EU AI Act Transparency Obligations Take Effect on 2 August 2026 as High-Risk Rules Slip

On 2 August 2026 the transparency obligations in Article 50 of the EU AI Act became applicable. Four duties took effect: providers must make clear when people are interacting with an AI system such as a chatbot, voice assistant or agent, unless it is obvious; providers of generative systems must embed machine-readable markings in synthetic audio, image, video and text and offer detection mechanisms; deployers must inform individuals when emotion recognition or biometric categorisation systems are used; and deployers must disclose artificially generated or manipulated content on matters of public interest, unless it has undergone substantive editorial review.

The European Commission adopted final guidelines on these obligations on 20 July 2026 and confirmed the voluntary Code of Practice on Transparency of AI-generated Content as an adequate route to compliance, with signatories receiving a presumption of conformity. Penalties for non-compliance can reach 15 million euros or 3% of worldwide annual turnover, whichever is higher.

The same date was originally the deadline for the Annex III high-risk regime. That no longer applies. The Digital Omnibus on AI, approved by the European Parliament on 16 June and by the Council at the end of June, moved Annex III standalone high-risk obligations to 2 December 2027 and Annex I product-embedded obligations to 2 August 2028. It also gave generative systems already on the market until 2 December 2026 to meet the marking and detection duties, added a prohibition on AI systems that generate non-consensual intimate imagery, clarified the "safety component" definition and centralised general-purpose AI enforcement in the AI Office.

Vendors have started to respond publicly. Anthropic said on 13 August that Claude products released from 2 August onward embed machine-readable watermarks, with older models updated by December, and that the feature rolls out globally rather than only in Europe.

Why it matters

The high-risk delay is real relief for organisations that were behind, but the Article 50 duties are not a soft landing. They are the first AI Act obligations that touch the customer-facing surface of almost every company using generative AI, regardless of whether any use case is high-risk. A marketing team publishing synthetic imagery, a contact centre running a voice bot, or an HR function using emotion analytics in interviews is now inside scope. The California AI Transparency Act became operative on the same day, so multinationals face parallel provenance and disclosure work on both sides of the Atlantic.

What it means for leaders

• Inventory every AI touchpoint that faces a person. Chatbots, voice agents, avatars and content pipelines each need a disclosure decision documented; NIST AI RMF's Map function is a natural home for this.
• Confirm your vendors' marking and detection support. Ask generative AI providers how they meet the machine-readable marking duty and whether detection tooling is available to you as deployer.
• Do not deprioritise high-risk readiness. December 2027 is one budget cycle away; use the extension to finish risk management, data governance and logging work rather than to pause it.
• Decide on the Code of Practice. Signing brings a presumption of conformity; not signing means demonstrating compliance another way and, per legal commentary, potentially closer scrutiny.
• Align with ISO/IEC 42001 documentation. Transparency measures and the rationale behind them should be traceable in the AI management system, not just in product notes.