By the end of this guide you will know which Article 50 notices apply to your company, which belong to your software vendor, and what the label should say. You need no tools to follow it: a text editor and an honest list of where your company uses AI.
Provider or deployer: find your role first
The whole article splits along two roles defined in Article 3 of the AI Act:
- Provider (Art. 3(3)): whoever develops an AI system, or has one developed, and places it on the market or puts it into service under its own name, paid or free.
- Deployer (Art. 3(4)): whoever uses an AI system under its authority, except for personal, non-professional use.
Most small companies are deployers. You use a chatbot product, an image generator or a writing assistant that someone else built.
There is one trap worth naming. If you assemble your own chatbot from an open model and put it on your website under your brand, you can be the provider of that system. Read the definition against your own setup before you decide which column of the table below is yours.
Who must do what under Article 50
This summary follows the text published on EUR-Lex. It is a reading aid, not a substitute for the article.
| Paragraph | Who | What they must do |
|---|---|---|
| 50(1) | Provider | Design systems that talk to people so those people are told they are dealing with an AI, unless that is obvious |
| 50(2) | Provider | Mark synthetic audio, image, video and text in a machine-readable, detectable way |
| 50(3) | Deployer | Tell people exposed to emotion recognition or biometric categorisation that the system is running, and follow the GDPR |
| 50(4) | Deployer | Disclose deep fakes (image, audio, video), and disclose AI-generated text published to inform the public on matters of public interest |
| 50(5) | Both | Give the information clearly, at the latest at the first interaction or exposure, meeting accessibility requirements |
The Commission’s Article 50 FAQ adds that the “obvious” exception in 50(1) is judged by a reasonably well-informed, observant and circumspect person, and that it should be read restrictively. In plain words: do not rely on it.
The dates, checked on 2026-10-08
- 2 August 2026. Article 50 applies, per Article 113 of the regulation.
- 2 December 2026. Generative AI systems already on the market before 2 August 2026 get until this date for the machine-readable marking in 50(2). This is a provider duty, added as Article 111(4) by the Digital Omnibus on AI, Regulation (EU) 2026/1744 of 8 July 2026. The Omnibus did not move the deployer duties in 50(3) and 50(4).
- 10 June 2026. The Commission published the final Code of Practice on Transparency of AI-generated Content. Signing it is voluntary; the obligations are not. The same page says the EU created a set of icons deployers may use to label AI-generated content.
- 20 July 2026. The Commission published its guidelines on the transparency obligations.
Our EU AI Act timeline places these next to the other milestones.
Step 1: list where people meet your AI
Open a spreadsheet with four columns: tool, who sees the output, type of output, your role. Fill one row per use. Typical rows for a 10-person office:
- A chatbot on the website or in WhatsApp that answers customers.
- An image generator used for social posts or ads.
- A voice tool that reads product videos.
- A writing assistant used for the newsletter, the blog or press notes.
- Any camera or call-centre tool that claims to read emotions.
Ask your AI assistant to help with the sorting, but decide the role column yourself. A prompt that works:
Here is a list of the AI tools my company uses and what we do with each output.
For each row, tell me which paragraph of Article 50 of Regulation (EU) 2024/1689
could apply (50(1) to 50(4)) and why. Mark any row where you are not sure.
Do not invent obligations that are not in the text.Step 2: label the chatbot, even if the vendor should
50(1) is the provider’s duty. Your chatbot vendor should already show a notice. Check that it does, in the languages your customers use. If you are the provider because you built it, the notice is yours to add.
Put it in the first message, before the user types anything. Text you can copy:
EN: You are chatting with an AI assistant, not a person. It can make mistakes,
so check anything important. To talk to someone on our team, type "person".
ES: Estás hablando con un asistente de inteligencia artificial, no con una persona.
Puede equivocarse, así que comprueba lo importante. Si prefieres hablar con
alguien del equipo, escribe «persona».Keep it visible on mobile and readable by screen readers (50(5) asks for accessibility). A tooltip that appears only on hover does not count as clear.
Step 3: label generated images, audio and video
50(4) covers deep fakes: content that resembles existing people, objects, places or events and would falsely appear authentic (Art. 3(60)). A product photo of your real shop with a sky painted in by AI can fall here. An obviously illustrated mascot usually does not.
The simplest rule is to label every generated or heavily manipulated asset. Wording you can use:
| Asset | EN label | ES label |
|---|---|---|
| Image | ”Image generated with AI” | «Imagen generada con IA» |
| Edited photo | ”Photo edited with AI” | «Foto editada con IA» |
| Voice | ”Synthetic voice generated with AI” | «Voz sintética generada con IA» |
| Video | ”This video contains AI-generated scenes” | «Este vídeo contiene escenas generadas con IA» |
For artistic, satirical or fictional work, the article limits the duty to disclosing that generated content exists, in a way that does not spoil the work. A line in the credits is the usual answer.
Step 4: decide what happens to AI text
The text duty in 50(4) is narrow. It covers AI-generated text published to inform the public on matters of public interest. It does not apply when a person reviewed or edited the text and a natural or legal person holds editorial responsibility for the publication.
For a company blog or newsletter, that gives you two clean options:
- A person edits every AI draft and your company signs the publication. Keep a note of who reviewed it.
- You publish without review. Then add a line such as “This text was generated with AI” / «Este texto se ha generado con IA».
Marketing copy about your own products is a different question from public-interest information. When in doubt, the label is cheaper than the debate.
When Article 50 is not your problem
- Personal use. The deployer definition excludes personal, non-professional activity.
- Assistive editing. 50(2) does not apply where the AI only performs an assistive function for standard editing or does not substantially alter the input.
- Law enforcement uses authorised by law have their own carve-outs in each paragraph.
- Internal tools that no customer or member of the public sees, such as a private writing aid, rarely trigger 50(4). Your AI literacy duties under Article 4 still apply to the staff who use them.
The one-page checklist
- Inventory of every AI use, with your role (provider or deployer) per row
- Chatbot notice visible in the first message, in every customer language
- Vendor confirmation that their outputs carry machine-readable marks (50(2))
- Labels on generated or manipulated images, audio and video
- A rule for AI text: human review with a named editor, or a label
- Notice for any emotion-recognition or biometric tool, plus its GDPR basis
- The date you last checked this list, and who checked it
Next steps
- Run the EU AI Act self-assessment scorecard to see where Article 50 sits among your other duties.
- Read the EU AI Act page for the full checklist we work from.
Work with us
If you want the inventory and the notices done with you, our entry offer, “EU AI Act: deployer-ready in two weeks”, covers exactly the notices that apply to your company. It is described on the EU AI Act page. Book a 15-minute call or see how we work in consulting.