If a single person in the EU can open your chatbot, upload to your image generator, or read text your agent published, the EU AI Act's Article 50 transparency rules apply to you from August 2, 2026 — and it does not matter where your company is registered. Like the GDPR, Article 50 binds anyone whose AI output is used in the Union (European Commission). The good news for a team of one: most of this is a disclosure you can ship in an afternoon, not an audit. Here's the whole surface, in the order it hits a typical AI startup, with the exemptions that let you skip work you don't owe.
The one-line version if you build a chatbot or agent: tell users they're talking to an AI, at the first interaction, in the interaction itself. Everything else below is conditional on whether you also generate media or publish AI-written content.
The five duties, at a glance#
Article 50 isn't one rule — it's five, split between providers (you built or supply the system) and deployers (you use it in your product or publish its output). Match yourself to the rows that describe what you actually ship. The table above is the whole map; the sections below are the actions.
1. Chatbot disclosure — the duty almost every founder owes (50(1))#
If your product includes an AI system designed to interact directly with people — a support chatbot, a voice agent, a conversational assistant — you are the provider of a 50(1) system. The obligation: ensure the person is informed they are interacting with an AI, clearly and distinguishably, at the latest at the point of the first interaction (Greenberg Traurig).
Ship this: a plain, visible notice in the conversation itself — the bot's opening message, a persistent label in the chat header, or a voice agent's first spoken line. A clause buried in your terms of service does not satisfy the duty; the disclosure has to live where the interaction happens.
The one exemption worth knowing: you can skip it when the AI nature is obvious to a reasonably well-informed person in the circumstances. A widget literally labeled "AI Assistant" that a user deliberately opened may already clear the bar — but the safe, cheap move is to disclose anyway. For the exact wording and placement patterns, we walked through the disclosure fix in detail in how to add an "I am an AI" disclosure before the deadline.
2. Synthetic-content marking — if you provide a generative model (50(2))#
If you are the provider of a system that generates synthetic audio, image, video, or text — your own image model, a voice-cloning API, a text generator you supply to others — Article 50(2) requires that outputs be marked in a machine-readable format and detectable as artificially generated or manipulated, in a way that is effective, interoperable, robust, and reliable as far as technically feasible (EU AI Act Service Desk).
Ship this: machine-readable provenance — C2PA content credentials, an embedded watermark, or signed metadata — on generated outputs. This is the one duty with real engineering weight, which is why it carries targeted transitional relief: providers whose systems were already on the market get additional time to bring marking into conformity, reported to run to December 2, 2026 (TechTimes). Don't treat that as a reason to ignore it — treat it as the window to implement C2PA properly.
Exemptions: systems that perform only an assistive function for standard editing, or that do not substantially alter the input data, are out of scope. Grammar cleanup and light photo retouching don't trigger 50(2); generating a new image or a synthetic voice does.
3. Deepfake labels — if you deploy generated media (50(4))#
Distinct from marking: if you are a deployer who publishes AI-generated or AI-manipulated image, audio, or video that constitutes a deepfake, you must disclose that the content is artificially generated or manipulated. Where 50(2) is a machine-readable mark the provider bakes in, this is a human-visible label the deployer puts on what they publish.
Ship this: a clear, visible "AI-generated" label on synthetic media in your product or marketing.
Exemptions that matter to builders: content that is evidently artistic, creative, satirical, or fictional only needs a disclosure that doesn't hamper the display or enjoyment of the work — a corner credit, not a full-screen interstitial. And uses authorised by law to detect, prevent, investigate, or prosecute a criminal offence are carved out entirely.
4. AI-written public-interest text — the editorial carve-out (50(4))#
If you deploy AI to generate or manipulate text published to inform the public on matters of public interest, you must disclose it's AI-generated — unless the content underwent human review and a natural or legal person holds editorial responsibility for it.
What it means: a fully automated "news" or public-affairs content pipeline needs an AI-generated label. Put a human editor with named responsibility in the loop and the disclosure duty lifts. (It's the same principle we bake into this publication — every piece is machine-drafted and human-reviewed, and we say so.)
5. Emotion recognition and biometric categorisation (50(3))#
Narrower, but if it's you: deployers of emotion-recognition or biometric-categorisation systems must inform the people exposed to them (subject to the usual law-enforcement carve-outs). Most solo AI products never touch this — but sentiment-on-camera or voice-emotion features do.
What it costs to get it wrong#
Article 50 breaches sit in the AI Act's transparency-obligation penalty tier under Article 99: up to €15 million or 3% of total worldwide annual turnover, whichever is higher (Bratby Law). Enforcement runs through national market-surveillance authorities, which member states must have stood up by August 2. That's the asymmetry that should drive your weekend: a disclosure line costs an afternoon; the downside is measured in millions.
This is the same global current we've tracked all month — China's persona law forced Doubao and Qwen to disclose or kill companion agents in July, and "tell the user it's an AI" is fast becoming a worldwide default, not an EU quirk. For the broader picture of how the Act treats autonomous systems, see the EU AI Act for AI agents.
The founder's Saturday checklist: (1) Add an "I am an AI" line to your chatbot's first message. (2) If you generate media, put C2PA/watermark marking on the roadmap now, with the transitional window as your deadline — not an excuse. (3) Label synthetic media you publish. (4) If you run an automated content pipeline, either label the output or put a named human editor in the loop. (5) If you do emotion/biometric inference, notify the people exposed. Four of the five are disclosures, not rebuilds — and every one of them is cheaper than the fine.
