Commission clarifies how chatbots and deepfakes must declare their artificial origin

The European Commission has published guidelines on how the transparency obligations of Article 50 of the AI Act must be applied, setting out when chatbots, synthetic media, deep fakes and emotion-recognition systems have to tell people that a machine is involved. The communication, registered as C/2026/6005, was published in the Official Journal of the European Union on 12 November 2026, just over three months after those obligations became applicable. It instructs national supervisory authorities, providers and deployers alike on how to comply in a “consistent, effective, proportionate and uniform manner”.

The AI Act, Regulation (EU) 2024/1689 of 13 June 2024, entered into force on 1 August 2024 and follows a risk-based approach in which transparency risks form one of four categories. The guidelines are issued under Article 96(1)(d) of that regulation and run to 43 pages, the result of a broad public consultation, input from Member States in the AI Board and comments on a draft published for consultation. The Commission states they are non-binding, since authoritative interpretation of the AI Act may ultimately be given only by the Court of Justice of the European Union.

Four obligations, four different addressees

Article 50 contains four distinct duties, each attached to a different actor or output:

• Article 50(1): providers of AI systems interacting directly with people must design them so users are informed they are talking to an AI system.
• Article 50(2): providers of systems generating or manipulating synthetic image, video, audio or text must mark outputs in a machine-readable format, detectable through solutions that are effective, interoperable, robust and reliable.
• Article 50(3): deployers of emotion recognition and biometric categorisation systems must inform the people exposed to them.
• Article 50(4): deployers must disclose that content is a deep fake or AI-generated text published to inform the public on matters of public interest.

Article 50(5) adds that all such information must be given clearly and distinguishably, at the latest at the first interaction or exposure, and in accessible formats.

What counts as adequate disclosure

For interactive systems the Commission accepts almost any technique as long as it is perceptible at the point of use: a chatbot opening a conversation by saying it is AI-based, a voice assistant announcing itself aloud, persistent icons or watermarks, or AI identifiers disclosed through digital identity wallets. A single prominent notice before the first interaction will usually suffice, though periodic reminders are expected in riskier settings such as financial, health or legal advice, AI companions, and services used by children, the elderly or people with disabilities.

The guidelines also list what does not work on its own: disclosures buried in terms and conditions or URLs, machine-readable metadata invisible to users, vague references to an “assistant”, and generic statements that “services on this website use AI”. The obvious-interaction exception is judged against a reasonably well-informed, observant and circumspect person, mirroring EU consumer law.

Deep fakes, satire and newsroom carve-outs

A deep fake requires four cumulative elements under Article 3(60) of the AI Act: appreciable resemblance, to an existing subject, whether persons, objects, places, entities or events, capable of falsely appearing authentic or truthful. Hyperrealistic avatars, digital replicas and de-aging qualify; dragons or humans flying unaided do not. Disclosure must not be retroactive for content generated before 2 August 2026, but text produced earlier and published after that date must be labelled.

Deep fakes inside evidently artistic, creative, satirical, fictional or analogous works are not exempt, but may be disclosed in a manner that does not hamper enjoyment of the work. Text that has undergone human review or editorial control and carries editorial responsibility is excluded from the labelling duty.

Codes of practice, supervisors and fines

Adherence to a Code of Practice on Transparency of AI-Generated Content, published on 10 June 2026 and assessed as adequate under Article 50(7), lets providers and deployers demonstrate compliance and is treated as a mitigating factor when fines are set. Non-signatories face more requests for information and must explain their alternative measures.

Enforcement falls to national market surveillance authorities, the AI Office and the European Data Protection Supervisor within the system of Regulation (EU) 2019/1020. Breaches carry fines of up to EUR 15 000 000 or 3 % of worldwide annual turnover, whichever is higher, with a lower ceiling for SMEs; EU institutions, bodies and agencies face up to EUR 750 000. The AI Omnibus gives generative systems already on the market before 2 August 2026 until 2 December 2026 to comply with the marking duty.

The guidance converts a one-line treaty-style obligation into auditable practice, and it lands in the middle of the first enforcement cycle: platforms, newsrooms, advertisers and model providers now know which labels supervisors will look for, and which excuses will not survive scrutiny. Because the Commission has promised to review the text as case law and market surveillance experience accumulate, the document is likely to become the reference point for every labelling dispute in the coming years.

Source: Official Journal of the European Union, C series, No C/2026/6005, 12 November 2026, section I (official reference: C/2026/6005).