Article 50 of the AI Act: what actually applies on 2 August 2026
The Omnibus agreement of 7 May 2026 has created confusion: the transparency obligations under Article 50 have not been postponed to 2027. Only the machine-readable marking requirement benefits from a stay of execution until 2 December 2026. Explanation.
Why Article 50 applies to all businesses using generative AI
Article 50 of the EU AI Regulation imposes transparency obligations for AI systems used in four specific scenarios. Unlike the provisions on high-risk AI systems, these rules apply to all actors, regardless of their sector or the size of their organisation.
The scope is broad: any interaction with an AI system, any AI-generated content, or any use of emotion recognition or biometric categorisation is covered. In practice, this includes chatbots, image, text or video generation tools, as well as facial or voice analysis systems.
The objective is twofold: to inform users that they are interacting with AI, and to enable the identification of synthetic content. These obligations aim to strengthen trust in AI technologies while limiting the risks of manipulation or disinformation.
The 4 transparency obligations in detail
Article 50 distinguishes four scenarios where transparency is required. Each imposes specific measures, tailored to the identified risks.
1. Disclosure of interactions with an AI system
Any AI system designed to interact directly with natural persons must inform those persons that they are communicating with AI. This applies to chatbots, voice assistants, or any other automated dialogue tool.
The information must be clear, accessible and provided before the interaction begins. It may take the form of a text message, an icon, or a voice notification, depending on the channel used.
2. Labelling of synthetic or manipulated content
Content generated or modified by an AI system must be identifiable as such. This includes images, videos, texts or audio files produced or altered by AI.
Article 50(2) requires machine-readable marking, i.e. metadata embedded in the file enabling automatic detection. This obligation is at the heart of current debates, as it requires standardised technical solutions.
3. Identification of deepfakes
Realistic content generated by AI, such as deepfakes, must be clearly labelled as artificial. This measure aims to prevent the risks of disinformation or identity theft.
The labelling must be visible and understandable to the end user. It may take the form of a watermark, a text notice, or a combination of both.
4. Information on AI-generated texts on matters of public interest
AI-generated texts disseminated publicly, particularly those dealing with matters of general interest (politics, health, finance, etc.), must be accompanied by a notice indicating their artificial origin.
This obligation also applies to emotion recognition or biometric categorisation systems, which must inform the persons concerned of their use.
What the Omnibus agreement changed (and what it did not)
The provisional agreement of 7 May 2026 between the European Parliament and the Council of the EU introduced a stay of execution for part of the obligations under Article 50. However, this modification is often misinterpreted.
The 4-month stay: only for machine-readable marking
The Omnibus agreement provides that providers of generative AI systems already placed on the market before 2 August 2026 benefit from an additional four-month period to comply with the machine-readable marking obligation (Article 50(2)). This measure would apply until 2 December 2026.
This stay applies only to systems put into service before the initial deadline. New systems must comply from their launch.
What remains unchanged: the 4 transparency obligations
The other provisions of Article 50, including the disclosure of interactions with AI, the labelling of deepfakes and the information on texts of public interest, remain applicable from 2 August 2026. No changes have been made to these obligations.
The Omnibus agreement also did not postpone the deadlines for deployers, who must inform persons exposed to AI-generated content or emotion recognition systems.
A provisional agreement, not yet final
The Omnibus agreement is a political compromise, but the consolidated text has not yet been published in the Official Journal. The dates mentioned (2 December 2026) are therefore subject to confirmation. However, Article 50 in its initial version is already in force and will apply on 2 August 2026.
Providers vs deployers: who must do what under Article 50
The obligations under Article 50 do not rest solely on AI system providers. Deployers also have a key role to play, particularly in informing end users.
Responsibilities of providers
Providers are required to:
- Inform users that they are interacting with an AI system (Article 50(1)).
- Implement machine-readable marking for synthetic content (Article 50(2)).
- Provide clear technical documentation on the system's functionalities and limitations.
These obligations apply from the design stage of the system and must be integrated into user interfaces.
Responsibilities of deployers
Deployers, i.e. organisations that use an AI system in the course of their activities, must:
- Inform persons exposed to deepfakes or AI-generated content (Article 50(4)).
- Ensure that emotion recognition or biometric categorisation systems are used transparently.
- Verify that providers comply with their obligations, particularly regarding content marking.
Deployers are also responsible for training their teams and implementing internal processes to ensure compliance with transparency rules.
Code of Practice and guidelines: where do we stand?
The European Commission is working on a Code of Practice to help stakeholders comply with transparency obligations. However, technical details remain under discussion.
A Code of Practice in the finalisation stage
On 17 December 2025, the Commission published a first draft of the Code of Practice on transparency of AI-generated content. This document, developed by the AI Office, focuses on Articles 50(2) and 50(4).
A second draft was published in March 2026, incorporating stakeholder feedback. The final version is expected in June 2026, two months before the 2 August deadline.
Although this Code is voluntary, it is likely to become a benchmark for market participants. Businesses are encouraged to draw inspiration from it to anticipate regulators' expectations.
Guidelines still under consultation
The Commission has launched a targeted consultation on operational guidelines for Article 50. Open until 3 June 2026, this consultation aims to gather feedback on practical implementation modalities.
The topics covered include:
- Technical standards for machine-readable marking.
- Best practices for labelling deepfakes.
- Modalities for informing users in different contexts (web, mobile, voice, etc.).
The results of this consultation will feed into the final guidelines, expected in summer 2026.
Legal framework and official sources
Article 50 is part of a broader regulatory framework, which includes the GDPR and other European texts. Here are the main references to be aware of.
Further reading:
How to prepare before 2 August 2026
Compliance with Article 50 is not just about adding a label. It involves overhauling internal processes and close collaboration with providers.
1. Identify the systems concerned
Start by mapping all AI systems used or deployed in your organisation. This includes:
- Chatbots and voice assistants.
- Content generation tools (text, image, video, audio).
- Emotion recognition or biometric categorisation systems.
For each system, identify its role (provider or deployer) and the resulting obligations.
2. Audit user interfaces
Check that the interfaces of AI systems clearly inform users of their artificial nature. This may require:
- Adding text notices or icons.
- Modifying interaction flows to include prior notification.
- Updating terms and conditions of use.
3. Collaborate with providers
If you are a deployer, require your providers to comply with their obligations, particularly regarding content marking. This may involve:
- Revising contracts to include compliance clauses.
- Verifying the metadata of generated content.
- Implementing quality control processes.
4. Train teams
Raise awareness among your teams about transparency issues and the obligations under Article 50. This concerns:
- Product and development teams, to integrate transparency from the design stage.
- Marketing and communications teams, to avoid misleading practices.
- Legal and compliance teams, to ensure rule compliance.
5. Anticipate regulatory developments
Monitor the evolution of the Code of Practice and guidelines. Participate in public consultations to contribute to the development of technical standards.
Also plan active monitoring of decisions by national authorities, such as the CNIL in France, which may clarify expectations regarding transparency.
Identify your obligations in 3 minutes
Our free assessment helps you understand what Article 50 means for your organisation.
Frequently asked questions
Answers to the most common questions about Article 50 of the AI Act.
No. The Omnibus agreement only concerns the machine-readable marking of synthetic content (Article 50(2)), and only for systems already placed on the market before 2 August 2026. This stay would apply until 2 December 2026. The other transparency obligations (disclosure of interactions, labelling of deepfakes, etc.) remain applicable from 2 August 2026.
Penalties under Article 99 of the AI Act apply in case of non-compliance. These can amount to up to €7.5 million or 1.5% of global turnover, whichever is higher. National authorities, such as the CNIL in France, are responsible for monitoring compliance and imposing penalties.
Yes. Article 50 applies to all organisations, regardless of size, as soon as they use or deploy an AI system in one of the four covered scenarios. No exemption is provided for small businesses or startups.
The transparency obligations under Article 50 complement those of the GDPR. For example, informing users about the use of an AI system can be integrated into the information notices required by Articles 13 and 14 of the GDPR. Both texts aim to strengthen user trust, but with different approaches: the GDPR focuses on personal data, while the AI Act covers all interactions with AI systems.
Several technical solutions are being developed to facilitate compliance:
- Watermarking libraries for marking synthetic content (e.g., C2PA, Adobe Content Credentials).
- Automatic detection tools for AI-generated content (e.g., deepfake detectors).
- Frameworks for integrating transparency notices into user interfaces (e.g., plugins for chatbots).
The Code of Practice and the Commission's guidelines should clarify the standards to be adopted.