Transparency Obligations under Art. 50 EU AI Act: Chatbots, Deepfakes, AI Content
TL;DR
- Art. 50 in force from Aug 2, 2026 — not affected by the Digital Omnibus proposal (Nov 19, 2025; trilogue ongoing, NOT adopted)
- Chatbot disclosure: "You are chatting with an AI" at the start of the conversation
- Deepfake labeling: realistic synthetic persons or scenes must be clearly marked
- Watermarking: Art. 50(2) requires synthetic content to be machine-readable as artificial (C2PA standard dominates)
- Fine risk: up to 15M EUR / 3% global revenue
1. What Art. 50 requires
Art. 50 covers transparency for four categories of AI content: chatbots, emotion-recognition / biometric-categorisation systems, deepfakes (synthetic audio/video/image), and AI-generated text on matters of public interest. Each category has its own disclosure rule.
| Subs. | What | Obligation |
|---|---|---|
| 1 | AI for interaction with humans (chatbots, voice bots) | Inform user at the start |
| 2 | Synthetically generated content (images, video, audio, text) | Machine-readable watermarking |
| 3 | Emotion-recognition / biometric-categorisation AI | Inform the data subject |
| 4 | Deepfakes | Clear labelling as artificial |
2. Chatbot labeling
Minimum requirement: "You are speaking with an AI / a bot. We use AI to provide faster answers." Position: first message of the bot, visible in the UI. Exception under Art. 50(1): if the AI nature is "obvious" from context (e.g., recognizably synthetic voice). In practice, always disclose explicitly — the exception is narrow.
3. Deepfake labeling
A deepfake is artificially generated or manipulated content that resembles a real person, object, or event in a deceptive manner. Obligations: visible "Created with AI" indication or equivalent, in or on the content (not hidden). For video: indication at the start or watermark. For images: visual marker or caption. Artistic and satirical exception: discreet but still recognizable labeling.
4. Watermarking under Art. 50(2)
Synthetic audio, video, image, and text content must be machine-readable as artificial. 2026 standards: C2PA (Coalition for Content Provenance and Authenticity) is dominant; ISO/IEC 22376 emerging. Implement at the generation-pipeline level so downstream platforms can detect AI provenance automatically.
| Standard | Backed by | Adoption |
|---|---|---|
| C2PA Content Credentials | Adobe, Microsoft, Sony, BBC | Dominant in 2026, open standard |
| SynthID | Google DeepMind | Google tools (Imagen, Veo) |
| OpenAI DALL-E watermark | OpenAI | OpenAI tools only |
| Stable Signature | Meta | Stable-Diffusion forks |
5. Exceptions and hardship clause
- Law enforcement: LEA powers not to be impaired
- Lawful processing in legal proceedings: exempted
- Artistic freedom: discreet but mandatory labeling (Art. 50(4))
- "Obvious from context": exception from disclosure (subsection 1) — apply with caution
6. Practical checklist
- All chatbots in the company identified
- Chatbot greeting carries an AI disclosure
- Marketing materials reviewed for AI-generated images / videos
- Watermarking pipeline in place for all AI image-generation tools
- Deepfake prohibition included in the Acceptable Use Policy
- HR tools with emotion detection (video-interview software) labeled?
- Customer-support voice bot carries a bot disclosure?
- Supervisor-compliant complaint address for AI content established?
Summary
Art. 50 is a small set of rules with broad reach: every chatbot, every AI-generated image, every deepfake. It enters force Aug 2, 2026 and is not affected by the Digital Omnibus proposal (Nov 19, 2025; trilogue ongoing, NOT adopted). Implement disclosures and C2PA watermarking now — this is the most visible compliance signal to customers and supervisors.
Frequently Asked Questions
Do we have to label our chatbot?
Do AI-generated marketing images have to be labeled?
What must the labeling contain?
Which watermarking standards exist?
Exception for artistic freedom?
Fines for violations?
Sources
- Regulation (EU) 2024/1689 — EU AI Act (Art. 50) (As of: 2026-05-02)
- AI Act Article 51 — GPAI threshold (As of: 2026-05-02)
- European Commission — EU AI Office (as of: ongoing)