EU AI Act Article 50 Transparency Deadline: What Lives August 2 2026

Article 50 of the EU AI Act is the transparency layer. Even after the May 2026 Digital Omnibus pushed the high-risk AI deadlines out, Article 50 mostly stayed put. The transparency obligations come into force on August 2, 2026. The one exception is the watermarking requirement, which slipped four months to December 2, 2026. If you run a chatbot, deploy synthetic media, or publish AI-generated text into the EU, August 2 is your live date.

This is a clean walkthrough of what Article 50 actually requires, the new dates after the AI Omnibus, the exceptions that survive, and the €15 million / 3% turnover penalty regime that backs it all up.

Article 50 obligations on the calendar

August 2, 2026: Article 50(1) chatbot disclosure, Article 50(3) emotion-recognition and biometric-categorisation disclosure, Article 50(4) deepfake disclosure, and the Article 50(4) public-interest text disclosure all come into force as originally scheduled (artificialintelligenceact.eu / Skadden).

December 2, 2026: Article 50(2) watermarking and machine-readable marking obligations for AI-generated content take effect. Under the Omnibus grandfathering rule, generative AI systems placed on the EU market before August 2, 2026 are not required to comply with watermarking until December 2, 2026 (Latham & Watkins).

That split matters. A chatbot you ship in July 2026 must carry the AI-disclosure message by August 2, 2026, but its watermarking obligation does not bite until December 2, 2026.

Article 50(1): chatbot disclosure

Providers of AI systems that interact directly with natural persons must design those systems so users are informed they are interacting with an AI (artificialintelligenceact.eu). The Commission's draft guidance published May 7, 2026 is unusually specific about how that disclosure should happen.

Disclosure must occur in the user interface itself, not buried in terms and conditions. It must happen on first interaction, before the conversation has begun (Skadden). And it must be tailored to the audience. Disclosures aimed at children or the elderly need to be more elaborate. Disclosures for tools used only by professional developers, such as coding assistants behind a paid SDK, can be lighter.

The exception is narrow. AI systems authorised by law to detect, prevent, investigate, or prosecute criminal offences are exempt from Article 50(1) disclosure, but only when the system is not made available for the public to report a criminal offence (artificialintelligenceact.eu). A police AI hotline open to the public still needs to disclose.

What compliant chatbot disclosure looks like in practice: a clear, unmissable line on the first screen of any AI-powered chat surface that says some variation of "you are chatting with an AI assistant." No buried fine print. No reliance on the user inferring it from the bot's name.

Article 50(2): watermarking and machine-readable marking

Providers of AI systems that generate synthetic audio, image, video, or text content must ensure outputs are marked in a machine-readable format and detectable as artificially generated or manipulated (artificialintelligenceact.eu / Skadden).

This is the technically heaviest part of Article 50. The required solution must be effective, interoperable, robust, and reliable. In practice that means two things working together: (i) a machine-readable watermark embedded in the output and (ii) a tool that lets humans, including downstream platforms, verify whether content is AI-generated.

C2PA Content Credentials is the most widely deployed candidate framework. Google's SynthID and Adobe's Content Authenticity Initiative work in adjacent spaces. The Commission has not blessed any specific technical standard; the obligation is outcome-based.

Exceptions to watermarking are narrow. Article 50(2) does not apply where the AI system performs an assistive function for standard editing, where it does not substantially alter the input data provided by the deployer, or where the use is authorised by law for law enforcement (artificialintelligenceact.eu). A spell-check tool is out. A generative tool that materially changes the input is in.

Article 50(3): emotion recognition and biometric categorisation

Deployers of emotion-recognition systems or biometric-categorisation systems must inform individuals of the system's operation (artificialintelligenceact.eu). The disclosure obligation falls on the deployer, not just the provider. If your business runs an emotion-recognition tool to gauge call-centre sentiment, you must disclose to the people whose voices you analyse, regardless of who built the model.

If the system also processes personal data, GDPR transparency obligations under Articles 13 and 14 apply in parallel. The combined effect is that operators need a layered notice: an Article 50(3) AI-operation disclosure plus a GDPR data-processing disclosure.

Article 50(4): deepfakes and public-interest text

Two distinct obligations sit under Article 50(4).

Deepfake disclosure. Deployers of AI systems generating image, audio, or video content that constitutes a deepfake of a real, identifiable person must disclose that the content is AI-generated (artificialintelligenceact.eu). The exception is law enforcement use authorised by law. The artistic-and-satirical carve-out is narrower than people often assume: the disclosure must still be made, but in a way that does not hamper enjoyment of the work. A deepfake parody video still needs a label. The label can be in the credits or a corner watermark.

Fanciful deepfakes that do not depict real people, like AI-generated cartoons of fictional characters, do not require disclosure (Skadden).

Public-interest text disclosure. Deployers generating text to inform the public on matters of public interest must label the text as AI-generated. That covers news content, corporate public reports, academic papers, government social media posts (Skadden).

The practical exception is significant. The obligation does not apply where AI-generated text has undergone human editorial review and a natural or legal person holds editorial responsibility for the publication (artificialintelligenceact.eu). A newsroom that runs every AI-drafted article past a human editor and stamps a byline on it is not required to add an AI-generated label. The minute you publish AI text with no human editor in the loop, you are.

The penalty regime

Violations of Article 50 transparency obligations carry administrative fines of up to €15 million or 3% of total annual worldwide turnover, whichever is higher (Latham & Watkins / AI CERTs).

That is the middle tier of the AI Act's three penalty bands. The Article 5 prohibited-practice tier is €35M / 7%. The catch-all for other violations is €7.5M / 1%. Article 50 sits in between because the transparency obligations are the structural counterweight to the Act's relatively permissive treatment of generative AI in the lower-risk categories.

Enforcement will come from national market-surveillance authorities and ultimately from the AI Office for cross-border cases. Expect the first wave of investigations to focus on consumer-facing chatbots and large-scale generative-AI platforms where non-compliance is visible without forensic effort.

Overlap with US state chatbot laws

Article 50 is not the first chatbot disclosure law to bite consumer-facing AI. California's chatbot disclosure law under Business & Professions Code section 17941 has required disclosure for bots that communicate with California users since 2019, with a narrow carve-out for entertainment and news media. Utah's Artificial Intelligence Policy Act, effective May 1, 2024, requires generative AI disclosures, with the Utah Division of Consumer Protection holding enforcement authority. Operators of generative AI products serving both US and EU users should treat Article 50 as the strictest common floor and design a single disclosure layer that satisfies California, Utah, and the EU at once.

A compliance checklist for Article 50

If you want a single-page operating list:

1. Inventory every AI surface your product exposes to EU users. Chatbots, voice agents, image/video generators, text generators, emotion-recognition and biometric-categorisation tools.
2. For each chatbot or interactive AI, add an unmissable disclosure on first interaction. Plain language. No fine print.
3. For each generative AI tool, plan your watermarking architecture now. C2PA, SynthID, or equivalent. Target deployment by December 2, 2026 with a buffer of at least 30 days for QA.
4. For deepfake-capable tools, add a default-on AI-generated label. Make the toggle for artistic/satirical use cases require explicit user confirmation.
5. For AI-text publishing workflows that touch public-interest content, document the human editorial review step. Keep an audit trail of who edited what and when.
6. For emotion-recognition and biometric-categorisation deployments, add the Article 50(3) disclosure to your existing GDPR transparency notice. Layer, do not duplicate.
7. Train customer-facing teams on the disclosure language. Most Article 50 enforcement risk comes from teams improvising messaging in the field.

If you are also planning for the broader EU AI Act timeline, our EU AI Act deadline extension explainer covers the HRAIS dates and the Omnibus changes. The EU AI Act framework page and our EU AI Act compliance checklist for 2026 round out the picture.

Article 50 is not the most technically complex piece of the AI Act. It is one of the easiest to fail. The compliance is visible on screen. So is the non-compliance.

References

1. artificialintelligenceact.eu. Article 50 — Transparency obligations for providers and deployers. https://artificialintelligenceact.eu/article/50/
2. Skadden. AI Act State of Play. May 12, 2026. https://www.skadden.com/insights/publications/2026/05/ai-act-state-of-play
3. Latham & Watkins. AI Act Update: EU Resolves to Change Rules and Extend Deadlines. May 13, 2026. https://www.lw.com/en/insights/ai-act-update-eu-resolves-to-change-rules-and-extend-deadlines
4. AI CERTs. EU Omnibus Extends HRAIS Deadlines. https://www.aicerts.ai/news/ai-regulation-eu-omnibus-extends-hrais-deadlines/
5. European Commission. Regulatory framework for AI. https://digital-strategy.ec.europa.eu/en/policies/regulatory-framework-ai