Article 50 of the EU AI Act requires that people know when they are dealing with AI. Since 2 August 2026 it has applied to chatbots and AI assistants, AI-generated content, emotion recognition and biometric categorisation, and deepfakes. Unlike the high-risk rules — delayed to December 2027 — Article 50 applies now, to everyday AI features, and breaches can be fined up to €15 million or 3% of worldwide annual turnover.
The four transparency obligations
| Situation | Who must act | What's required |
|---|---|---|
| People interact with an AI system (chatbots, assistants, agents) | Provider | Design the system so people are told they're interacting with AI, unless obvious from context |
| AI generates synthetic audio, image, video or text | Provider | Mark outputs in a machine-readable format so they're detectable as AI-generated |
| Emotion recognition or biometric categorisation | Deployer | Inform the people exposed to it |
| Deepfakes, or AI-generated text published to inform the public on matters of public interest | Deployer | Disclose that the content is AI-generated or manipulated |
There is one transitional exception: generative AI systems already on the market before 2 August 2026 had until 2 December 2026 to meet the content-marking requirement.
Does it apply to your SaaS product?
If EU users interact with your AI features, or AI-generated output from your product reaches people in the EU, assume yes. Article 50 isn't limited to high-risk systems — it covers ordinary features such as:
- an in-app AI assistant or support chatbot
- AI-drafted emails, documents or marketing copy
- AI-generated images or avatars
- voice agents
See EU AI Act for US SaaS companies for scope and roles.
Practical steps
- Inventory every AI feature that interacts with people or generates content.
- Assign roles — provider or deployer — for each.
- Add AI-interaction notices at the start of chat and voice experiences ("You're chatting with an AI assistant"). Make them clear, not buried in terms.
- Mark generated content — use machine-readable marking (metadata or watermarking) for generated media; check what your model provider supports.
- Label deepfakes and public-interest AI text where you publish them.
- Inform people before any emotion recognition or biometric categorisation.
- Document and evidence it — screenshots, design decisions, test records. Regulators expect transparency to be built in before release, not added afterwards.
- Contract for it — confirm your AI vendors support the marking and disclosure you need.
Good AI disclosure looks like
- Shown before or at the start of the interaction
- In plain language, in the user's language
- Visible on every channel (web, mobile, voice)
- With a path to a human where it matters
How it fits your wider program
Article 50 controls slot into an AI governance program alongside your AI inventory, policy and risk assessments — and into SOC 2 through change management (disclosures don't get removed in a redesign) and product documentation. See AI governance for SaaS startups.
FAQ
Is a line in our terms of service enough?
Generally no. The intent is that people are informed at the point of interaction, in a clear and distinguishable way.
Does an AI disclosure hurt conversion?
Clear disclosure, done well, builds trust. Keep it short and friendly and offer a human route.
Is this legal advice?
No. It's a practical overview; confirm your obligations with EU legal counsel. We help build the inventory, controls and evidence.
Need an Article 50 readiness check? Book a free scoping call.