Labelling AI Content: Implementing Art. 50 AI Act in Practice

Labelling AI content under Art. 50 of the AI Act means: anyone who uses AI to generate or edit synthetic text, images, audio or video must make this transparent. Providers mark outputs in machine-readable form; deployers visibly label deepfakes and certain public texts. The transparency obligations of the AI Act (Regulation (EU) 2024/1689) apply from 2 August 2026 – regardless of whether the system is high-risk.

For content teams and agencies working with AI agents, Art. 50 is the central compliance layer: most marketing and content workflows do not end up in the full high-risk regime, but precisely here.

• Who: Providers (model/system suppliers) and deployers (users) have separate obligations – the role determines what you must do.
• What: Machine-readable marking of outputs (provider) plus a visible notice for chatbots, deepfakes and public-interest texts (deployer).
• When: Substantive obligation from 2 August 2026; technical marking solutions (watermarks, metadata) by 2 December 2026 at the latest.

Note: This article classifies the legal situation in a practical way and is no substitute for legal advice. All article numbers and deadlines reflect the position as of May/June 2026; they may still shift through the formal adoption of the Digital Omnibus.

Why Art. 50 is the decisive lever for content automation

The AI Act has no dedicated legal term for "AI agent". A content agent is captured via the general definition of an AI system (Art. 3(1)), via the rules for general-purpose AI models, and via the transparency layer of Art. 50. Within the risk model, Art. 50 sits between the unregulated minimal-risk tier and the high-risk regime, and applies independently: many limited-risk agents – such as a copywriting agent or a customer service bot – are subject exclusively to Art. 50, without any further high-risk obligations.

That is the good news for the DACH SME sector and agencies: anyone who does not automate recruitment, creditworthiness assessment or insurance pricing need not go through a full conformity procedure. The bad news: in practice, Art. 50 is underestimated, because the obligation applies differently depending on the role (provider vs. deployer) and the type of content.

The four transparency obligations of Art. 50 at a glance

Art. 50 bundles four obligations. Two fall on the provider (supplier of the system or model), two on the deployer (the company that uses the system under its own responsibility).

The key difference: Art. 50(2) requires machine-readable marking (technical, within the file), whereas Art. 50(3) and (4) require a human-readable, visible notice. Both may be necessary in parallel – for example, where an AI tool generates an advertising video (provider marks it in machine-readable form) and your agency publishes it as a deepfake spot (deployer labels it visibly).

Machine-readable marking vs. visible notice

Machine-readable marking (provider obligation, Art. 50(2)): The output itself carries a technical label. In practice, this is done via digital watermarks or provenance metadata bound to the file. The AI Act does not prescribe a specific technology, but requires the solution to be effective, interoperable, robust and reliable. Across the industry, the open C2PA standard (Content Provenance, Coalition for Content Provenance and Authenticity) has become established for provenance attestation; it is a common way to achieve interoperability, but is not named explicitly in Art. 50. Responsibility for this marking lies with the supplier of the content-generating system, not with the deploying company.

Visible notice (deployer obligation, Art. 50(3) and (4)): This is directed at people. Examples: the "generated by AI" label on a synthetic image or video, the information "You are chatting with an AI assistant", or notice that an emotion-recognition system is in use.

The EU Commission is fleshing out the technical implementation step by step: a first draft of a Code of Practice on labelling and marking AI-generated content was published on 17 December 2025, and a second on 3 March 2026; draft guidelines on implementing the Art. 50 transparency obligations followed on 7/8 May 2026 (under consultation). These documents address watermark robustness, detection tools, and a taxonomy distinguishing between "fully AI-generated" and "AI-assisted".

The exceptions – and their limits

Art. 50 is not absolute. Three exceptions are relevant in practice:

• "Obvious from the context" (Art. 50(1)): Can the chatbot notice be dropped where it is clear anyway that an AI is responding? This exception is construed narrowly. A human-sounding voice agent must disclose – "obviousness" is precisely not assumed where the voice is deceptively realistic.
• Editorial responsibility (Art. 50(4), 2nd subparagraph): For public-interest texts, the obligation does not apply where the text underwent editorial review and a natural or legal person assumes editorial responsibility. This exception is comparatively broad – often applicable to professionally edited B2B content with clear accountability.
• Art and satire (Art. 50(4), deepfake): For artistic or satirical works, disclosure is limited to an "appropriate" extent that does not disrupt enjoyment. This exception is construed narrowly and does not cover commercial advertising.

Important: anyone relying on an exception should document the reasoning. In the event of a review, that documentation is the evidence.

Practical example: an agency publishes an AI campaign

A Vienna-based agency produces a campaign for a B2B client with three content types. The obligation landscape looks as follows:

1. 30 AI-generated LinkedIn posts (text): The writing tool (provider) marks the outputs in machine-readable form (Art. 50(2)). The agency edits each post and names a responsible person → the visible text obligation under Art. 50(4) does not apply to the editorially controlled texts.
2. A synthetic product video with an AI avatar (deepfake): Here Art. 50(4) first subparagraph applies in full. As deployer, the agency must visibly disclose that the video was artificially generated – the art/satire exception does not apply to advertising.
3. An AI customer service bot on the landing page: The bot supplier (provider) must build in the notice "You are interacting with an AI" (Art. 50(1)). If the agency brands the bot under its own name, it may become a provider itself under Art. 25 – in which case the obligation passes to it.

If the missing deepfake labelling is challenged, the penalty risk sits within the second tier: up to EUR 15 million or 3 percent of worldwide annual turnover (Art. 99(4)); for SMEs, the lower amount applies (Art. 99(6)).

A tick-box implementation checklist

• Clarify the role per system: Are you a deployer (the standard case) or, through rebranding/modification, a provider (Art. 25)?
  
• Compile an inventory of all customer-facing AI touchpoints (chatbots, voicebots, generated text/images/videos).
  
• Art. 50(1): Verify that the supplier's chatbot/voicebot notice is present and visible.
  
• Art. 50(2): Confirm with the tool supplier that outputs are marked in machine-readable form (watermarks/metadata).
  
• Art. 50(3)/(4): For every deepfake and every public-interest text, implement deployer-side, visible disclosure.
  
• Document exceptions: Record in writing the reasoning for "obvious from the context" or editorial responsibility.
  
• Update privacy notices, terms and conditions, and AI usage notices.
  
• Deadlines in the roadmap: substantive obligation 2 August 2026, technical marking by 2 December 2026.
  
• Check interactions: DSA obligations (very large platforms) and unfair-competition law (UCPD) apply in addition to Art. 50.
  

Interaction with adjacent regimes

Art. 50 does not stand alone. Very large online platforms and search engines are additionally subject to the risk assessment obligations of the Digital Services Act for AI-generated content. Unfair-competition law (UCPD) applies on top: an AI bot that passes itself off as a human salesperson breaches both Art. 50 and the prohibition on misleading commercial practices. For DACH SMEs below the platform thresholds, the DSA obligations are limited, but the users' expectation of transparency nonetheless holds.

For agencies and B2B decision-makers

Treat Art. 50 as an integral part of every content-automation workflow, not as a downstream legal question. In concrete terms, this means: maintaining a role and use-case register, contractually securing supplier commitments to machine-readable marking, building in deployer-side deepfake and chatbot notices as a default, and documenting the reasoning for exceptions. Agencies that deliver transparency by design reduce their clients' penalty risk and build trust – DACH audiences increasingly reward clear AI labelling. Those who set up the processes now will be robustly positioned ahead of 2 August 2026. For a legally binding assessment of your specific case, consult qualified legal counsel.