AI Act, article 50: transparency

Article 50 of the European AI Act (Regulation 2024/1689) sets transparency duties for AI systems that interact with people and for AI-generated or manipulated content. People must know they are talking to AI, and generated content must be recognisable in a machine-readable way.
What article 50 asks
- Interaction: whoever offers an AI system that interacts with people must make sure they know they are dealing with AI, unless that is obvious from the context.
- Generated content: providers of systems that generate synthetic audio, images, video or text must mark the output in a machine-readable format as artificially generated or manipulated.
- Deepfakes: whoever uses them must disclose that the content is artificially generated or manipulated.
The regulation entered into force in August 2024 and applies in stages. The exact start dates and any transition periods are still being adjusted in Europe. Follow the current text and the European Commission's guidelines.

Provider and deployer
The duties differ per role. A provider of an AI system must make the marking technically possible. An organisation that publishes such content (the deployer) must in some cases disclose itself that something was made by AI. So make sure you know which tools in your organisation generate content.
What you can arrange now
- List which tools generate AI content (video, audio, images, text).
- Ask your suppliers how they mark generated content.
- Choose deliberately which AI model is used for what and record it.
- Agree internally how you make AI use visible to recipients.
In CoVince, AI-generated audio and video get a machine-readable marking in the file metadata. We do not add a visible label in the picture automatically; how you present that to recipients is up to you. Read more in marking AI-generated content.
Examples from work
- A chatbot on your website that answers questions: visitors must know they are talking to AI.
- A training video with an AI-generated voice: the audio must be marked in a machine-readable way.
- An AI-made image that appears to show a real event: that deserves extra care.
Put it in your policy
Describe in a short AI guideline which tools may generate, who approves, how you inform recipients and where to find the origin of a file. Whoever knows in advance which model runs for what can answer faster when a customer or supervisor asks. See also European AI models.
An example: the communications team of a care institution
The communications team of a regional care institution makes an information video for patients. An AI-generated voice reads the text and an AI tool makes the background images. Before publishing, the team goes through four questions: which parts were made by AI, is the output marked in a machine-readable way, must viewers know (and if so, how do we say it without disturbing the video) and who approved it? The answer is recorded in one line in the project file. That takes five minutes and prevents an awkward conversation later.
Which model, and why it matters
Transparency is not only about the output but also about the chain behind it: which model produced the text or image, and where does it run? In CoVince an organisation can choose European models, such as Mistral AI (through the EU endpoint) and Scaleway (Europe), and local models, in the model picker. That does not mean all AI runs in the EU: it is a choice you make as an organisation and record.
Checklist for your AI guideline
- Which AI tools are allowed, and for which tasks?
- Who approves AI-generated content before publication?
- How and when do you inform recipients that something was made by AI?
- Where can the origin of a file be found?
- Who answers questions from customers or supervisors about this?
What article 50 is not
Article 50 is about transparency, not about whether an AI application is permitted. That question follows from the risk classes of the regulation, in which some applications are prohibited and others, as high risk, face extra requirements. A chatbot that helps customers can therefore fall under article 50 and also under other rules, depending on what it does. So look per application at both the transparency duty and the risk class, and record both. That prevents a correctly marked video from turning out to be an application that demands more.
Common mistakes
- Thinking article 50 applies only to AI providers and not to organisations that publish.
- Treating a visible label as a replacement for the machine-readable marking, or the other way round.
- Having no overview of which department uses which AI tool.
What you can do on Monday
Ask each department which AI tools they use to make something that goes outside. Put the answers in one list and discuss with your lawyer which cases fall under article 50. An organisation that has this overview can react quickly when start dates or guidelines change. This is general information, not legal advice, and a platform does not make you compliant by itself.
Frequently asked questions
What is the AI Act?
Regulation (EU) 2024/1689, the European law that regulates AI systems by risk, with transparency rules in article 50.
Must AI content be visibly labelled?
Article 50 asks for a machine-readable marking by the provider and, in some cases, a disclosure by whoever publishes the content. Assess your situation with your lawyer.
When do the rules apply?
Start dates and transition periods are being adjusted. Consult the current text of the European Commission.
Does CoVince do anything about this?
Yes, AI-generated audio and video get a machine-readable marking in the metadata. That is a basis, not a guarantee that your organisation is compliant.
Can I choose a European AI model in CoVince?
Yes. In the model picker you can choose European models (Mistral AI through the EU endpoint, Scaleway) and local models. It is a choice, not automatically all AI.
Who is responsible when an employee publishes AI content?
The organisation that publishes. Record in an AI guideline who approves and how you inform recipients.
