A customer chats with an AI assistant. A communications team publishes an AI-generated image, while an editor uses AI to help draft an article. All three involve AI, but Article 50 of the EU AI Act does not treat them in the same way.
If Article 50 is new to you, the idea is simple: in certain situations, people should know when AI is involved. That may mean telling someone they are interacting with an AI system or disclosing that content was artificially generated or manipulated.
Since 2 August 2026, these transparency obligations have applied to providers and deployers of certain AI systems across the European Union and, in some cases, to organizations based outside it. For governance, compliance, legal, risk, product, and communications teams, the practical question is where those obligations apply.
So what does that mean for your organization? Are you a provider or a deployer? Does every AI-generated image now need a label?
AI provider vs. AI deployer
Under Article 50, your responsibilities depend partly on whether your organization is a provider or a deployer.
A provider develops an AI system, or has one developed, and offers it under its own name or trademark. Think of OpenAI offering ChatGPT, Anthropic offering Claude, or Google offering Gemini.
A deployer uses an AI system in its professional activities. A consultancy using Claude to analyze client material, an insurer using third-party AI to support claims processing, or a manufacturer using an external computer-vision system on its production line would typically be acting as deployers.
The same company can be both. Microsoft, for example, can act as a provider when it offers Copilot and as a deployer when it uses AI systems internally. The role depends on the specific system and how the organization uses or offers it.
When people interact with AI
Once you know your role, look at the moment the rule becomes real.
A customer visits your website with a question about an invoice. A chat window answers immediately. If nothing indicates that AI is responding, the customer may reasonably assume there is a person on the other side.
For AI systems designed to interact directly with people, providers must now ensure that users know they are interacting with AI, unless that is already obvious from the context. Where disclosure is required, it must be clear, distinguishable, and provided from the start of the first interaction.
If your organization deploys that chatbot, the legal duty under Article 50(1) still sits with the provider. But you should check the customer experience in practice: has branding, configuration, or integration made the AI notice less visible or removed it altogether?
That question becomes more complex when AI is generating content rather than holding a conversation.
Not every AI-generated image is a deepfake
Using AI to create an image does not automatically mean you need to add a visible label.
The key question is whether the content qualifies as a deepfake. Under Article 50, that means AI-generated or manipulated image, audio, or video content that resembles an existing person, object, place, entity, or event and could falsely appear authentic or truthful.
A fictional illustration created for a campaign, for example, is not automatically a deepfake simply because AI was used. A realistic video that appears to show your CEO announcing something they never said is a much clearer case.
If the content is a deepfake, the deployer must tell people that it was artificially generated or manipulated. The disclosure must be clear and perceivable to the audience.
For evidently artistic, creative, satirical, or fictional works, disclosure is still required when the content is a deepfake, but it can be provided in a way that does not interfere with the experience of the work.
Not every AI-assisted article needs a label
Written content follows a different rule.
Article 50(4) focuses on AI-generated or manipulated text published to inform the public about matters of public interest. That can include areas such as politics, public health, security, fundamental rights, or significant economic and scientific developments.
But using AI somewhere in the writing process does not automatically mean the final article needs a label.
If the text has undergone substantive human review or editorial control, and a person or organization takes editorial responsibility for publishing it, the disclosure obligation does not apply.
The important part is what that review actually involved. Correcting grammar or running a spell-check is not enough. Someone needs to assess the substance, fact-check where necessary, make changes, reject material if needed, and ultimately take responsibility for what is published.
For communications and publishing teams, the practical question is therefore not simply, “Did we use AI?” It is, “Who reviewed the content, what did they review, and who took responsibility for publishing it?”
Machine-readable marking is not the same as disclosure
So far, we have looked at transparency meant for people. However, Article 50(2) adds a different layer: certain AI-generated or manipulated outputs must also be detectable by machines.
Where Article 50(2) applies, providers must ensure that outputs are marked in a machine-readable format and detectable as artificially generated or manipulated. In other words, the information may not be visible to you, but software should be able to identify the content’s AI origin.
When AI analyzes people
Imagine now a company uses an AI tool to infer visitors’ emotions from facial expressions during an event. If people are exposed to that system, the organization deploying it must tell them it is being used. The same applies to biometric categorization systems covered by Article 50, whether the analysis happens in real time or afterwards.
But disclosure is not permission. Some uses of emotion recognition, including certain applications in workplaces and educational institutions, are prohibited under the AI Act. Before asking how to inform people, organizations first need to establish whether the use itself is allowed.
Does Article 50 apply to Swiss organizations?
At this point, a Swiss organization may have an obvious question: we are outside the European Union, so do these rules still concern us?
Potentially, yes. A Swiss company can fall within scope if it provides an AI system on the EU market, or in certain cases when the output of its AI system is used in the EU.
Consider a Zurich-based company offering its own AI service to customers in Germany. Being headquartered in Switzerland does not, by itself, settle the question. What matters is where the system is offered or its output is used, and what role the organization plays.
What should your organization check?
The easiest way to work with Article 50 is to start with the use case. For each AI system, ask four things:
- Where does AI meet people? Map customer interactions, employee-facing systems, and AI-generated or manipulated content.
- What role do we play? Are we the provider, the deployer, or both?
- What does this use require? Check whether disclosure, machine-readable marking, human review, or another measure applies.
- Who owns it? Decide who implements the measure, checks that it works, and documents the process.
For teams already using AI, the starting point is the systems already in use. Map them, establish your role, determine what people need to know, and give someone ownership. Making those checks part of existing AI governance processes helps prevent transparency from becoming an afterthought.
Build practical AI governance skills
Article 50 is one example of the decisions AI governance teams increasingly need to make in practice.
Our AI Governance Professional training helps professionals apply the EU AI Act and other governance frameworks to real organizational contexts, with a practical focus on risk, accountability, and implementation.
