EU AI Act · Article 50
Article 50 — the transparency rule almost every SME has to meet
If you use an AI chatbot, an AI content generator, an accessibility overlay, or anything that produces deepfaked images or audio, you have transparency obligations under Article 50 — enforceable from 2 August 2026.
If a person interacts with your AI, you tell them. If your AI makes content, you label it. That's the whole idea — the detail is just where it applies.
"You're chatting with an AI assistant. For complex queries, ask to be transferred to a person."
"Drafted with AI assistance and edited by [Name]." — on AI-written articles.
"AI-GENERATED" — a visible label on deepfaked or synthetic imagery.
The four Article 50 obligations, in plain English
Article 50 carries four discrete obligations, each scoped to a different kind of AI use. Most SMEs trip at least one of them.
1. AI-system interaction disclosure — Article 50(1)
If a user is going to interact with an AI system — a chatbot, a virtual assistant, a phone voicebot — you have to tell them. The disclosure must be made at the first interaction, in a clear and distinguishable manner. It must be obvious — not buried in a privacy notice or revealed only after the user has shared personal details.
There's a carve-out for cases where the AI nature is obvious from context (a clearly-labelled AI demo, for example), and another for law-enforcement use. Neither helps a typical SME chatbot.
2. Marking AI-generated content — Article 50(2)
If your system produces synthetic audio, image, video, or text, the output must be marked as AI-generated in a machine-readable way (watermarks, content credentials, metadata). This is on the provider of the AI system — typically the foundation-model vendor — but as the deployer you're on the hook for honouring that marking and not stripping it out.
3. Emotion recognition + biometric categorisation — Article 50(3)
If you use AI to infer emotion or to categorise people biometrically (race, age, gender etc.) in any way that isn't already prohibited under Article 5, you must inform the affected individuals and process the data on a GDPR-lawful basis. Workplace emotion recognition is already prohibited under Article 5(1)(f); this rule catches the consumer-facing uses that aren't.
4. Deepfake labelling — Article 50(4)
AI-generated or AI-manipulated audio or visual content that resembles real people, objects, places or events — and would falsely appear authentic — must be explicitly labelled as AI-generated or manipulated. There are exceptions for satire, parody, and certain journalistic uses, but the safe default is to label.
A typical SME landing pattern. A small services company runs a lead-capture chatbot on its homepage and uses an AI tool to draft blog posts. They trip 50(1) (chatbot disclosure) and 50(2) (AI-generated content marking). Compliance is a one-line disclosure above the chatbot and a content-credentials policy in the blog CMS — not a six-month consulting project.
What a compliant disclosure looks like
The Act doesn't prescribe exact wording. The standard is "clear and distinguishable." In practice, the patterns that work:
- Chatbot opener: "You're chatting with an AI assistant. For complex issues, ask to be transferred to a person."
- Voice agent intro: "Hello — you're speaking to an AI assistant from [Company]. This call may be recorded for quality."
- AI-written article footer: "Drafted with AI assistance and edited by [Name]." (Belt-and-braces — combines Article 50(2) with editorial honesty.)
- Deepfake label: A visible "AI-generated" overlay or sticker in the corner of the image or video, plus content-credentials metadata.
The penalty
Up to €15 million or 3% of global turnover, whichever is higher. For SMEs the cap is the lower of the two — Article 99(6) is a deliberate SME protection. The fine ceilings are theoretical maxima; enforcement priorities will fall on systematic non-compliance and on businesses that ignore warnings, not on a small company that fixes its chatbot label within a fortnight.
What to do this month
- List every AI system that touches a user. Chatbots, voicebots, AI image generators, AI-written content publishing flows.
- For each, decide which Article 50 sub-rule applies. Most trip 50(1) or 50(2); fewer trip 50(3) or 50(4).
- Add the disclosure. One line of copy or a small label in most cases.
- Capture proof. Screenshot the live disclosure, save it. When the August deadline arrives you want evidence that you complied before it, not after.
On WordPress? Our free plugin adds a compliant Article 50 notice in one click — and detects the AI tools already running on your site. See the Klarvo WordPress plugin →
The shortcut. Klarvo's free Transparency Checker walks you through your AI system and tells you exactly which Article 50 sub-rule applies, with ready-to-deploy disclosure text. Free, no login.
Run the Transparency Checker →
Klarvo organises and explains EU AI Act compliance. It is not legal advice. For specific legal situations, consult a qualified professional.
Wondering about a specific tool? The Klarvo AI tool registry answers the Article 50 question per tool — 2,800+ entries, grown weekly from the WordPress plugin directory, each with the clause that applies and a disclosure you can adapt.
Frequently asked
When does Article 50 apply? +
From 2 August 2026. There is no transitional grace period and the date has not slipped. The Digital Omnibus proposal may delay parts of the Annex III high-risk regime to December 2027, but Article 50 transparency remains on 2 August 2026.
Does it apply to my company if we're outside the EU? +
Yes, if your AI system affects people in the EU. Like GDPR, the AI Act has extraterritorial reach — what matters is where the user is, not where the business is registered.
What does a compliant disclosure actually look like? +
It must be clear, distinguishable, machine-readable where feasible, and provided at the latest at the first interaction. A small line above a chatbot saying "You're chatting with an AI assistant" is sufficient for a chatbot. AI-generated images and videos need marking under Article 50(2). Deepfakes carry an explicit "this is an AI-generated or manipulated image/video" obligation under Article 50(4).
What's the fine? +
Up to €15M or 3% of global turnover, whichever is higher. For SMEs the cap is the lower of the two — a deliberate SME protection in Article 99(6).
Who's responsible — me or my vendor? +
Disclosure is on the deployer — the business that puts the AI system in front of the user. If your vendor provides the chatbot, you still own the obligation to disclose it. The vendor must make this technically possible.
Add Article 50 to your evidence vault in 10 minutes.
Klarvo's free tier classifies your chatbot or content tool, drafts the Article 50 disclosure text, and stores the proof — all in one flow.
Free tier · Full KlarvoEngine classification · No credit card