EU AI Act Article 50: what chatbots and AI content must disclose now

If your company runs a chatbot, publishes AI-generated images or produces AI video for marketing, the EU AI Act now affects you directly. Not through the heavily discussed rules for high-risk systems — but through Article 50, the transparency obligations that apply to everyday AI use.

This article explains what Article 50 requires, who is responsible for what, how it applies to typical business scenarios, and how to check your chatbots and AI content in an afternoon. It's a practical overview, not legal advice — for edge cases, talk to your lawyer.

What is Article 50 of the EU AI Act?

Article 50 sets transparency rules for AI systems that interact with people or generate content. Its goal is simple: people should know when they're dealing with a machine, and AI-generated content should be recognisable as such. It applies regardless of whether the system is "high-risk".

When does Article 50 apply?

The transparency obligations apply from 2 August 2026, together with most other provisions of the AI Act. In late 2025 the European Commission proposed simplifications to the AI Act (the "digital omnibus"), including a transition period for marking content from systems already on the market. Check the current status of the proposal before relying on transition periods — and don't plan on Article 50 going away.

What exactly does Article 50 require?

Article 50 contains four groups of obligations:

The information must be given clearly and at the latest at the first interaction or exposure, and it must be accessible — for example, also understandable for people with disabilities.

Provider or deployer — who is responsible for what?

The AI Act distinguishes between the provider, who develops an AI system or has it developed and puts it on the market under their name, and the deployer, who uses an AI system in their own business. Many companies are both: if you build your own customer support bot, you're the provider of that bot and you deploy it.

• Building your own chatbot or content pipeline? You're responsible for AI disclosure in the conversation and, for generated media, for machine-readable marking.
• Using a third-party tool? The tool's provider handles technical marking; you remain responsible for disclosing deepfakes and how you publish the content. Check that the tool actually marks its output and your workflow doesn't strip it.

Which AI use cases need a disclosure? Six real-world scenarios

The table below reflects our reading of Article 50 for common setups — a starting point for your own assessment, not a legal opinion. Context can change the answer.

The pattern: anything that talks to people needs an AI notice, and the visible label on content is mostly about realism — the more it could be mistaken for a real person or event, the clearer the label.

What does this mean for chatbots and AI agents?

For customer-facing bots, the requirement is straightforward: tell people they're talking to an AI. In practice that means a clear first message or label — not a line hidden in the terms.

Disclosure is worth doing well rather than grudgingly. In our live support-agent project, our AI support agent resolves 80% of requests and 70% of customers rate their conversations positively; disclosing that it's an AI doesn't change what customers care about: fast, correct answers.

What should an AI disclosure actually say?

A good disclosure is short, comes first and tells people what they can do next. Adapt these examples to your brand and channel.

Website or messenger chatbot

Hi, I'm the virtual assistant of [Company] — an AI, not a human. I can help with orders, returns and opening hours. If you'd rather talk to our team, just type "human".

Voice agent on the phone

Hello and welcome to [Company]. You're speaking with our AI voice assistant. If you'd like to talk to a person at any time, just say "colleague".

AI-generated ad or avatar video

AI-generated video. The person shown is not real.

For stylised visuals, a small "Created with AI" tag is usually enough. For realistic avatars, put the label where people see it while watching, not only in the post text. For an AI-assisted article you choose to disclose voluntarily, a line such as "Drafted with AI support, reviewed and edited by our editorial team" works well.

Common mistakes we see: a human-sounding persona ("Anna from support") with no AI note, a disclosure that only appears in the privacy policy, and a website bot that is labelled while the same bot on WhatsApp is not.

What does this mean for AI-generated marketing content?

For images and video, two layers matter:

• Technical marking. Generation tools and pipelines must embed machine-readable signals — such as metadata, content credentials (C2PA) or watermarks — that identify content as AI-generated. If you run your own content factory, this belongs in the pipeline.
• Visible disclosure for deepfakes. If content shows realistic people, objects, places or events that could be mistaken for real, you must disclose that it's AI-generated. For clearly artistic, satirical or fictional works, a lighter disclosure that doesn't spoil the work is allowed.

Typical AI ad formats — product shots, stylised animations, obviously synthetic scenes — are usually low-risk. AI avatars that look like real people talking about a product deserve a clear label.

How does machine-readable marking work in practice?

The AI Act asks for marking solutions that are effective, interoperable, robust and reliable as far as this is technically feasible, taking into account the state of the art. It doesn't prescribe one technology. In practice, three methods are combined:

• Metadata. Fields in the file (for example IPTC or XMP) that state the content was created by a generative model. Easy to add, but many platforms, screenshots and re-exports strip them.
• C2PA content credentials. A cryptographically signed manifest recording how a file was created and edited. Tampering is detectable, but like metadata it can be removed.
• Invisible watermarks. Signals embedded in the pixels or audio. They survive compression better but usually need the vendor's detector, and heavy editing weakens them. For text, watermarking remains unreliable.

Where marking sits in the pipeline matters too. In our experience, it rarely fails at generation — it fails at a later editing or export step that silently drops it.

In the Virality Bomb content factory, which brought video costs from around $10 to $0.20–0.40 per video, marking is built into the pipeline rather than added afterwards. The European Commission has also been working on a code of practice for marking and labelling AI content; check the latest version before finalising your setup.

How does Article 50 relate to GDPR transparency?

They are separate obligations that apply side by side. Article 50 is about recognising AI; the GDPR's information duties are about personal data — what your chatbot collects, why, for how long and who receives it. A chatbot that processes names, emails or order data typically needs both — the AI notice in the conversation and a link to the privacy information. Neither replaces the other. For keeping conversation data in-house, see our guide to private AI and local LLMs under GDPR.

What are the fines for violating Article 50?

Violations of the transparency obligations can be fined up to €15 million or 3% of total worldwide annual turnover, whichever is higher. For small and medium-sized companies, the lower amount applies. Beyond fines, missing disclosure is a reputational risk — "they didn't tell us it was AI" is not a headline anyone wants.

How do you make your AI use Article 50-ready?

Most companies can do this inventory in a day. If your systems were built by us, AI disclosure and content marking are part of the setup by default — see our security & data principles and our AI agents & chatbots service. Questions about your own systems? Book a free call.