Article 50: Transparency obligations for providers and deployers of certain AI systems

Disclosure duties independent of high-risk status: inform users they interact with AI; mark AI-generated/manipulated content in machine-readable form; disclose deepfakes; notify subjects of emotion-recognition/biometric-categorisation.

Obligations under this article

• Disclose AI interaction to users — Provider
• Mark AI-generated content machine-readably — Provider, GPAI model provider
• Disclose deepfakes — Deployer

Article 50 disclosures: what to show, and how to prove it

Article 50 has applied since 2 August 2026. There is no single official form — but each disclosure has a specific job to do, and the part most guides skip is that you must be able to evidence that you did it. The wording below is illustrative, to adapt with your own counsel — not legal advice.

1. AI people interact with — chatbots & voice agents (Art. 50(1))

Who: the Art. 50(1) duty sits with the provider, who builds the disclosure in; as good practice the deployer should not suppress it.

Example notice (illustrative): “You’re speaking with an AI assistant, not a person. It can make mistakes — please check anything important.”

Evidence an auditor expects: the exact disclosure text, a screenshot of where and when it appears (at the start of the interaction), and your reasoning for any “it’s obvious” exemption.

2. AI-generated (synthetic) content — image, audio, video, text (Art. 50(2))

Who: the provider of the generative system.

Timing: systems placed on the market from 2 August 2026 must comply now. Generative systems (including general-purpose AI systems) already on the market before 2 August 2026 have a grace period until 2 December 2026 to comply with Art. 50(2) (Art. 111(4), inserted by Regulation (EU) 2026/1744). Content generated before 2 August 2026 need not be labelled retroactively. The grace period covers 50(2) only — not 50(1), (3) or (4).

What compliant looks like: output marked as artificially generated in a machine-readable, interoperable format (e.g. metadata or watermarking such as C2PA), robust enough to survive normal editing — not just a visible caption.

Evidence an auditor expects: which marking standard you implemented, proof the mark persists after ordinary processing (re-encoding, cropping, re-upload) — and, if you rely on the grace period, a dated record of when the system was placed on the market.

3. Deep fakes (Art. 50(4))

Who: the deployer who publishes the content.

Example label (illustrative): “This image/video has been artificially generated or manipulated.” Where the content is evidently artistic, satirical or fictional, the disclosure may be made in a way that does not spoil the work — but it must still be present.

Evidence an auditor expects: the label as published, applied at first exposure (not retrofitted), and your reasoning for any artistic or satirical framing used.

Also under 50(4): a deployer that publishes AI-generated or manipulated text to inform the public on matters of public interest must disclose that too — unless the content underwent human review and a person or organisation holds editorial responsibility.

4. Emotion recognition & biometric categorisation (Art. 50(3))

Who: the deployer operating the system.

What compliant looks like: the people exposed to the system are informed that it is operating, before or at the point of exposure.

Evidence an auditor expects: the notice given, how it was delivered, and when.

The compliance route the Commission recognises — and what it means for your evidence

Three official instruments arrived in June–July 2026 that most Article 50 guides still omit:

• Code of Practice on Transparency of AI-Generated Content — final version published 10 June 2026, drawn up by independent experts in a process facilitated by the AI Office. It covers the marking and detection duties of providers (Art. 50(2) and (5)) and the deep-fake / public-interest-text labelling duties of deployers (Art. 50(4)), as commitments and numbered measures (e.g. Measure 1.1 machine-readable marking, 1.2 non-removal of markings, 2.1 detection mechanisms, 3.1–3.4 effectiveness, reliability, robustness, interoperability). Signing is voluntary; the underlying Article 50 duties are not.
• Commission adequacy Opinion (8 July 2026) and AI Board Adequacy Assessment (9 July 2026) — the Code was assessed as adequately covering Art. 50(2), (4) and (5) under Art. 50(7). The Commission’s own caveat: adherence “does not constitute conclusive evidence of compliance”.
• Commission Guidelines on Article 50 (C(2026) 5054 final) — content approved 20 July 2026; the Commission states they will be formally adopted, and become applicable, once all language versions are available. They clarify scope and give worked examples.

Why this matters for an audit file (Guidelines, section 8.1): for signatories, the Commission and market surveillance authorities say they will focus supervision on whether the signatory has adhered to the Code and implemented its measures; opting out of sections forfeits that benefit for those sections. Non-signatories are expected to demonstrate compliance with Art. 50(2), (4) and (5) “through other adequate means” and to explain how their measures achieve it — the Guidelines name a gap analysis against the Code’s measures as the example, and warn that non-signatories will likely face more information and access requests. Commitments implemented in line with the Code may count as a mitigating factor when fines are set (Art. 99(7)(e)).

Evidence an auditor expects: if you signed — the dated signatory record, and per-measure implementation evidence (which marking technique, the detection interface, the robustness/interoperability testing behind Commitment 3). If you did not sign — a dated, documented gap analysis mapping each of your measures to the Code’s commitments, plus the rationale for any alternative approach.

Three clarifications from the Guidelines worth recording in your file: (1) standard editing falls outside the Art. 50(2) marking duty — the examples include grammar and spell-checking, minor stylistic polishing, format conversion, minor cropping or colour correction, and AI-generated translations of text that do not change the substance; (2) AI agents that interact with people are within Art. 50(1) and, per the Guidelines, should disclose both their artificial nature and the person on whose behalf they act, including at key steps such as authorisation, reporting and validation; (3) systems that are partly interactive and partly generative can use the 2 December 2026 transition for the 50(2) marking duty only — the 50(1) disclosure has applied since 2 August 2026. These are Commission guidance, not amendments to the Regulation; document which reading you rely on.

The EU AI Office publishes a tested set of icons for labelling AI-generated content for the visual side of these duties. Article 50 is independent of high-risk status — a limited-risk system can still trigger it.

Legal source

Art. 50 AIA · European Commission — Article 50 FAQ (grace period for Art. 50(2)) · Commission Guidelines on Article 50 transparency obligations (C(2026) 5054 final, content approved 20 Jul 2026) · Code of Practice on Transparency of AI-Generated Content (final, 10 Jun 2026) · Commission Opinion on the adequacy of the Transparency Code (8 Jul 2026) + AI Board Adequacy Assessment (9 Jul 2026)

Not legal advice. This is a practitioner reference generated from a cited knowledge base built on the primary legal text and official Commission guidance. Verify against the cited primary sources before relying on it for a compliance decision. See editorial standards & methodology.