The European Union begins enforcing, starting August 2, 2026, the transparency obligations outlined in Article 50 of the AI Regulation. The regulation affects providers of generative models, platforms, chatbot developers, and companies publishing content created or manipulated with AI. The rule does not require marking every use of these tools, but it does mandate informing when an interaction or file could be mistaken for authentic due to automatic generation or synthetic content.
The key points of AI content labeling in 30 seconds
- Chatbots must disclose that the user is interacting with artificial intelligence unless it is obvious.
- Providers of generative systems will need to embed detectable technical signals in texts, images, audio, and videos.
- Deepfakes must include a visible warning when they could appear real.
- Informational texts reviewed by a person and published under editorial responsibility are exempt from specific labeling requirements.
- Penalties can reach €15 million or 3% of the company’s global annual turnover.
Article 50 introduces two levels of transparency. The first is technical and mainly concerns developers or vendors of these systems. The second is visible to the public and impacts users of these tools who provide services or distribute certain content.
This distinction is relevant for digital media, PR agencies, social platforms, software companies, marketing departments, and AI providers. Not all will have to do the same; many will need to review their products and internal processes.
Providers must embed machine-readable signals
Companies offering systems capable of generating or manipulating text, images, audio, or video must incorporate mechanisms to detect the artificial origin of content.
The regulation does not specify a single technology. These signals may include metadata, digital watermarks, provenance credentials, or other methods that enable recognition that the file was generated or altered using AI.
The solution must be effective, interoperable, sufficiently robust, and reliable within the current state of technological development. This prevents demanding the same level of resistance across all formats and systems but also discourages simple, easily removable marks or signals that only the provider can interpret.
Technical requirements do not necessarily mean adding a visible label to each file. A synthetic image, for instance, can contain information detectable only by automated systems without showing an explicit warning to the user. Visible labeling occurs in specific situations, especially when content is a deepfake.
Exceptions are allowed for tools that perform minor edits or changes that do not substantially alter the original material’s meaning. Color correction, noise reduction, or certain automated adjustments are treated differently from creating an entirely fictitious scene.
What content qualifies as deepfake
The European regulation uses a broad definition. Deepfake refers to an image, audio, or video generated or manipulated with AI that resembles real people, objects, locations, entities, or events and that could be mistaken for authentic.
It is not limited to fake videos of politicians or celebrities.
It can also involve cloned voices, real photographs with substituted persons, edited videos adding actions that never occurred, or recreated locations presented as genuine recordings.
In the business context, less obvious cases may arise. For example, a real estate agency showing an AI-generated furnished apartment that doesn’t exist, a company using synthetic voice of a real individual, or a campaign recreating an event could fall under this obligation if the audience might believe the content is genuine.
Professionally published content must clearly and visibly disclose if it was generated or manipulated with AI.
The warning should appear from the first exposure. Hiding it in terms of use, on a secondary page, or in a disappearing disclaimer during download may not be sufficient.
Visible labeling must survive content sharing
One technical challenge is maintaining the warning when content leaves the platform where it was published.
An image can be downloaded, cropped, and shared on other social media, and a video can circulate via messaging apps without preserving any accompanying text. Therefore, the European Commission recommends that labels be embedded within the content itself or linked through provenance mechanisms that travel with the file.
This presents challenges for platforms and developers. Metadata can be lost during recompression or conversion. Visible marks may be cropped out. Invisible signals may degrade after multiple transformations.
The regulation does not guarantee that a label cannot be removed, but it does require responsible parties to adopt reasonable and sufficiently resilient solutions based on the format used.
The European Commission has introduced three free icons: one general icon indicating AI involvement, another for content fully generated by AI, and a third for content partially modified by humans.
Use of these icons is voluntary. Companies can create their own labels, provided they are clear and understandable. Displaying the EU icons does not guarantee compliance if they are shown briefly, are illegible, or do not specify what was manipulated.
Chatbots must identify themselves from the first interaction
Article 50 also governs systems that interact directly with people.
Support chatbots, sales assistants, or conversational agents must disclose that the user is talking with AI, unless it is obvious to a reasonably informed and attentive person.
This disclosure should occur no later than the first interaction.
Applicable to web services, mobile apps, voice assistants, messaging bots, and systems embedded in customer service platforms. It should not be limited to mentions in privacy policies or legal documents that users do not review during conversations.
The regulation also applies to certain emotion recognition and biometric categorization systems. Users must be informed when these technologies are used, provided the use aligns with the regulation.
Exceptions for informational texts and reviewed content
The rules for AI-generated texts are less strict than for deepfakes.
Visible labeling is mainly required when an AI-created or manipulated informational text is published to inform the public about important topics and has not undergone human review or does not have a responsible publisher.
A media outlet can use AI for transcribing interviews, organizing documents, summarizing reports, correcting style, or preparing initial drafts without necessarily labeling the final article.
This exception requires real human review and identifiable editorial responsibility. It is not an automatic process where anyone simply hits “publish” without verifying the output.
For newsrooms, the focus is on who used AI, what role it played, who verified the information, and who takes responsibility for the published content.
Artistic, creative, satirical, or fictional works have a more flexible approach. They should disclose AI involvement appropriately, but warnings should not unnecessarily detract from the artistic experience.
Limited transition period for existing systems
Obligations under Article 50 start on August 2, 2026. However, AI systems already marketed before this date have an extra period to adapt certain marking mechanisms.
This does not mean a delay of all obligations. Newly launched systems must comply from the start, while platforms and companies should begin assessing which visible warnings apply now.
Preparation involves more than just adding an icon. Organizations must inventory their AI tools, differentiate fully synthetic from partially edited content, retain provenance information, and assign who approves each publication.
Coordination with external providers is also necessary. A company may be responsible for warnings even if the content were generated by third-party systems.
Fines can reach up to 3% of global turnover
Non-compliance can result in penalties of up to €15 million or, for companies, up to 3% of their annual worldwide turnover from the previous year, whichever is higher.
Authorities will consider factors like the severity, duration, and intent of violations, as well as the size of the organization and corrective measures taken.
The practical application of Article 50 will depend on future decisions by national authorities and the evolution of technical standards. Nonetheless, the core principle is clear: when content or interactions could be mistaken for real or human, adequate disclosure is required.
The EU is not demanding labeling for every retouched photo, edited text, or minor AI use. Instead, it seeks transparency where technology might influence perceptions of authenticity.
Frequently Asked Questions
Is it necessary to label all AI-generated images?
Not necessarily with a visible warning. Technical signals must be embedded, but explicit signage is mainly required when content could be mistaken for a deepfake or appears convincingly real.
Does a media outlet need to disclose using ChatGPT?
Not always. If the content was reviewed by a person and the publisher takes editorial responsibility, the specific requirement for labeling AI-generated informational texts does not apply.
Must chatbots always identify themselves?
They must do so by the first interaction at the latest, unless it is obvious to a reasonably attentive person that they are interacting with AI.
Are the European Commission icons mandatory?
No. They are voluntary tools. Companies can design their own labels, provided they are clear, visible, and understandable.
