of readers. Private messages and internal documents are not covered.

In practice, this catches a wider range of content than most teams expect. A corporate sustainability report published on your website, a health-related blog post, a commentary on economic policy, or a press release on a matter of public concern could all fall within scope - even if you are not a news organisation.

The Editorial-Review Exemption: Narrower Than You Hope

The AI Act does provide an exemption for AI-generated public-interest text that has undergone human review and editorial control. But the Commission's draft guidelines make clear this is a high bar.

Human review must involve "deliberate examination of the substance of the content" by persons with "relevant competence and professional judgement." Superficial, solely formal or procedural checks - such as spell-checking, grammatical correction, or cursory editorial approval without substantive engagement - cannot satisfy the exception.

The exemption also requires clearly attributable editorial responsibility with publicly accessible contact details. The exception for public-interest text effectively requires a documented editorial workflow with identified responsible persons, rather than a mere assertion that a human review occurred.

The editorial exemption checklist — all three must be true:

Substantive review: A person with relevant expertise has genuinely examined the content's substance — not just its grammar or formatting.
Documented workflow: The review process is recorded and auditable, with identified responsible persons.
Public editorial accountability: A named natural or legal person holds editorial responsibility, with publicly accessible contact details.

If any one of these is missing, the exemption does not apply and disclosure is required.

Penalties

Non-compliance is not a theoretical risk. Penalties for Article 50 infringements may reach up to €15 million or 3% of total worldwide annual turnover, whichever is higher. Penalties for infringement may reach up to EUR 15 million or 3% of total worldwide annual turnover, whichever is higher. These are the same penalty bands that apply to other non-high-risk AI Act violations under Article 99.

The Code of Practice: A Compliance Safe Harbour

The AI Office published the first draft of its Code of Practice on marking and labelling of AI-generated content on 17 December 2025; adherence may help demonstrate compliance with Art. 50(2), (4), and (5). The Commission prepared these guidelines in parallel with the Code of Practice on marking and labelling of AI-generated content. The guidelines clarify the scope of the legal obligations and address aspects not covered by the code.

The Guidelines are non-binding, but they are the first Commission instrument to provide interpretive guidance across the full scope of Article 50. They were prepared in parallel with the related, but more narrowly scoped, Code of Practice on Transparency of AI-Generated Content, the second draft of which was published on March 5, 2026.

Organisations that sign up to and adhere to the Code will be well-positioned to demonstrate compliance with the marking and labelling obligations. The final version of the Code is expected in June 2026 - monitor the AI Office's digital strategy page for publication.

Practical Compliance Checklist

Use this decision tool to identify which obligations apply to your organisation and what actions to take before 2 August 2026.

For Marketing and Creative Teams

Audit all AI-generated visuals. Any photorealistic synthetic person - even a fictitious one - requires a deepfake label. Fantasy and clearly stylised content is out of scope.
Label AI-generated video and audio. Synthetic voiceovers, AI-generated spokesperson videos, and AI-manipulated brand imagery all require disclosure.
Review public-facing AI-generated copy. If it touches health, environment, public policy, or consumer rights and is accessible to a large audience, the public-interest text obligation may apply.
Don't rely on metadata alone. Machine-readable watermarks satisfy Art. 50(2) for providers, but deployers must also provide visible, human-readable disclosure.

For Product and Engineering Teams

Embed chatbot disclosure at the UI layer. The notice must appear at the point of first interaction - not in the privacy policy, not in the onboarding flow three screens back.
Implement machine-readable marking in your generative outputs (Art. 50(2)). If your system was on the market before 2 August 2026, you have until 2 December 2026 under the provisional Omnibus agreement - but only for legacy systems.
Audit agentic AI workflows. Any autonomous AI agent that contacts third parties on behalf of users must identify itself as AI.
Check vendor contracts. Deployers should map their content workflows to identify where labelling obligations arise. Both providers and deployers should review their contractual arrangements across the AI supply chain, since responsibility is not assumed to rest with the original model provider alone.

For Newsrooms and Editorial Teams

Map every AI-assisted article published to a general audience. If the topic touches public interest (broadly defined), the editorial-review exemption must be genuinely earned.
Document your editorial workflow. Spell-check and a quick read-through are not enough. The review must be substantive, by someone with relevant expertise, and the responsible editor must be publicly identifiable.
Consider labelling as the default. Given how broadly "public interest" is read, it is safer to disclose AI involvement and invest in genuine editorial oversight than to rely on an exemption that may not hold.

For Compliance Leads

Map all AI systems in use against the four Article 50 triggers. Map all AI systems in use across the organisation and identify those covered by Article 50. Check if any exemptions or carve-outs may apply.
Track the Code of Practice. Organisations that comply with the Code will be well-positioned to demonstrate Article 50 compliance.
Note the Omnibus status. The Art. 50(2) transitional relief is a provisional political agreement - formal adoption is still pending. Do not treat December 2026 as a firm deadline until the Omnibus is formally adopted.
Consider GDPR overlap. When personal data are processed, the transparency and information requirements of GDPR Articles 12 to 14 continue to apply alongside Article 50.

Start your Article 50 audit now — free tools on AI Act Navigator:

Risk-Tier Classifier — answer 6 questions to get a provisional risk-tier assessment with a plain-English rationale you can share with stakeholders.
Obligations Checker — once you know your tier, map every applicable obligation to practical compliance steps.
Subscribe to The AI Act Brief — weekly plain-English updates on EU AI Act developments, including the final Code of Practice and Omnibus adoption news.

The Bottom Line

Article 50 is deceptively simple in its headline requirements and genuinely demanding in its detail. The Commission's draft guidelines - published on 8 May 2026, with stakeholder consultation closing on 3 June 2026 - make clear that the bar for compliance is higher than most organisations currently meet.

Transparency obligations will be the next to take effect on 2 August 2026. Providers and deployers of AI systems must adhere to strict transparency obligations, where transparency must be built into the full lifecycle of AI systems, from the moment a user first interacts with the system through to marking relevant outputs.

The organisations that will navigate this most smoothly are those that treat transparency not as a legal checkbox but as a design principle - building disclosure into the user experience, the content workflow, and the editorial process from the ground up. The deadline is fixed. The guidance is now available. The work starts today.

help_outlineDoes Article 50 apply to my organisation if we only use AI internally?expand_more

Generally, no — Article 50(4) only applies to text that is 'published' and accessible to an indefinitely large number of readers. Private messages and internal documents are not covered. However, Art. 50(1) applies to any AI system that interacts directly with people, including employees, so internal chatbots or AI assistants deployed to staff may still trigger the chatbot disclosure obligation.

help_outlineWe use an AI tool to draft marketing copy, which a human then edits. Do we need to label it?expand_more

It depends on the content and the quality of the review. For most marketing copy (product descriptions, ad copy, social posts), the public-interest text obligation under Art. 50(4) is unlikely to apply — it targets content published to inform the public on matters of public interest. However, if your copy touches health claims, environmental matters, or consumer rights, and is published to a large audience, the obligation may apply. In that case, the editorial-review exemption requires genuine substantive review by a person with relevant expertise — not just a quick proofread.

help_outlineWe use a photorealistic AI-generated model in our advertising. Is that a deepfake?expand_more

Almost certainly yes, under the Commission's draft guidelines. A realistic synthetic depiction of a fictitious but natural-looking person still counts as a deepfake under Article 50(4). You must disclose that the content has been artificially generated. The label must be visible and clear — not buried in small print or limited to metadata.

help_outlineWhat does a compliant chatbot disclosure actually look like?expand_more

The guidelines don't prescribe exact wording, but the disclosure must be clear, distinguishable, and presented at the point of first interaction. A notice such as 'You are chatting with an AI assistant' displayed prominently at the start of the conversation window is the baseline. It must be accessible (meeting applicable accessibility requirements) and must not be hidden in a footer, a tooltip, or the terms of service.

help_outlineDoes the Art. 50(2) Omnibus relief mean we don't need to act until December 2026?expand_more

Only for legacy systems — generative AI already on the EU market before 2 August 2026 — and only for the machine-readable marking obligation under Art. 50(2). The other three obligations (Art. 50(1), (3), and (4)) apply from 2 August 2026 with no transitional relief. New systems launched after 2 August 2026 must comply with Art. 50(2) from day one. And the Omnibus relief is still a provisional political agreement — formal adoption is pending.