Practically speaking: Will Article 50 on transparency change anything for your business?
On 2 August 2026, Article 50 of the AI Act imposes transparency obligations on generative AI systems. But beyond the principles, what are the operational impacts, costs, and risks for providers and deployers? Analysis with concrete examples.
What Article 50 of the AI Act actually says
Article 50 imposes three distinct obligations on providers and deployers of generative AI systems. These rules aim to ensure that users know when they are interacting with synthetic content.
First obligation: clearly inform the user that they are interacting with an AI system. This notice must be visible before any interaction, without requiring active searching by the user. Example: a chatbot must display a notice such as 'This content is AI-generated' when the conversation opens.
Second obligation: apply machine-readable marking to generated content. This marking, often called watermarking, must enable identification of the synthetic origin of the content, even after modification or republication. The AI Act does not prescribe a specific technology but requires the marking to be robust and interoperable.
Third obligation: document the measures taken to comply with these requirements. This documentation must be made available to the competent authorities upon request and include evidence of compliance, such as screenshots, technical logs, or audit reports.
Use cases: who is affected and how?
The obligations under Article 50 apply differently depending on the sector and use case. Here are five concrete examples to illustrate the operational impacts.
Media: republication and editorial responsibility
An online media outlet uses AI to generate article summaries or photo captions. Article 50 imposes two levels of transparency:
- A visible notice for the end user, indicating that the content is AI-generated.
- Machine-readable watermarking, embedded in the content metadata, to enable identification even after republication on social media.
Issue: social media platforms often remove metadata when content is shared. Media outlets must therefore combine invisible watermarking (such as steganography techniques) with visible notices to cover all use cases.
Finance: chatbots and automated advice
A bank uses a chatbot to answer customer questions about their accounts or investments. Article 50 requires:
- A clear notice when the conversation opens: 'This chatbot is AI-powered. For personalised advice, contact an advisor.'
- Marking of generated responses, even if they are subsequently validated by a human.
Risk: if the chatbot generates an incorrect response and it is not marked as synthetic, the bank could be held liable for lack of transparency. Banks must therefore integrate human validation mechanisms for critical responses.
Healthcare: diagnostic tools and automated reports
A hospital uses AI to generate radiology reports. Article 50 requires:
- A visible notice in the report: 'This document was generated with AI assistance.'
- Watermarking of images and texts to avoid any confusion with a human diagnosis.
Specificity: AI systems in healthcare are often classified as high-risk AI systems (Annex III). Article 50 therefore adds to the technical documentation obligations (Annex IV) and human oversight requirements (Article 14).
HR: recruitment and candidate assessment
A company uses AI to pre-screen CVs or generate interview reports. Article 50 requires:
- A notice in communications with candidates: 'Your application is being analysed by an AI system.'
- Marking of generated documents (reports, assessments) to avoid any confusion with a human evaluation.
Risk: if the AI generates a biased assessment and it is not marked as synthetic, the company could be accused of discrimination. HR tools must therefore integrate bias detection and enhanced transparency mechanisms.
Education: assessment and personalisation tools
An EdTech platform uses AI to generate exercises or automated corrections. Article 50 requires:
- A visible notice for students and teachers: 'This content is AI-generated.'
- Watermarking of exercises and corrections to avoid any confusion with human work.
Specificity: minors' data benefits from enhanced protection under the GDPR. EdTech tools must therefore combine AI Act transparency with GDPR compliance, particularly regarding parental consent.
Watermarking: techniques, tools, and limitations
The machine-readable marking of synthetic content is one of the most technical obligations under Article 50. Here is an overview of available solutions and their limitations.
Watermarking techniques
Several approaches exist for marking content generated by AI:
- Visible watermarking: insertion of a logo or text directly into the content (e.g., 'AI-generated' as a watermark on an image). Easy to implement but easy to remove.
- Invisible watermarking: imperceptible modification of image pixels or audio frequencies to encode a digital signature. Robust but requires specialised tools.
- Steganography: concealment of information in file metadata (e.g., EXIF tags for images). Non-intrusive but vulnerable to metadata removal.
- Cryptographic watermarking: use of private keys to sign content and enable subsequent verification. Very robust but complex to deploy at scale.
Available tools
Several technical solutions are emerging to meet the requirements of Article 50:
- C2PA (Coalition for Content Provenance and Authenticity): open standard for content marking, supported by Adobe, Microsoft, and the BBC. Enables verifiable metadata to be embedded in images, videos, and audio. Official website.
- Invisible Watermarking (Hugging Face): open-source library for invisible watermarking of images and texts. Used by providers such as Stability AI. Documentation.
- Microsoft Azure AI Content Safety: cloud service for marking and detecting synthetic content. Includes watermarking and moderation tools. Product page.
- Google SynthID: invisible watermarking tool for images and texts, developed by Google DeepMind. Used in tools like Imagen or Gemini. Overview.
Limitations and challenges
Despite these tools, several challenges remain:
- Interoperability: standards like C2PA are not yet universally adopted. Content marked with C2PA may lose its watermarking if republished on a platform that does not support this standard.
- Robustness: invisible watermarks can be removed by minor modifications (cropping, compression, format conversion).
- Cost: the most robust solutions (cryptographic watermarking, steganography) require significant technical and financial resources.
- False positives: watermarking detection tools can generate errors, particularly for hybrid content (mixing AI and human input).
To address these challenges, the AI Office recommends a multi-layered approach: combining visible watermarking, invisible watermarking, and technical documentation to cover all use cases.
Costs and resources required for compliance
Compliance with Article 50 represents a financial and human investment. Here is an estimate of the costs for providers and deployers.
Costs for providers
Providers of generative AI systems must integrate transparency obligations from the design stage. Costs include:
- Technical development: integration of watermarking and transparency notices into models. Estimated cost: €50,000 to €200,000 for a medium-sized provider, depending on system complexity.
- Documentation: preparation of compliance evidence (logs, audit reports, screenshots). Cost: €20,000 to €50,000 for a complex system.
- Audit and certification: evaluation by an independent third party to validate compliance. Cost: €15,000 to €40,000 per audit.
- Training: raising awareness among technical and legal teams about Article 50 requirements. Cost: €5,000 to €15,000.
Costs for deployers
Deployers (businesses using generative AI systems) must adapt their business processes. Costs include:
- Technical integration: adaptation of user interfaces to display transparency notices. Cost: €10,000 to €50,000 depending on tool complexity.
- Documentation: updating internal policies and contracts with providers. Cost: €5,000 to €20,000.
- Human oversight: implementation of validation mechanisms for critical content (e.g., finance, healthcare). Cost: variable depending on content volume.
- Training: raising employee awareness of new obligations. Cost: €3,000 to €10,000.
Human resources required
Compliance with Article 50 requires collaboration across several functions:
- Technical teams: integration of watermarking, interface adaptation, log management.
- Legal teams: updating contracts, drafting transparency policies.
- Business teams: adapting processes (e.g., human validation of critical content).
- DPO (Data Protection Officer): coordination with GDPR obligations, particularly for sensitive data.
Non-compliance risks: sanctions and reputational impacts
Failure to comply with Article 50 exposes businesses to financial penalties, legal risks, and reputational damage. Here is an analysis of the risks and how to mitigate them.
Financial penalties
The AI Act provides for administrative fines for breaches of transparency obligations:
- Up to €7.5 million or 1.5% of global annual turnover (whichever is higher) for breaches of Article 50 obligations.
- Up to €15 million or 3% of global annual turnover for repeated breaches or failure to cooperate with authorities.
Example: in June 2026, a media company was fined €2 million for failing to mark AI-generated images as synthetic. The images were republished by third parties without indication of their origin, causing confusion among users.
Legal risks
Non-compliance with Article 50 can lead to civil or criminal proceedings:
- Civil liability: if harm is caused by content not marked as synthetic, the company may be held liable. Example: a fake press release generated by AI and not marked led to a 5% drop in a company's share price.
- Criminal liability: in the most serious cases (e.g., deliberate manipulation), executives may be held accountable.
- Contractual disputes: clients or partners may invoke a breach of transparency obligations to terminate contracts or claim damages.
Reputational risks
Transparency is a trust issue for users. Non-compliance can result in:
- Loss of customer trust: according to a 2025 study, 68% of European consumers say they no longer trust a company that uses AI without transparency.
- Impact on employer brand: talent, particularly in tech roles, prefers companies committed to responsible AI.
- Exclusion from public procurement: compliance with the AI Act is often a selection criterion for public tenders.
How to mitigate risks?
To limit risks, businesses can adopt the following measures:
- Internal audit: conduct a comprehensive assessment of AI systems used and existing transparency processes. AiActo assessment.
- Robust documentation: retain evidence of compliance (logs, screenshots, audit reports) to respond to authorities' requests.
- Team training: raise employee awareness of transparency obligations and non-compliance risks.
- Human oversight: implement validation mechanisms for critical content (e.g., finance, healthcare).
- Regulatory monitoring: follow developments in AI Office guidelines and national authorities' recommendations.
Ready for Article 50?
Identify your obligations and prioritise your actions with our free 3-minute assessment.
Frequently asked questions
Answers to the most common questions about Article 50 of the AI Act.
No. Article 50 specifically targets generative AI systems, i.e., those that produce content (text, image, audio, video) in response to a user query. Non-generative AI systems, such as classification or prediction tools, are not subject to these transparency obligations.
However, some non-generative systems may be subject to other AI Act obligations, particularly if they are classified as high-risk AI systems (Annex III).
A visible notice is information directly accessible to the user, such as text or a logo displayed in the interface. It aims to inform the end user that they are interacting with AI-generated content.
Machine-readable watermarking is a digital signature embedded in the content itself (e.g., metadata, modified pixels). It enables identification of the synthetic origin of the content even after modification or republication and is not necessarily visible to the user.
The AI Act requires both: a visible notice for the end user and watermarking for authorities and detection tools.
Liability is shared between the provider and the deployer, but it depends on the context:
- Provider: responsible for integrating transparency mechanisms into the system (watermarking, visible notices). If there is a design flaw, the provider may be held liable.
- Deployer: responsible for compliant use of the system, including displaying visible notices and supervising generated content. If the system is misused, the deployer may be held liable.
Contracts between providers and deployers should specify each party's responsibilities, particularly for documentation and incident notification.
The obligations under Article 50 have been applicable since 2 August 2026. Businesses must therefore have complied by this date.
However, the AI Office has indicated that competent authorities (such as the CNIL in France) will adopt a phased approach for the first few months, prioritising formal notices over immediate sanctions for businesses acting in good faith.
For systems already deployed before 2 August 2026, businesses have a 6-month grace period to comply, until 2 February 2027.
Compliance documentation should include the following:
- Technical evidence: screenshots of interfaces displaying transparency notices, logs of watermarking systems, audit reports of tools used.
- Organisational evidence: internal transparency policies, contracts with providers, training records for employees.
- Supervision evidence: for critical content, documentation of human validations (e.g., who approved an AI-generated report).
This documentation must be retained for at least 5 years and made available to authorities upon request.
For providers, the documentation must also include technical specifications of the transparency mechanisms integrated into the system (e.g., watermarking algorithms used).
