AI Transparency Article 50, a guide for businesses.
On 2 November 2026, the AI Act's transparency obligations come into force. Providers and deployers must identify relevant systems, implement visible markings, and document their processes. Here's how to prepare without delay.
Which AI systems are covered by Article 50?
Article 50 targets three categories of AI systems, regardless of their risk level.
Unlike the obligations for high-risk AI systems, transparency requirements apply to widely used tools. The Regulation distinguishes:
- Generative AI systems producing text, audio or video,
- AI systems interacting with humans through conversational interfaces,
- Systems manipulating existing content to create deepfakes.
The AI Office clarifies that automatic translation or summarisation tools are not covered unless they generate original content. The official glossary provides concrete examples for each category.
Who must comply? Providers vs deployers
Obligations differ depending on the role in the AI value chain.
Providers of AI systems must:
- Integrate synthetic content marking mechanisms,
- Provide technical documentation describing these mechanisms,
- Inform deployers of their residual obligations.
Deployers, for their part, must:
- Verify that the systems used properly integrate transparency mechanisms,
- Train their teams to recognise and use these markings,
- Document their supervision processes.
The AI Office published guidelines in June 2026 clarifying these shared responsibilities. Businesses using SaaS solutions must contractually verify that their provider complies with these obligations.
How to mark synthetic content?
The Regulation requires both visible and machine-readable marking.
Two methods are authorised for visible marking:
- A watermark embedded directly in the content,
- A clear textual statement indicating the synthetic nature of the content.
For machine-readable marking, technical standards are being finalised by CEN-CENELEC. The AI Office recommends using the following formats:
- For images: EXIF tags or XMP metadata,
- For videos: metadata in the stream or ancillary files,
- For text: HTML tags or JSON-LD metadata.
The CNIL published a practical guide in April 2026 detailing these methods with concrete examples. Businesses must adapt their production and dissemination tools to integrate these markings.
What are the risks of non-compliance?
Sanctions vary depending on the severity of the breach and the actor's role.
For providers, fines may reach:
- Up to €15 million or 3% of global turnover,
- With possible increases for repeated breaches.
For deployers, sanctions include:
- Administrative fines proportionate to the company's size,
- Orders to comply within a set timeframe,
- Temporary suspension of the use of non-compliant systems.
According to several studies, reputational risks could be more costly than financial penalties. Businesses involved in scandals related to unmarked synthetic content may face lasting loss of trust from customers and partners.
The AI Office has indicated that initial inspections will target high-impact public sectors: media, advertising, financial services and public administration. The official timeline sets out key dates for each type of obligation.
Concrete examples by sector
Each industry must adapt transparency obligations to its specific uses.
Media and communications
Press agencies and media organisations must:
- Systematically mark AI-generated images and videos,
- Clearly indicate in articles when AI tools are used for writing,
- Train journalists in transparency best practices.
Example: The Le Monde Group announced in 2025 the automatic integration of watermarks in its image production tools.
Financial services
Banks and insurers must:
- Inform clients when automated decisions are made,
- Mark AI-generated reports for credit analysis,
- Document human supervision processes.
Example: BNP Paribas deployed in 2026 a system for marking financial analysis report summaries.
Healthcare
Healthcare establishments must:
- Clearly indicate when diagnoses are AI-assisted,
- Mark automatically generated medical reports,
- Train medical staff in the transparent use of these tools.
Example: AP-HP implemented in 2025 a transparency protocol for its radiology diagnostic support tools.
Are your AI tools compliant?
Identify your exact obligations under the AI Act with our free assessment.
Frequently asked questions
Answers to the most common questions about transparency obligations.
Yes, Article 50 applies to all relevant AI systems, including those used internally. The Regulation does not distinguish between internal and external use for transparency obligations. Businesses must therefore mark synthetic content even when it is intended for strictly internal use, such as automated reports or AI-generated presentations.
Visible marking is intended for human users. It may take the form of a watermark, textual statement or recognisable icon. Machine-readable marking, on the other hand, is embedded in the content's metadata and enables automatic detection by technical tools. Both types of marking are mandatory for synthetic content, except where exemptions are provided by the Regulation.
No, transparency obligations apply equally to all businesses, regardless of size. However, SMEs may benefit from additional time to comply in cases of first-time, unintentional non-compliance. The AI Office has published specific guides to help smaller organisations apply these obligations with limited resources.
Documentation must include: a description of the systems in scope, the marking methods used, supervision processes implemented, and training provided to teams. Businesses must retain these documents for at least 5 years and make them available to supervisory authorities upon request. Our obligations guide details the elements to include in this documentation.
In the event of accidental dissemination of unmarked synthetic content, the business must: notify the AI Office within 72 hours, correct the marking of the affected content, and document measures taken to prevent recurrence. Deployers must also inform their providers to enable upstream correction. The CNIL recommends implementing an internal incident management procedure now.
