AI Act Article 50: labelling AI-generated content
What Article 50 requires for chatbots and AI content: provider marking, deployer disclosure, deepfakes and the human editorial-control exception.
AI ACT · CONTENT TRANSPARENCY
“Label everything touched by AI” is not the rule in Article 50. The Regulation separates direct interaction with a system, technical marking of synthetic outputs and visible disclosure of certain materials. A company’s role and the way content reaches the public change the obligation.
Four situations, not one universal label
Article 50 groups several different obligations. The first concerns systems interacting directly with people. The second concerns providers of systems generating synthetic content and technical marking of outputs. The third concerns deployers using emotion recognition or biometric categorisation. The fourth concerns deepfakes and certain texts published to inform the public. [EU-REG] [EU-A50-G]
These obligations are not interchangeable. A marker embedded in a file does not automatically replace a visible disclosure required from the deployer. A visible notice written by a publisher does not satisfy the provider’s duty to design the system so that certain outputs can be detected as artificially generated or manipulated. [EU-A50] [EU-A50-C]
When users must be told they are interacting with AI
Providers of systems intended to interact directly with people must design them so that individuals are informed they are interacting with an AI system, unless this is obvious to a reasonably well-informed and observant person. The rule targets products such as chatbots, conversational agents and interactive avatars. [EU-REG] [EU-A50]
Information should appear no later than the first interaction or exposure, in a clear and distinguishable manner, and must comply with applicable accessibility requirements. A sentence hidden in the privacy policy is not a sound solution for an interface that appears to be a human operator. The message should be placed where users can understand it before forming the wrong impression. [EU-REG] [EU-A50-G]
Technical marking is a provider obligation
Providers of AI systems generating synthetic audio, image, video or text must ensure that outputs are marked in a machine-readable format and detectable as artificially generated or manipulated. The solution should be effective, interoperable, robust and reliable to the extent technically feasible. [EU-REG] [EU-A50-G]
The guidance excludes certain outputs that do not have the relevant nature for this duty, such as source code, short sequences of symbols or exclusively machine-to-machine communication. Standard editing functions and assistance that do not substantially alter content or meaning may also fall outside it. The exception should be assessed by actual function, not claimed for every product called an “assistant”. [EU-A50]
Deepfake disclosure is a deployer obligation
An organisation using an AI system to generate or manipulate image, audio or video that appreciably resembles real people, objects, places, entities or events and could appear authentic must disclose that the material was artificially created or manipulated. Here the duty falls on the deployer publishing or exposing the material. [EU-REG] [EU-A50]
For evidently artistic, creative, satirical, fictional or similar works, disclosure may be made in a way that does not hinder display or enjoyment of the work. This does not mean the artificial origin may be hidden completely. In an advertisement, case study or realistic media appearance, context and the risk of misleading the audience should be assessed before publication. [EU-REG] [EU-A50-G]
The special rule for public-interest text
Deployers using AI to generate or manipulate text published to inform the public about matters of public interest must, in principle, disclose its artificial origin. This does not automatically cover every product description, email, commercial post or SEO page. The publication purpose and nature of the subject must be assessed in context. [EU-REG] [EU-A50]
There is an important exception where the text has undergone human review or editorial control and a natural or legal person holds editorial responsibility for publication. The exception should not be reduced to formally pressing an “Approve” button. Credible control means checking claims, sources and context and genuinely assuming the publication decision. [EU-REG] [EU-A50-G]
What happens to existing material and systems
Article 50 applies from 2 August 2026. Certain systems placed on the market before that date have a limited transition until 2 December 2026 for the technical marking duty in paragraph 2. This is not a general postponement of chatbot information or deployer disclosure obligations. [EU-A50]
Material generated before 2 August 2026 does not have to be labelled retroactively under Article 50, although the Commission encourages voluntary labelling where possible. Editorial decisions may also consider other rules, audience expectations and reputational risk; the absence of a retroactive duty does not prohibit useful transparency. [EU-A50]
A practical publishing matrix
A publisher or SaaS company should assess separately the system, output type, its own role, audience, purpose of the material and editorial control. The answer should then become a product requirement or editorial step rather than remaining in a legal memo unseen by the team. [EU-A50-G] [EU-A50-C]
- Chatbot or conversational agent: a clear notice at first interaction unless its AI nature is obvious.
- Generator provider: detectable technical marking for synthetic outputs within scope.
- Deployer publishing a deepfake: visible and distinguishable disclosure of artificial origin.
- Public-interest text without editorial control: assess the disclosure duty before publication.
- Human-reviewed text with editorial responsibility: retain evidence of the process, not only final status.
Applying the rule to SEO and Dosinescu products
For AYSA.RO and an editorial workflow, the central question is whether material goes through research, verification, rewriting and human responsibility. For AYSA.AI, transparency should also be assessed as a product function: what the system generates, which metadata it retains, what users see and what control they have before publishing to WordPress. [EU-A50] [EU-A50-C]
For CanUHelp APP or a chatbot installed on a client website, the message that the user is speaking with AI should be designed into the interface and tested on mobile, not later added to a document. If the conversation can be transferred to a person, escalation and the change of counterpart should be easy to understand. [EU-A50-G]