Starting August 2, 2026, Article 50 of the AI Act imposes transparency obligations on all users and providers of generative AI systems. The Omnibus agreement only shifts machine-readable marking to December 2, 2026; other measures remain effective from August 2026.
Article 50 of the European AI Act imposes transparency requirements on all actors who use or deploy generative AI systems in four specific situations. Unlike obligations reserved for high-risk systems, this provision applies regardless of the sector or size of the organization.
Its scope is vast: any interaction with an AI, any content produced or modified by an AI system, or the use of emotion recognition or biometric categorization falls within the scope of the article. Conversational assistants, image, text, and video generators, as well as voice or facial analysis devices, are therefore affected.
Why Article 50 of the AI Act concerns all generative AI users
The main objective is to ensure that the user is informed when interacting with an AI or viewing synthetic content. This is intended to strengthen trust in AI systems and limit the risks of manipulation or disinformation.
Contrary to popular belief, Article 50 is not limited to high-risk systems. Any company or organization that uses generative AI, even occasionally, is affected.
The four transparency obligations detailed
Article 50 identifies four cases in which transparency is mandatory, each involving specific measures adapted to the level of risk.
1. Informing during an interaction with an AI system
Any AI system designed to interact with natural persons must signal to the user that they are interacting with an artificial intelligence. This particularly concerns chatbots, voice assistants, and other automated dialogue tools.
Information must be clear, easily accessible, and communicated before the exchange begins, in the form of a message, icon, or voice alert, depending on the medium used.
2. Reporting synthetic or modified content
Any content generated or altered by an AI must be identifiable as such. This concerns images, videos, texts, or sounds produced or retouched by an AI system.
Article 50(2) imposes machine-readable marking, i.e., integrated metadata allowing for automatic identification. This requirement raises significant technical challenges and requires shared standards.
3. Labeling deepfakes
Highly realistic AI-generated content, such as deepfakes, must be explicitly marked as artificial. The goal is to prevent disinformation and identity theft.
Labeling must be visible and understandable to any user, for example via a watermark or a textual mention.
4. Mention on AI texts of general interest
Texts generated by AI and disseminated publicly, particularly on topics of general interest (politics, health, economy, etc.), must carry a mention specifying their artificial origin.
This obligation also extends to emotion recognition or biometric categorization systems, which must inform the persons concerned of their use.
**Summary of obligations:**
- **Art. 50(1)**: Information on interaction with an AI
- **Art. 50(2)**: Machine-readable marking of synthetic content
- **Art. 50(4)**: Labeling of deepfakes and texts on public interest topics
What the Omnibus agreement changed (and what remains unchanged)
The provisional agreement of May 7, 2026, known as Omnibus, between the European Parliament and the EU Council, introduced a partial postponement regarding Article 50. This point is often misunderstood.
A 4-month delay limited to machine-readable marking
According to the Omnibus, providers of generative AI systems already marketed before August 2, 2026, benefit from an additional 4-month period to comply with the machine-readable marking obligation (Article 50(2)), i.e., until December 2, 2026.
This reprieve only concerns systems existing on the date of entry into force. Any new system must be compliant as soon as it is placed on the market.
Other obligations remain unchanged
All other measures provided for by Article 50, including information on the use of an AI, the labeling of deepfakes, and the mention on texts of public interest, apply in full from August 2, 2026. No postponement has been granted for these points.
The reprieve does not apply to deployers either, who must continue to inform persons exposed to AI content or emotion recognition systems.
A political compromise not yet official
The Omnibus agreement remains provisional, as the final text has not yet been published in the Official Journal. Dates, such as December 2, 2026, are therefore to be confirmed. Nevertheless, the initial version of Article 50 will indeed apply from August 2, 2026.
In summary: only machine-readable marking benefits from an additional delay; all other obligations take effect from August 2, 2026.
Providers and deployers: respective obligations under Article 50
The requirements of Article 50 do not only concern providers of AI systems. Deployers, i.e., organizations that use these systems, also have specific responsibilities.
Obligations of providers
Providers must:
- Warn users that they are interacting with an AI (Article 50(1))
- Integrate machine-readable marking on synthetic content (Article 50(2))
- Provide detailed technical documentation on the system's operation and limitations
These obligations must be taken into account from the design stage and integrated into the user interface.
Obligations of deployers
Deployers are responsible for:
- Informing any person exposed to deepfakes or AI content (Article 50(4))
- Ensuring transparency when using emotion recognition or biometric categorization systems
- Ensuring that providers comply with their obligations, particularly regarding content marking
They must also train their teams and implement internal procedures to ensure compliance.
**Reference articles:**
- Art. 3(3): Definition of provider
- Art. 3(4): Definition of deployer
- Art. 28: Responsibilities of deployers
Code of Practice and guidelines: current status
The European Commission is developing a Code of Practice to support stakeholders in implementing transparency, but all technical details remain under discussion.
Code of Practice: schedule and content
A first draft of the Code of Practice focused on the transparency of AI content was published on December 17, 2025, by the European Commission, via the AI Office, mainly regarding Articles 50(2) and 50(4).
An enriched version was proposed in March 2026, after consultation with stakeholders. The final publication is scheduled for June 2026, two months before the August 2, 2026 deadline.
This Code, although non-binding, will become a reference for the market. Companies are advised to comply with it in anticipation of regulatory expectations.
Guidelines under consultation
A targeted consultation on operational guidelines is open until June 3, 2026. It covers:
- Technical standards for machine-readable marking
- Methods for labeling deepfakes
- Information modalities according to media (web, mobile, voice, etc.)
The results of this consultation will guide the drafting of the final guidelines, expected for the summer of 2026.
Companies must closely monitor the evolution of these documents to adjust their practices.
For further reading:
Legal references and regulatory framework
Article 50 is part of a broader European legal framework, including the GDPR and other key texts.
- **AI Act – Art. 50**: Transparency for AI systems
- **AI Act – Art. 50(2)**: Machine-readable marking of AI content
- **AI Act – Art. 50(4)**: Labeling of deepfakes and public interest texts
- **AI Act – Art. 5**: Prohibited practices (nudifiers, CSAM)
- **GDPR – Art. 13-14**: Information of persons and articulation with AI transparency
- **AI Act – Art. 99**: Penalties for non-compliance
Preparing for Article 50 before August 2, 2026: operational method
Compliance with Article 50 requires more than just adding mentions: it is a comprehensive approach involving organizational and technical adaptations.
1. Identify the systems concerned
Carry out an exhaustive mapping of the AI systems used or deployed in your structure:
- Chatbots and voice assistants
- Content generators (text, image, video, audio)
- Emotion recognition or biometric categorization systems
For each system, specify your role (provider or deployer) and the associated obligations.
2. Verify interface transparency
Check that user interfaces clearly inform about the artificial nature of the system. This may involve:
- Adding explicit mentions or icons
- Adapting interaction paths to include prior notification
- Updating terms of use
3. Collaborate with providers
As a deployer, require your providers to comply with obligations, particularly content marking. To do this:
- Integrate compliance clauses into contracts
- Verify the presence of metadata on generated content
- Implement quality control processes
4. Train and raise awareness among teams
Ensure team training on transparency issues and Article 50 requirements:
- Product and tech teams: integrate transparency by design
- Marketing and communication: avoid any ambiguity
- Legal and compliance: ensure rule application
5. Anticipate regulatory developments
Stay informed about the publication of the Code of Practice and guidelines. Participate in public consultations and monitor decisions by national authorities (e.g., CNIL) to adjust your practices.
---