Artificial intelligence can write, speak, design, and respond in a way that is deceptively realistic. That’s impressive. But it’s also the moment when transparency is no longer just a matter of good manners.
Article 50 of the EU AI Act sets forth the circumstances under which people must be informed that they are interacting with AI or viewing AI-generated content.
The basic idea is delightfully simple: Anyone who encounters AI should be able to recognize it in certain situations.
In practice, it's a little less straightforward.
After all, not every use of AI requires the same disclosure. Not every instance of image editing is a deepfake. Not every AI-generated text needs to be marked with a warning label. And a visible label is not the same as a machine-readable identifier.
Transparency therefore requires more than just the statement: “This content may have been created using AI.”
She needs a system.
Much of the discussion surrounding the EU AI Act centers on high-risk AI: candidate selection, creditworthiness, medical applications, and critical infrastructure.
Article 50 works differently.
Transparency requirements may also apply to companies that do not use high-risk AI at all. A chatbot in customer service, an AI-generated promotional video, or an automatically generated informational text may be enough to trigger these requirements.
Article 50 addresses four situations:
The key question, therefore, is not just: How risky is the system?
But also: Can a person understand what role AI plays here?
When an AI system interacts directly with people, it must always be clear that there is no human on the other end.
This applies, for example, to:
This obligation applies first and foremost to the system provider. The provider must design the system in such a way that affected individuals are informed.
The notice should be provided no later than the start of the first interaction. It must not be buried somewhere in the terms of use, hoping to be noticed occasionally.
One possible way to phrase it would be, for example:
"You are communicating with an AI-powered assistant system."
It's neither dramatic nor particularly literary. But it serves a purpose.
An exception may apply if it is obvious to a reasonably informed and attentive person that they are interacting with AI. However, companies should not rely on this too readily.
What may seem completely obvious to the IT department can come across as surprisingly human to customers.
With regard to generative AI systems, Article 50 establishes a technical requirement.
Providers of systems that generate synthetic text, images, audio, or video content must ensure that the output is labeled in a machine-readable format and is recognizable as artificially generated or manipulated.
This involves, for example:
This requirement should not be confused with a visible notice intended for the public.
A machine-readable tag is primarily intended for technical systems. It is designed to enable automated verification of a piece of content's artificial origin.
A visible label, on the other hand, is intended for people.
Therefore, content may be technically marked without any visible indication at first glance. Conversely, a visible indication may be present even though the necessary technical marking is missing.
Companies should consider these two levels separately.
The statement “We’ll just put ‘AI’ under it” does not automatically fulfill the provider’s technical obligations.
The EU AI Act does not intend to turn every spelling correction into a regulatory event.
If AI merely serves a supporting role in standard processing or does not substantially alter the input or its meaning, the technical labeling requirement does not necessarily apply.
Typical examples might include:
However, things get interesting when AI no longer just enhances content but actually generates new content or significantly alters the message.
From a regulatory standpoint, “Please correct typos” has a different implication than “Please draft a complete statement and come up with a suitable justification.”
The software used alone is therefore not the deciding factor. What matters is what it actually does with the content.
Anyone who uses systems for emotion recognition or biometric categorization must inform the individuals affected.
This may include systems that, for example, attempt to infer moods, stress, or certain characteristics from voice, facial expressions, or other biometric features.
The first step is to determine whether the use of this technology is even permitted. Certain applications of emotion recognition in the workplace and in educational institutions are already prohibited in principle, unless a strictly limited exception applies.
Where surveillance remains permissible, it is not enough to simply let the technology run discreetly.
Data subjects must be able to clearly recognize that such a system is in use.
In addition, the GDPR remains applicable. Transparency under the AI Act does not replace a legal basis, a privacy notice, or an assessment of special categories of personal data.
A notice does not make unlawful processing lawful. It merely makes it better labeled.
Special obligations apply to operators who use AI to create or manipulate deepfakes.
This refers to image, audio, or video content that resembles real people, places, objects, or events and may appear to be authentic when it is not.
Such content must always be disclosed as artificially generated or manipulated.
In the case of works that are clearly artistic, satirical, or fictional, disclosure can be handled with greater discretion. It must not unnecessarily spoil the enjoyment of the work. Regulators should also recognize that satire rarely benefits from a permanent warning banner.
In addition, there are AI-generated or manipulated texts that are published to inform the public about matters of public interest.
Disclosure may be required in this case as well.
An important exception applies if the text has been subject to actual human review or editorial oversight and a natural person or legal entity assumes editorial responsibility.
That means more than just a quick glance at the headline.
Human oversight must focus on content. Someone who simply clicks “Publish” may be taking on responsibility, but has not necessarily edited the content.
Article 50 requires that information be provided in a clear and distinguishable manner. It must also be provided no later than the first interaction or exposure and must comply with accessibility requirements.
Therefore, the following are regularly insufficient:
Transparency must not be structured in such a way that it exists in theory but is invisible in practice.
A good tip should be appropriate for the situation in which it is used.
For a chatbot, it can appear at the beginning of the conversation. For a video, it can be displayed at the start. For audio content, an audible notice may be required. For an image, the label must be clearly visible.
The location of the label is not just a minor design consideration. It is a matter of compliance.
Companies should not leave Article 50 entirely up to the marketing department. Nor should the legal department be asked just five minutes before publication whether the new AI video is actually a deepfake.
It makes sense to have a clear process:
The AI Officer can coordinate this process. Working with Legal, Data Protection, Communications, IT, and business units, the AI Officer can develop consistent labeling rules and ensure that not every team comes up with its own little version of a transparency philosophy.
The European Commission has now published a code of conduct on transparency for AI-generated content.
Participation in this program is voluntary. The legal obligations under Article 50 are not.
The code of conduct includes practical measures for providers and operators, particularly regarding machine-readable tagging and the labeling of deepfakes and certain AI-generated texts.
Companies are not required to sign the code. However, those who choose a different approach should be able to document how their own measures meet the legal requirements in an equivalent manner.
So "voluntary" here does not mean "nice, but unimportant."
It means: You can choose the path. But not the destination.
Article 50 of the EU AI Act does not require blanket labeling of all AI use.
He calls for targeted transparency in situations where people encounter AI, consume synthetic content, or are analyzed by certain systems.
For companies, this means:
The real task, therefore, is not to distribute as many small AI signs as possible.
It involves knowing when guidance is needed, who is responsible for implementing it, and how to ensure it is understandable, technically sound, and visible.
Because transparency isn't just a label.
It is a design choice.