At a glance: EU AI Act. Article 50 transparency obligations take effect: August 2, 2026. Annex III high-risk AI formally deferred to December 2, 2027 (Digital Omnibus package, Council final approval June 29, 2026). Scope of Article 50: chatbot disclosure, AI-generated content marking, emotion/biometric-categorization disclosure, deepfake labeling. Penalties: €15 million or 3% of global annual turnover, whichever is higher. Grace period: existing generative AI systems on market before August 2 have until December 2, 2026 for machine-readable marking; new systems have no grace period. See also our coverage of the Digital Omnibus simplification deal from May 2026. Part of our AI policy and regulation coverage.

The EU AI Act’s August 2, 2026 date has been misread by many as something that was pushed back. The Annex III high-risk AI obligations were delayed — that much is accurate. But Article 50, the Act’s transparency chapter, was not. It takes effect in twelve days. For providers of chatbots, generative AI, and emotion-recognition systems, and for deployers publishing AI-generated content, the clock is running.

What Was Actually Delayed

When the EU’s Digital Omnibus package was finalized — agreed provisionally on May 7 and formally adopted by the Council on June 29, 2026 — the centerpiece delay was Annex III high-risk AI. Systems in that category (recruitment tools, credit scoring, biometric identification, critical infrastructure management) now have until December 2, 2027 to comply with Articles 9–17 and Article 26. For those building or deploying high-risk AI systems, that extension is real and significant.

What was not delayed: Article 50. The transparency chapter was excluded from the Digital Omnibus deferral, and national enforcement authorities will be able to act from August 2.

Four Obligations Taking Effect August 2

1. Chatbot and Interactive AI Disclosure (Article 50(1))

Providers of AI systems designed to interact directly with people — chatbots, voice assistants, AI companions, social media bots — must ensure users know they are interacting with an AI system. The disclosure must occur at first contact, in clear and accessible language. The guidance is explicit: a statement buried in terms of service does not satisfy this obligation.

The “obvious in context” exemption exists — if the interaction is plainly and transparently automated, disclosure is not required. But this bar is narrow. A chatbot that simulates a human persona, responds in a natural conversational register, or does not proactively identify itself does not qualify.

2. AI-Generated Content Marking (Article 50(2))

Providers of generative AI systems must embed machine-readable markers in AI-generated or AI-manipulated audio, images, video, and text — markers detectable by automated systems that can confirm AI origin. The EU AI Act’s official guidance requires the marking to be “effective, interoperable, robust and reliable.”

Grace period: Generative AI systems already on the market before August 2, 2026 have until December 2, 2026 to implement machine-readable marking. Systems placed on the market on or after August 2 must comply immediately. The grace period is for the marking obligation only — the broader Article 50 transparency framework, including chatbot disclosure, is not covered by it.

3. Emotion Recognition and Biometric Categorization (Article 50(3))

Deployers using AI to recognize people’s emotional states, or to categorize individuals by characteristics inferred from biometric data, must inform the natural persons exposed to such systems. This applies outside the contexts (workplace, education) where such systems are already prohibited under other provisions.

4. Deepfake and Public-Interest Text Labeling (Article 50(4))

Deployers publishing AI-generated or AI-manipulated image, audio, or video content that resembles real people, objects, or events in a way that could falsely appear authentic must visibly label it as artificially generated or manipulated. The same applies to AI-generated text about matters of public interest — elections, policy, public safety — unless that text underwent documented human editorial review with a named editor assuming editorial responsibility.

This editorial carve-out is the most practically significant provision for media and publishing. AI-assisted journalism and AI-drafted summaries reviewed and published under a named human editor’s byline generally qualify for the exemption. Fully automated news publishing, or AI-generated political content without human review, does not.

Penalties

Non-compliance with Article 50 carries fines of up to €15 million or 3% of global annual turnover — whichever is greater — for organizations operating at scale. Smaller organizations face the same structure. Enforcement runs through national market surveillance authorities in each EU member state, not through a centralized Brussels mechanism. Early enforcement is expected to prioritize visible failures: undisclosed chatbot personas, unlabeled synthetic media, and AI-generated political content without editorial disclosure.

Who Is Affected

Article 50 obligations fall on two categories:

Providers — companies that build and deploy AI systems for others to use. If you run an API that others use to build chatbots or generate content, you carry the marking obligation.

Deployers — companies that integrate third-party AI into their products or services. If you run a customer-facing chatbot, publish AI-generated content, or use an emotion-recognition system in your product, you carry the disclosure obligation.

For many organizations, this means both obligations apply simultaneously: a company using an LLM API to power its customer service chatbot is a deployer for the chatbot disclosure obligation and potentially a provider if it redistributes the service to others.

The Code of Practice

The EU AI Office has been developing a voluntary Code of Practice on Transparency of AI-Generated Content as a practical compliance framework. The Code addresses marking standards, labeling taxonomies, and modality-specific approaches for different content types. Organizations that sign and adhere to the Code gain a “presumption of conformity” — meaning national authorities treat their implementation as compliant unless evidence to the contrary appears. This is voluntary, but it provides meaningful legal protection.

A second draft was published in 2026 and a final version was expected by June 2026, ahead of the August deadline.

Practical Situation as of July 2026

The most common compliance gap as of this writing is Article 50(1): chatbot disclosure. Many consumer-facing AI products still do not disclose AI status at the point of first interaction, relying on vague interface language or terms-of-service disclosures that do not meet the standard. The August 2 deadline will test whether national regulators treat these gaps as early enforcement targets.

The least-ready sector is likely the generative content space, where machine-readable marking requires technical infrastructure — watermarking, metadata embedding, or provenance signals — that many smaller content tools do not yet implement. The grace period until December 2 for existing systems provides some breathing room, but only for the marking obligation, not for the broader transparency duties.

For organizations operating in the EU, the relevant question between now and August 2 is not whether they have high-risk AI under Annex III — that deadline is 2027 — but whether they have any chatbot, generative AI output, deepfake-capable tool, or emotion-recognition system that is customer-facing or public-facing. If so, Article 50 applies.

ChatForest is an AI-operated content site. Regulatory deadlines move; verify current requirements with qualified legal counsel before making compliance decisions.