Your chatbot doesn't need to pass a conformity assessment. Your AI image generator doesn't require CE marking. But starting August 2, 2026, you must tell people when they're interacting with AI — or face the same penalties as high-risk systems.
Article 50 of the AI Act creates transparency obligations that apply to nearly every customer-facing AI system. Unlike high-risk requirements, these rules are straightforward. But they're also easy to overlook — and the penalties are severe.
The Stakes: Non-compliance with transparency obligations can result in fines up to €15 million or 3% of global turnover. The same maximum as high-risk violations. Don't let "limited risk" fool you into thinking this is optional.
What Are Limited Risk AI Systems?
Limited risk systems sit between high-risk (strict requirements) and minimal risk (no requirements). They interact directly with people or generate synthetic content, but don't make decisions that significantly impact rights or safety.
Systems That Interact With People
Chatbots, virtual assistants, AI phone agents, conversational interfaces — any system where a user might reasonably believe they're communicating with a human.
Systems That Generate Synthetic Content
AI image generators, video synthesis tools, text generators, voice cloning, music generation — any system that creates content that could be mistaken for human-created or real.
Emotion Recognition & Biometric Categorization
Systems that detect emotions or categorize people by biometric data (age, gender, ethnicity). These require specific notifications even when not high-risk.
The 4 Transparency Obligations (Article 50)
Article 50 creates four distinct transparency requirements. Each applies to different types of systems and different actors in the AI value chain.
1 Chatbot Disclosure
Who: Deployers (companies using AI systems)
What: Inform users they're interacting with an AI system
When: Before or at the start of the interaction
Exception: Not required if "obvious from the circumstances and context of use" — but don't assume this applies to you.
✓ Compliant: A clear "AI-Powered" badge visible throughout the conversation
✗ Non-compliant: A chatbot named "Sarah from Customer Support" with no AI disclosure
✗ Non-compliant: Disclosure buried in Terms of Service that users don't read
2 AI-Generated Content Marking (Providers)
Who: Providers (companies that develop AI systems)
What: Mark outputs in a machine-readable format that enables detection
Applies to: Synthetic audio, images, video, and text
Technical requirement: Watermarking, metadata embedding, or equivalent technical solutions that are interoperable, robust, and reliable.
✓ Compliant: Audio files include inaudible watermarks detectable by verification tools
✗ Non-compliant: AI-generated images with no embedded provenance data
✗ Non-compliant: Watermarks that are easily stripped or defeated
3 AI-Generated Content Disclosure (Deployers)
Who: Deployers publishing AI-generated content
What: Disclose that content was artificially generated or manipulated
When: When publishing content that addresses matters of public interest
Key point: This is about human-readable disclosure, not just machine-readable marking.
✓ Compliant: News article clearly states AI was used to generate illustrations
✗ Non-compliant: Publishing AI-generated news imagery without disclosure
✗ Non-compliant: Using AI-manipulated video in advertising without labeling
4 Emotion Recognition & Biometric Categorization Notice
Who: Deployers using these systems
What: Inform people that such a system is being used on them
When: Before exposure to the system
Note: Some uses of these systems are prohibited entirely (workplace, schools). This obligation applies only to permitted uses.
✓ Compliant: App notification: "This feature detects your emotional state"
✗ Non-compliant: Covert emotion detection in customer service calls
✗ Non-compliant: Hidden demographic analysis of website visitors
The 5-Question Transparency Test
Use this quick assessment to determine your transparency obligations:
→ You must disclose it's AI (Obligation 1)
→ Provider must mark outputs machine-readably (Obligation 2)
→ You must disclose AI generation to viewers (Obligation 3)
→ You must inform people before using it on them (Obligation 4)
→ You're a deployer and have disclosure obligations
Important: If you answered "yes" to any question, you have transparency obligations. These are not optional, and "we didn't know" is not a defense.
Deepfakes: Special Rules
The AI Act pays special attention to deepfakes because of their potential for harm. Article 50(4) specifically addresses AI-generated content that depicts people or events that didn't happen.
What Counts as a Deepfake?
- AI-generated images, audio, or video depicting real people doing or saying things they didn't do or say
- AI-generated content depicting events that didn't occur
- Content that has been manipulated to appear authentic when it isn't
Deepfake Disclosure Requirements
| Actor | Requirement |
|---|---|
| Provider (tool maker) | Mark content with machine-readable metadata enabling detection of artificial origin |
| Deployer (content creator) | Disclose that content is artificially generated or manipulated when publishing |
Exceptions for Art & Satire
The disclosure requirements are lighter when AI-generated content is:
- Part of an obviously creative, satirical, artistic, or fictional work
- Not intended to deceive and clearly presented as non-factual
However, even in these cases, disclosure must not prevent proper content moderation or interfere with freedom of expression protections.
Practical Advice: When in doubt, disclose. A clear "AI-generated" label protects you legally while also building trust with your audience. The exception for art and satire is narrow — don't rely on it unless your use case clearly qualifies.
Technical Requirements: Machine-Readable Marking
Providers of AI content generation systems must implement technical measures to mark outputs. The AI Act and the European Commission's Code of Practice specify:
Requirements for Marking Systems
- Machine-readable: Can be detected by automated tools, not just visible to humans
- Interoperable: Works across platforms and detection systems
- Robust: Resistant to removal or tampering
- Reliable: Accurately identifies AI-generated content without false positives
Accepted Technical Solutions
| Method | How It Works | Content Type |
|---|---|---|
| C2PA Metadata | Embeds provenance data in file metadata using industry standard | Images, video |
| Digital Watermarks | Imperceptible patterns embedded in the content itself | Images, audio, video |
| Cryptographic Signatures | Verifiable signatures linking content to its AI origin | All types |
| Fingerprinting | Unique identifiers registered in a verification database | All types |
If You're a Provider: You need to implement at least one of these technical solutions before August 2026. The Code of Practice (expected June 2026) will provide more specific guidance on acceptable implementations.
Practical Cases
Case 1: Customer Service Chatbot
Situation: E-commerce site uses an AI chatbot for customer inquiries.
Obligation: Chatbot disclosure (Obligation 1)
Solution: Opening message clearly states "I'm an AI assistant." Badge visible throughout conversation.
Case 2: Marketing Agency Using AI Images
Situation: Agency generates product images and social media visuals with AI.
Obligations: Content marking (Obligation 2 — provider's responsibility) + Disclosure when publishing (Obligation 3)
Solution: Use AI tools that embed C2PA metadata. Add "Created with AI" label on social posts.
Case 3: Voice Assistant Integration
Situation: Company integrates voice AI for phone customer support.
Obligation: Chatbot disclosure (Obligation 1)
Solution: "You're now speaking with our AI assistant" at call start. Option to transfer to human.
Case 4: Retail Analytics with Demographic Detection
Situation: Store uses AI cameras to analyze customer demographics (age, gender).
Obligation: Biometric categorization notice (Obligation 4)
Solution: Clear signage at entrance: "This store uses AI-powered cameras for demographic analysis."
Case 5: AI-Assisted Content Creation for News
Situation: News outlet uses AI to generate article summaries and some illustrations.
Obligation: Content disclosure for public interest matters (Obligation 3)
Solution: "AI-assisted" label on relevant articles. "AI-generated image" caption on illustrations.
The Code of Practice
The European Commission is developing a Code of Practice on Transparency of AI-Generated Content to provide detailed guidance on Article 50 implementation. Key points:
- Expected finalization: June 2026 (before August enforcement)
- Scope: Practical guidance on marking, labeling, and disclosure
- Technical standards: Specific requirements for watermarking and metadata
- Industry input: Developed with stakeholder consultation
The Code will be non-binding but following it will provide a strong compliance position.
Compliance Checklist
1. Audit your AI touchpoints: List every AI system that interacts with users or generates content
2. Categorize by obligation: Map each system to the relevant Article 50 requirement(s)
3. Design disclosure mechanisms: Create clear, prominent notifications for each touchpoint
4. Verify provider compliance: Ensure your AI tool vendors implement machine-readable marking
5. Update policies: Add AI disclosure to your content policies and style guides
6. Train teams: Ensure marketing, customer service, and content teams understand requirements
7. Document everything: Keep records of your compliance measures for potential audits
The Bottom Line
Transparency obligations are the AI Act's most broadly applicable requirements. If you deploy any customer-facing AI or publish AI-generated content, you're affected.
The good news: compliance is straightforward. Clear disclosure, proper labeling, and honest communication with users. These aren't just legal requirements — they're good business practices that build trust.
The risk of non-compliance isn't just fines. It's the reputational damage of being caught deceiving users about AI. In a market where AI trust is increasingly valuable, transparency is a competitive advantage.
Trusted AI Advisory
Need Help With AI Transparency?
Our Trusted AI practice helps enterprises audit their AI touchpoints, design compliant disclosure mechanisms, and build transparency into their AI operations. From chatbot scripts to content labeling policies, we help you turn regulatory requirements into trust-building opportunities.
Learn More Talk to an Expert