Article 50 disclosure: templates and examples.

Article 50 of the EU AI Act asks for very little text. One sentence in a chatbot greeting, a label on a generated image, a line under a synthetic video. The hard part is not the wording, it is putting the right words in the right place at the right moment. This page gives you templates you can adapt, the placement rules behind them, and examples of wording that works and wording that does not.

Article 50(1) has applied since 2 August 2026. Under Article 50(5) the information must be clear and distinguishable, given at the latest at the first interaction, and meet accessibility requirements. A line in the privacy policy does not meet that test. The greeting does.

If you want the wording generated for your own product, the free Article 50 disclosure generator fills these templates with your product name and gives you a copy-paste HTML snippet and a policy clause. It runs in your browser and sends nothing anywhere.

Which paragraph applies to you

Provider and deployer are the Act's roles, not your job titles. If you ship the chatbot under your own name, you are its provider, even when the model underneath is someone else's. Provider or deployer? works through the edge cases.

Chatbots and assistants: Article 50(1)

The duty is to design the system so that people are informed they are interacting with an AI system, unless that is obvious to a reasonably well-informed, observant and circumspect person. Read the exception narrowly. A widget with a human first name and a stock photo is designed not to be obvious.

Template: chat greeting

You are chatting with an AI assistant from [Product], not a person. Its answers can be wrong, so check anything important.

Put it in the first message the user sees, before they type. If your bot has a name, keep the name and add the fact: "Hi, I'm Ava, [Product]'s AI assistant."

Template: persistent label in the chat header

AI assistant

A short label next to the bot's name, visible for the whole conversation. Useful when users scroll past the greeting or return to an old thread. It supports the greeting; it does not replace it.

Template: writing assistant or copilot inside an app

This feature uses AI to suggest and write text. Review suggestions before you use them.

Show it in the panel where the feature starts, not in onboarding the user saw once three months ago.

Template: voice agent on calls

Hello, you are speaking with an AI voice assistant from [Company], not a person.

Say it first, before the agent asks anything. If callers can reach a person, say how in the same opening.

Placement rules that hold up

• Before or at the first interaction. Not after the third message, not in a tooltip.
• In the interface, in the language of the interface. A disclosure in English on a German chat widget is not clear to the person reading it.
• Visible on a phone. Check the widget at 360 pixels wide. Collapsed headers and truncated greetings are where disclosures disappear.
• Readable by assistive technology. Real text, not an image of text, with enough contrast. The generator's snippet uses role="note" for this reason.
• Backed by a clause, not replaced by one. Your privacy policy or terms should describe the assistant too. That is supporting documentation, not the disclosure.

Synthetic content: Article 50(2), from 2 December 2026

Providers of AI systems that generate audio, image, video or text must mark the output in a machine-readable format so it can be detected as artificially generated. This is a technical duty. A visible "AI-generated" badge helps people, but a machine cannot read a badge. Marking means metadata such as content credentials, a watermark, or both, as far as technically feasible.

Who carries it: the provider of the generative system. If you call a model provider's API and pass the output through unchanged, the marking duty for that output generally sits with them. If you generate, assemble, composite or re-encode media yourself, check rather than assume: re-encoding and cropping routinely strip provenance metadata. Where the provider line sits.

Template: visible label on generated media

AI-generated image

Short, on or directly under the asset, and still there when the asset is downloaded or shared from your product.

Template: marking statement for your documentation

Images, audio and video generated in [Product] carry machine-readable marking: [describe the method, for example C2PA content credentials and an invisible watermark]. We check that the marking survives our export formats. Last reviewed: [date].

Buyers ask for exactly this in vendor questionnaires. Keep it dated. How to answer the Article 50 question in a questionnaire.

Deepfakes and public-interest text: Article 50(4)

This one binds deployers, so it lands on whoever publishes the content, whatever tool produced it. Image, audio or video that resembles real people, objects, places or events and would falsely appear authentic must be disclosed as artificially generated or manipulated.

Templates: deepfake labels

This video was generated with AI. [Name] did not record it.

AI-generated voice. This is not a recording of [Name].

Place the label where the content is seen: on the video, in the first seconds of the audio, in the caption of the image. A note on a separate page does not travel with the content.

For evidently artistic, creative, satirical or fictional work, the duty is adapted, not removed: you disclose in a way that does not spoil the work, for example in the credits or the description. "It is an advert" is not one of the carve-outs.

Templates: AI-generated text on matters of public interest

Text published to inform the public on matters of public interest must be disclosed as AI-generated, unless it underwent human review and a natural or legal person holds editorial responsibility for its publication.

This article was generated with AI and published without editorial review.

If an editor reviews the text and your organisation takes responsibility for it, the duty does not apply. Many publishers say so anyway, which costs nothing:

Written with AI assistance. Reviewed and edited by [Name], who is responsible for this publication.

Emotion recognition and biometric categorisation: Article 50(3)

Deployers must inform the people exposed to the system that it is operating, and GDPR duties apply on top. A short template:

This [service or area] uses an AI system that [analyses facial expressions / categorises people by biometric data] to [purpose]. More: [link].

In the workplace and in education, emotion recognition is prohibited outright under Article 5, with narrow medical and safety exceptions. A disclosure does not make a prohibited practice lawful.

Wording that works, and wording that does not

Keep a record

Disclosure is quick to add and easy to lose in a redesign. Keep a short, dated note per feature: which paragraph applies, the exact wording, a screenshot of where it appears, and for generated media how it is marked. Note the features you decided were exempt too, and why. That note is what you hand to a buyer, an investor or an authority.

Breaches of Article 50 fall under Article 99(4): up to €15,000,000 or 3% of worldwide annual turnover, whichever is higher, or for SMEs and start-ups, whichever is lower. The full fine structure. For the duties themselves in full, see Article 50 explained.

Disclosure wording is one obligation. Whether Article 50 applies to your product, which paragraphs, whether you or your model vendor is the provider, and what else the Act asks of you depend on the product. The free scan gives your likely risk tier in about two minutes, and the sample report shows what the full written position looks like. It is a self-assessment, not legal advice.

Frequently asked questions

Where should an Article 50 chatbot disclosure appear?

In the interface, at the latest at the first interaction (Article 50(5)): the chatbot greeting, the panel of the AI feature, or the opening words of a call. A clause in the privacy policy supports it but does not replace it.

Is a visible AI-generated label enough for images, audio and video?

Not on its own. Article 50(2) requires providers of generative systems to mark output in a machine-readable format, such as content credentials or a watermark, from 2 December 2026. A visible label helps people; marking is for machines.

Do I have to label AI-generated text?

Under Article 50(4), deployers must disclose AI-generated text published to inform the public on matters of public interest, unless it underwent human review and a person or organisation holds editorial responsibility for its publication.