On 2 August 2026, the transparency obligations of Article 50 of the European AI Regulation (Regulation (EU) 2024/1689, the “AI Act”) become applicable. A targeted transition applies nonetheless: providers of generative systems placed on the market before that date have until 2 December 2026 to implement the machine-readable marking required by Article 50(2) (Article 111(4), inserted by Regulation (EU) 2026/1744).
Article 50 does not cover every professional use of AI. It targets four specific situations: systems interacting directly with individuals, systems generating synthetic content, emotion recognition and biometric categorisation, and the generation or manipulation of deep fakes and certain texts of public interest.
In practice: individuals must be informed when they interact with an AI system; providers must implement machine-readable marking of synthetic outputs; and deployers must perceptibly disclose deep fakes and certain texts of public interest. These obligations are distinct and do not necessarily fall on the same actor.
The four transparency obligations of Article 50
1. Chatbots and AI assistants must identify themselves as such
Any AI system intended to interact directly with natural persons must be designed so that those persons know they are interacting with an AI system (Article 50(1)). The only notable exception: where this is obvious to a reasonably well-informed and observant person, taking the context into account. When in doubt, display it.
Who is concerned: the provider of the system. Beware: if your company commissioned the chatbot and puts it into service under its own name or trademark, it may itself qualify as the provider (Article 3(3)). It cannot then simply pass compliance on to the technical contractor. If it merely integrates a third-party system, it must at the very least verify, contractually and technically, that the required information is actually displayed.
2. Synthetic content must be technically marked — by the provider
Providers of generative AI systems — including general-purpose AI systems — must ensure that outputs (audio, image, video, text) are marked in a machine-readable format and detectable as artificially generated or manipulated (Article 50(2)). The marking must be effective, interoperable, robust and reliable, as far as technically feasible. Exceptions: standard editing assistance functions or modifications that do not substantially alter the input content.
Transitional regime: for systems generating synthetic content placed on the market before 2 August 2026, the provider has until 2 December 2026 to comply with this marking obligation (Article 111(4)). This transition only concerns paragraph 2: it postpones neither the information duty for chatbot interactions (paragraph 1), nor the information duty for emotion recognition and biometric categorisation (paragraph 3), nor the perceptible labelling of deep fakes and certain texts of public interest (paragraph 4). The Commission’s Guidelines expressly confirm this.
Who is concerned: the provider of the generative system. For user companies, the reflex is contractual: does your provider ensure this marking, and within what timeframe?
3. Emotion recognition and biometric categorisation: inform the persons exposed
Deployers of emotion recognition or biometric categorisation systems must inform the persons exposed to them of the operation of the system — not merely of its existence — and process personal data in accordance with the GDPR (Article 50(3)). A useful reminder: transparency does not cure an otherwise prohibited use — in the workplace and in education, emotion recognition remains prohibited, except for medical or safety reasons.
4. Deep fakes and informational texts: label visibly — by the deployer, in these cases only
Deployers who generate or manipulate image, audio or video content constituting a deep fake within the meaning of the Regulation must disclose that the content has been artificially generated or manipulated (Article 50(4)). For evidently artistic, creative, satirical or fictional works, disclosure is still due, but in an appropriate manner that does not hamper the display or enjoyment of the work.
The same logic applies to AI-generated text published with the purpose of informing the public on matters of public interest: labelling is mandatory, unless the text has undergone human review or editorial control and a natural or legal person holds editorial responsibility. On this point, the Guidelines are demanding: the review must be genuine, carried out by a competent person, and cannot be reduced to a superficial grammar check.
Beware of the shortcut: not every piece of content produced with AI must carry a visible label. The perceptible disclosure obligation of paragraph 4 only covers deep fakes in the legal sense and the aforementioned texts of public interest. A fantasy illustration generated by AI is not necessarily a deep fake and does not automatically need to carry a visible icon — even though the technical marking of paragraph 2 does apply on the provider’s side. The two obligations only cumulate where their respective conditions are simultaneously met. The Commission’s official page on the icons expressly states this.
In all cases: clear, immediate, accessible
The information must be provided in a clear and distinguishable manner at the latest at the time of the first interaction or exposure, and comply with applicable accessibility requirements (Article 50(5)).
The Commission’s Guidelines of 20 July 2026
On 20 July 2026, the European Commission published its Guidelines on Article 50 (Guidelines on the implementation of the transparency obligations for certain AI systems under Article 50 of the AI Act). Without creating new obligations — they are not binding — they clarify:
- the scope of the provision: which systems fall under each of the four regimes;
- its exceptions: obviously apparent interaction, standard editing assistance, creative works, uses authorised by law;
- the modalities of the information: when, where and how to inform so that disclosure is clear and distinguishable from the first exposure;
- the distinction between technical marking and perceptible information: machine-readable marking (watermarks, metadata — Article 50(2)) falls on the provider and is addressed to machines; perceptible disclosure (labels, notices — Article 50(4)) falls on the deployer and is addressed to the public, only in the cases covered by that paragraph;
- the scope of human review and editorial control: the conditions under which an AI-assisted text of public interest, genuinely reviewed and editorially assumed, escapes the labelling obligation.
▶️ These Guidelines are complemented by the Code of Practice on Transparency of AI-Generated Content, whose final version was published on 10 June 2026 and which the Commission found adequate on 8 July 2026, after the opinion of the AI Board. Adherence is voluntary (Article 50(7)) and the Code covers paragraphs 2, 4 and 5 of Article 50 — not the entire article. Signing up facilitates the demonstration of compliance without creating any irrebuttable presumption; non-signatories will have to demonstrate the adequacy of their measures by their own means.
How to label? The EU provides official icons
For the cases where perceptible disclosure is due, the Commission has published official icons for labelling AI content. Three pictograms — “AI” (base icon), “AI GENERATED” (fully generated content) and “AI MODIFIED” (partially modified content) — in four graphic variants (black, white, black at 50% transparency, white at 50%), in SVG and PNG formats.
- free to use, without attribution to the Commission or the AI Office;
- optional: what is binding is the labelling obligation, not the choice of icon — an equivalent clear label will do, and using the icon does not, by itself, constitute proof of compliance;
- perceptible from the first exposure; embedding in the content and persistence upon sharing stem notably from the Code of Practice, for its signatories;
- accompanied, where possible and in accordance with applicable accessibility requirements, by appropriate alt text or an ARIA label.
What sanctions?
Non-compliance with the transparency obligations of Article 50 exposes providers and deployers to administrative fines of up to — these are ceilings — EUR 15,000,000 or 3% of total worldwide annual turnover for the preceding financial year, whichever is higher (Article 99(4)(g)). The actual amount must remain effective, proportionate and dissuasive, and depends on the circumstances of the infringement.
For SMEs, including start-ups, the lower of the two amounts applies instead (Article 99(6)) — a favourable regime that Regulation (EU) 2026/1744 now extends to small mid-cap enterprises (Article 99(6a)).
Your checklist
- Inventory your AI systems falling within the four targeted situations: chatbots, content generators, emotion recognition or biometric categorisation, deep fakes.
- Qualify your role for each system: provider (including where you put into service under your own brand a system developed for you) or deployer — the obligations differ.
- Chatbots: check that a “you are talking to an AI” notice appears from the first interaction.
- Providers of generative systems: implement machine-readable marking — by 2 December 2026 at the latest for systems placed on the market before 2 August 2026, immediately for the others.
- Deployers: identify what, in your content, constitutes a deep fake or a text of public interest within the meaning of the Regulation — that is what requires a visible label, not every AI-generated visual.
- Published texts: where an AI-generated or manipulated text of public interest undergoes genuine human review or editorial control — by a competent person — and a natural or legal person holds editorial responsibility, document that process and the identity of the person responsible.
- Accessibility: clear information from the first exposure, alt text or ARIA label where possible.
- Reference documents: read the Guidelines of 20 July 2026 for borderline cases, and assess whether adhering to the Code of Practice makes sense for you.
📦 To make your life easier, we have prepared a practical resource: our guide “when to label, which icon to choose, where to place it”, with access to the official EU icons (SVG + PNG, all variants). See the practical guide and download the icons (in French).
The firm assists companies with the qualification of their AI systems and AI Act compliance — contact.
Sources
- Regulation (EU) 2024/1689 of 13 June 2024 (Artificial Intelligence Act), Articles 3, 50, 99 and 111 — EUR-Lex, CELEX 32024R1689
- Regulation (EU) 2026/1744 of 8 July 2026 (“Digital Omnibus on AI”), published on 24 July 2026, in force since 27 July 2026 — new Articles 111(4) and 99(6a) — EUR-Lex, CELEX 32026R1744
- European Commission, Guidelines on the implementation of the transparency obligations for certain AI systems under Article 50 of the AI Act, 20 July 2026 — digital-strategy.ec.europa.eu
- European Commission / AI Office, Code of Practice on Transparency of AI-Generated Content, final version of 10 June 2026 — digital-strategy.ec.europa.eu
- European Commission, EU icons for labelling AI-generated content — digital-strategy.ec.europa.eu
This post provides general information and does not constitute legal advice. For an analysis tailored to your situation, contact the firm.