On July 20, 2026, the European Commission published official Guidelines on the implementation of the transparency obligations for certain AI systems under Article 50 of the AI Act. The guidelines clarify who must disclose AI interaction and label AI-generated content under Regulation (EU) 2024/1689, with those Article 50 duties starting to apply on August 2, 2026, thirteen days after publication.
The Commission announcement and companion Guidelines on Transparency of AI-Generated Content are the first Commission soft-law package dedicated to Article 50. Providers of interactive and generative AI systems must implement disclosure and machine-readable marking designs now. Deployers must prepare deepfake, public-interest text, emotion recognition and biometric categorisation notices for the same August 2 start date, subject to a narrow grace window for pre-existing systems on marking only.
What exactly changed on July 20, 2026?
Article 50 itself has been on the statute book since the AI Act entered into force. What is new is Commission guidance that defines scope, exemptions and compliance pathways for those obligations. The guidelines spell out four cumulative criteria for when a provider must inform a person that they are interacting with an AI system: the system qualifies as an AI system; it is designed for a genuine two-way exchange; the interaction is direct (the AI communicates with the person, not via a human intermediary); and the counterpart is a natural person. Background-only or machine-to-machine systems fall outside Article 50(1).
For generative outputs, Article 50(2) requires machine-readable marks that make synthetic audio, image, video or text detectable as AI-generated or manipulated. The guidelines carve out short token sequences, source code, pure machine-to-machine outputs, closed-loop industrial environments (unless the output is final), and a narrow standard-editing assistive function. They also describe a limited B2B / industrial-context exemption when the stated conditions are met. The voluntary Code of Practice on Transparency of AI-Generated Content is the Commission-endorsed route to demonstrate marking and labelling compliance; non-signatories must show equivalently adequate alternative means.
Who is hit, and what must they do by when?
Providers (including third-country providers whose system output is used in the EU) must meet Articles 50(1), (2) and (5) before placing systems on the market or putting them into service. Deployers (professional users, not purely personal use) must inform people exposed to emotion recognition or biometric categorisation systems under Article 50(3), and must clearly label deepfakes and AI-generated or manipulated public-interest text without human review or editorial control under Article 50(4). Machine-readable marks from the provider do not discharge the deployer's human-readable disclosure duty for deepfakes.
| Obligation | Actor | Application date |
|---|---|---|
| Inform users of direct AI interaction (Art. 50(1)) | Providers | August 2, 2026 |
| Machine-readable marking of synthetic content (Art. 50(2)) | Providers | August 2, 2026; systems placed on the market before that date: December 2, 2026 |
| Emotion recognition / biometric categorisation notice (Art. 50(3)) | Deployers | August 2, 2026 |
| Deepfake and public-interest text labelling (Art. 50(4)) | Deployers | August 2, 2026 |
| Content generated before August 2, 2026 | Deployers | No retroactive labelling duty (voluntary catch-up encouraged) |
Fines for breaches can reach EUR 15 million or 3% of total worldwide annual turnover, whichever is higher, with proportionality for SMEs and small mid-caps. National market surveillance authorities lead enforcement; the AI Office covers a narrow set of GPAI-linked or DSA VLOP/VLOSE-integrated systems; the European Data Protection Supervisor covers EU institutions. The Commission Article 50 Q&A confirms there is no general grace period beyond the December 2, 2026 marking deferral for systems already on the market before August 2.
How should compliance teams operationalise the guidelines in the next two weeks?
Map every customer-facing chatbot, agent, avatar and generative product against the four Article 50(1) criteria and decide whether the "obvious interaction" exception can honestly apply (the guidelines say interpret it restrictively). Inventory generative pipelines that produce audio, image, video or text for EU users and select either Code of Practice adherence or a documented alternative marking stack that is effective, reliable, robust and interoperable. For deployers, build first-exposure labels for deepfakes that a person can perceive without technical tools, and define when public-interest text has undergone real human review or editorial control versus spell-check theatre.
Treat the guidelines' practical examples (standard editing, movie VFX backgrounds, professional versus personal deepfake dissemination) as the audit checklist national authorities will use. Continuous, per-jurisdiction real-time monitoring surfaces this kind of Commission soft-law drop the moment it publishes, so product and legal owners are not discovering Article 50 clarifications after the application date.
Take advantage of this real-time watch
Verify whether each AI product you place or use in the EU triggers Article 50(1) to (4), lock the August 2, 2026 and December 2, 2026 dates into your compliance calendar, and brief product, trust and safety, and legal teams on the disclosure and marking designs that must ship. Decide within days whether to sign the transparency Code of Practice or document an equivalent marking programme. Obsidian keeps the AI Act Article 50 track visible as Commission guidance and national enforcement signals continue to land.
