AI Act Transparency Obligations: What You Must Do Now

The EU AI Act’s transparency obligations under Article 50 apply from 2 August 2026. Here is exactly who must do what, when, and how, with practical steps and…

· 5 min read

What the AI Act transparency obligations are and who they apply to

The EU AI Act’s Article 50 transparency obligations apply from 2 August 2026. They require providers and deployers of certain AI systems to inform people when they are interacting with AI, mark AI-generated content, and disclose specific uses like deepfakes or biometric categorisation. These duties are legal requirements, not guidelines, and breaches can lead to fines of up to EUR 15 million or 3% of worldwide annual turnover, whichever is higher, with lower ceilings for SMEs.

Article 50 contains four distinct duties. First, providers must tell people when they are interacting with an AI system, such as a chatbot or virtual assistant. Second, providers of generative AI systems must mark synthetic audio, image, video and text output in a machine-readable format so it can be detected as artificially generated or manipulated. Third, deployers of emotion recognition or biometric categorisation systems must inform the people exposed to them. Fourth, deployers must disclose deepfakes and AI-generated or manipulated text published to inform the public on matters of public interest.

Marking AI-generated content: the Article 50(2) duty for providers

Providers of generative AI systems must ensure that synthetic audio, image, video and text output is marked in a machine-readable format so it can be detected as artificially generated or manipulated. The marking must be effective, interoperable, robust and reliable as far as technically feasible. This means the mark should survive common edits, compression, or sharing without being stripped or altered.

The Digital Omnibus on AI gives generative AI systems placed on the market before 2 August 2026 until 2 December 2026 to meet this marking duty. Content generated and published before 2 August 2026 does not need to be labelled retroactively. There is a narrow exemption for outputs used in business-to-business or industrial contexts, but artistic, creative, satirical or fictional works only need disclosure in a way that does not hamper the display or enjoyment of the work.

Disclosing deepfakes and AI-generated public interest text: the Article 50(4) duty for deployers

Deployers must disclose deepfakes and AI-generated or manipulated text published to inform the public on matters of public interest. This duty applies to anyone publishing such content, whether they are a media outlet, a business, or an individual. The disclosure must be clear, distinguishable, provided at the latest at first interaction or exposure, and accessible to the intended audience.

There is an exemption for AI-generated text that has undergone human review with editorial responsibility. This means that if a human editor has reviewed and taken responsibility for the content, it does not need to be disclosed as AI-generated. However, the exemption only applies to the text disclosure duty and does not affect other obligations under Article 50.

Emotion recognition and biometric categorisation: the Article 50(3) duty for deployers

Deployers of emotion recognition or biometric categorisation systems must inform the people exposed to these systems. This duty applies regardless of whether the system is used in a public or private context. The information must be provided in a clear and distinguishable way, at the latest at first exposure, and must be accessible to the people affected.

The requirement does not apply to systems that do not involve emotion recognition or biometric categorisation. It also does not apply to systems used for purely internal purposes where no individuals are exposed to the system’s outputs.

How to give the information: Article 50(5) and practical steps

Article 50(5) sets out how the information must be given: it must be clear, distinguishable, provided at the latest at first interaction or exposure, and accessible. This means the disclosure should be easy to understand, visually or audibly distinct from other information, and available to the intended audience without unnecessary barriers.

For marking AI-generated content, a multi-layered approach is recommended. This includes metadata embedded in the file, such as C2PA manifests (also known as Content Credentials), and invisible watermarks that persist even if metadata is stripped. The EU has also created a voluntary set of icons for labelling AI content, which can be used to visually indicate whether content is fully AI-generated, partially AI-modified, or involves AI in some other way.

The Code of Practice and Guidelines: what they mean for compliance

The Code of Practice on Transparency of AI-Generated Content was finalised on 10 June 2026 and assessed as adequate by the Commission and the AI Board. Signing is voluntary, but signatories can rely on it to demonstrate compliance with Article 50(2), (4) and (5). The Code asks for a multi-layered approach to marking, with metadata that is digitally signed and time-stamped in a secure, tamper-evident way.

The Commission published Guidelines on Article 50 on 20 July 2026. These are non-binding but provide practical advice on how to meet the transparency obligations. The Guidelines do not change the legal requirements but offer clarification on how to interpret and apply them.

What you should do now: a practical checklist

First, determine whether you are a provider or deployer of AI systems covered by Article 50. Providers are responsible for marking AI-generated content, while deployers must disclose deepfakes, AI-generated public interest text, and the use of emotion recognition or biometric categorisation systems.

For providers of generative AI systems, implement a marking solution that is machine-readable, effective, interoperable, robust and reliable. Consider using C2PA manifests for images and video, and combine metadata with invisible watermarks to ensure the mark persists. For audio and text, explore watermarking solutions that embed the mark directly into the content.

For deployers, review your use of AI systems and identify any that require disclosure under Article 50(3) or (4). Ensure that disclosures are clear, distinguishable, and provided at the latest at first interaction or exposure. Use the EU’s voluntary icons where appropriate to visually label AI content.

Finally, keep up to date with the Code of Practice and the Commission’s Guidelines, as they provide valuable guidance on best practices for compliance. While the Code is voluntary, adhering to it can help demonstrate compliance with Article 50.

Frequently asked questions

Do the AI Act transparency obligations apply to content created before 2 August 2026?

No. Content generated and published before 2 August 2026 does not need to be labelled retroactively. However, any new content created or published after this date must comply with the marking and disclosure duties.

What is the deadline for generative AI systems placed on the market before 2 August 2026?

The Digital Omnibus on AI gives these systems until 2 December 2026 to meet the Article 50(2) marking duty. This extension does not apply to other transparency obligations.

Are there exemptions for artistic or creative works?

Yes, but only for the marking duty. Artistic, creative, satirical or fictional works only need disclosure in a way that does not hamper the display or enjoyment of the work. This exemption does not apply to other obligations under Article 50.

What happens if I fail to comply with Article 50?

Breaching the Article 50 obligations can lead to fines of up to EUR 15 million or 3% of worldwide annual turnover, whichever is higher. Lower ceilings apply to SMEs.

Can I use C2PA or watermarking to comply with the marking duty?

Yes. Providers can use C2PA manifests (Content Credentials) for images and video, and combine them with invisible watermarks to ensure the mark persists. The marking must be effective, interoperable, robust and reliable as far as technically feasible. The Code of Practice recommends a multi-layered approach.

Put provenance on your next file

L'AItelier signs images and video with C2PA, adds an invisible watermark, and issues a certificate anyone can check. Your file is deleted as soon as the certificate is issued.

General information about the law as it stood when this was published, not legal advice for your situation.