Ales Frelih

EU AI Act Article 50: The Transparency Rules That Apply If You Use a Chatbot or Publish AI Content

In the last post I covered Article 4 — the documentation you need to prove your team understands how to use AI.

Article 50 is different. It is not about internal processes. It is about what you tell people outside your business: your customers, your website visitors, your social media audience.

The question Article 50 answers is simple: if someone is talking to an AI on your website, do they know it? If they are looking at an image you published, do they know an AI made it?

These are not philosophical questions. From 2. August 2026, they are legal ones.

What Article 50 actually requires

Article 50 covers two separate obligations. Most businesses will be affected by one or both.

Obligation 1: Chatbot disclosure

If you deploy any AI system that interacts with humans — a chatbot on your website, an AI customer service agent, a voice assistant — you must inform users, in a clear and distinguishable way, that they are interacting with an AI.

The requirement applies unless it is obvious from the context. A button on your website labelled “Chat with our AI assistant” probably satisfies this. A widget that launches with “Hi, I’m Maya, how can I help you today?” probably does not — unless “Maya” is clearly identified as AI somewhere in that same interface.

The rule applies to real-time interaction: live chat, voice, instant messaging with AI on the backend.

Obligation 2: AI-generated and AI-manipulated content

If you publish content that is:

  • An image generated by AI
  • A video generated or significantly manipulated by AI
  • An audio file generated or significantly manipulated by AI

…you must label it as artificially generated or manipulated.

The label needs to be machine-readable. For practical purposes, this means embedding metadata into the file itself — not just writing a caption. The regulation anticipates that platforms will eventually read this metadata automatically. Until that infrastructure exists, a visible caption or disclosure also works and provides the human-readable layer.

There is an explicit exception for satire, parody, and artistic expression — but only if the labelling does not risk serious harm to public interest. For most business content (marketing images, product photos, social posts), the exception does not apply. Label the content.

Who this affects

The honest answer is: most businesses who use modern AI tools.

If you have any of these, Article 50 is relevant:

  • A chatbot on your website (Tidio, Intercom bot, a custom ChatGPT widget, Voiceflow, Landbot, or any similar tool with AI behind it)
  • AI-generated marketing images (Midjourney, DALL-E, Stable Diffusion, Adobe Firefly, Canva AI image generation)
  • AI-generated or AI-edited video (Sora, RunwayML, HeyGen, Synthesia, CapCut AI features)
  • AI voice content published publicly (ElevenLabs audio on your website, AI voiceover in a YouTube video, synthetic voice in a podcast)

If none of these apply to your business today, you are currently outside Article 50’s scope. But this changes the moment you add a website chatbot or use AI to generate a social media image.

The practical steps

For businesses with a chatbot

  1. Identify every entry point where customers interact with your AI: website chat, WhatsApp bot, email auto-responder with AI drafting, voice IVR with AI routing.
  2. At each entry point, add a disclosure before or at the moment interaction begins. The wording does not need to be elaborate. Examples:
  • “You are chatting with an AI assistant.”
  • “This conversation is handled by AI. For human support, contact us at…”
  • “AI-powered chat — [Your Company] Customer Service”
  1. Document where you added the disclosure and when. One line per chatbot implementation: tool name, URL or channel, disclosure text, date added.

That is the complete implementation for most businesses. No technical rebuild required. Add a line of text to your chat widget configuration.

For businesses that publish AI-generated content

  1. Establish a simple labelling convention for your team. The most practical approach: any AI-generated image used in a social post gets a caption element or alt-text tag that includes “AI-generated image.” Any AI voice audio published publicly gets a note in the description or accompanying text.
  2. For the machine-readable requirement, use whatever metadata your tool supports. Adobe Firefly embeds Content Credentials automatically. Midjourney does not do this natively yet, but you can use tools like C2PA-compliant editors to add it. At minimum, visible human-readable disclosure satisfies your obligation today.
  3. Write one paragraph in your AI Use Policy (from Article 4) describing your content labelling approach: “When we publish AI-generated images, we label them as such in captions. When we use AI-generated voice in public audio or video content, we disclose this in the description. This applies to all external-facing channels including our website, social media, and marketing materials.”

That paragraph does two things: it documents your compliance intent, and it signals to staff what the expectation is.

What the mistakes look like

Mistake 1: Treating the chatbot name as sufficient disclosure

Calling your bot “Aria” or “Max” is not a disclosure. The name needs to communicate AI, not just imply it. “Aria — AI Assistant” works. “Aria” alone does not. Check every chatbot implementation for this.

Mistake 2: Caption-only disclosure for AI images without considering reach

A caption on an Instagram post that says “Created with AI” satisfies the human-readable requirement. But if you are a business publishing content at scale — a news aggregator, a marketing agency publishing for clients, an e-commerce site with hundreds of product images — the machine-readable layer becomes more important. Start thinking about how you embed metadata now, before regulators start enforcement.

Mistake 3: Assuming your AI tool handles the labelling automatically

Some do. Most do not. Adobe and Microsoft have invested in C2PA content credentials. Most image generation tools have not implemented this yet. Do not assume your tool is compliant — check its documentation, and add visible labelling regardless.

Mistake 4: Not distinguishing AI interaction from AI assistance

Article 50’s chatbot rule applies to systems that interact with humans in real time. It does not require you to disclose every time your team uses AI internally to draft a response. The disclosure obligation is about the customer’s experience — not your team’s workflow. If your customer service agent uses ChatGPT to draft replies and then sends them as their own, Article 50 does not apply to that internal workflow (though Article 4 literacy obligations still do).

CTA — Templates and checklists

If you want to confirm exactly whether Article 50 applies to your business — including which tools and channels trigger the obligation — the free German and French checklists cover this specifically.

Kostenlose Prüfung — Bin ich betroffen? https://frelih.gumroad.com/l/slwzch (DE)
Verification gratuite — Suis-je concerne? https://frelih.gumroad.com/l/bghydm (FR)

If you want the done-for-you templates — including the AI Use Policy with the content labelling section already written, the chatbot disclosure text in multiple languages, and the training records template — the full kits include all of this.

EU AI Act Compliance Kit — EUR 149 https://frelih.gumroad.com/l/mjsaqg (German)
Kit Conformite Loi IA — EUR 149 https://frelih.gumroad.com/l/xdyenu (French)

The timeline and the risk profile

August 2, 2026 is the deadline for both Article 4 and Article 50. Seven weeks from the time this is written.

Article 50 enforcement will initially focus on visible, obvious violations — a business using a chatbot with no disclosure, or a public figure using AI deepfakes without labelling. Small businesses using a website chatbot are not the primary target of enforcement actions in year one.

That said, the obligation exists from day one. Adding a two-line disclosure to a chat widget takes five minutes. Not doing it is a compliance gap that costs nothing to close and creates unnecessary exposure.

The realistic risk for a small business is not a regulatory fine in August 2026. The realistic risk is a complaint from a user who felt misled, or a B2B client whose own compliance requirements extend to their suppliers. A visible disclosure protects you in both cases.

Where to go next

This series has now covered the two obligations that matter for most SMEs before August: Article 4 (AI literacy documentation) and Article 50 (transparency disclosures).

The next significant compliance date after August 2026 is 2 December 2027 for high-risk AI system requirements — Article 6 and Annex III. High-risk applies to a narrow set of AI applications: hiring tools that make autonomous decisions, creditworthiness assessment, biometric categorisation. Most SMEs are not in scope for December 2027.

The post after next will cover how to check whether any of your AI tools fall into high-risk categories, and what to do if they do.

Ales Frelih writes about EU AI Act, forex trading, money, and AI tools from a desk in Norway. This is informational. Talk to a lawyer before relying on it for compliance.

Leave a comment