目錄
The information in this article is current as of August 2026. Enforcement practices, technical standards and regulatory approaches adopted by individual EU member states may continue to change.
Article 50 of the EU AI Act became applicable on August 2, 2026. AI systems that interact directly with people must disclose that users are communicating with AI, while generative AI outputs must include machine-readable markings. Additional disclosures to the general public may also be required when organizations use emotion-recognition systems, biometric categorization, deepfake content or public-interest text that has not undergone meaningful human review.
The changes are likely to become most visible in customer service, telephone support, AI-generated virtual characters, advertising and social media content. Services where users could not previously tell whether AI was operating behind the scenes must now provide a clear notice at the first point of interaction or exposure.
As AI labels become more common, another issue emerges: disclosure fatigue. If every website, customer service channel and social media post includes an AI notice, will people eventually close or ignore them in the same way they dismiss cookie consent banners?
This article explains the scope of Article 50, its exceptions, transitional periods and the key issues that ordinary users and small and medium-sized businesses need to understand.
When Did EU AI Act Article 50 Take Effect?
The EU AI Act entered into force on August 1, 2024, but its provisions apply according to a phased implementation schedule.
The transparency obligations under Article 50 became applicable on August 2, 2026. Enforcement is carried out by the European AI Office, the competent authorities of EU member states and, in certain circumstances, the European Data Protection Supervisor.
Article 50 divides responsibilities between providers and deployers.
A provider is generally a company that develops an AI system or places it on the EU market under its own name or trademark. A deployer is an organization that uses the AI system in a website, customer service operation, media activity, marketing campaign or other professional context.
The same company may act as both a provider and a deployer.
A company that develops its own customer service chatbot and offers it to other businesses may be considered a provider. If it integrates a third-party generative AI system into its own website support service, it is more likely to be acting as a deployer.
Which Situations Are Covered by EU AI Act Article 50?
Article 50 mainly covers four categories of transparency obligations.
It also requires all relevant notices to be clear and distinguishable and to appear no later than the first interaction with the AI system or the first exposure to the relevant content.
AI Systems That Communicate Directly with People
AI systems that engage in two-way communication with people must inform users that they are interacting with AI.
This may apply to business chatbots, telephone support systems, virtual assistants and certain AI agents.
If an ordinary person can clearly recognize from the context that the system is AI, an additional notice may not be required.
However, EU guidance applies this exception narrowly. A company cannot omit the disclosure simply because it assumes that users should be able to tell that the service is automated.
Machine-Readable Markings for Generative AI Outputs
AI systems capable of generating or manipulating text, images, audio or video must generally ensure that their outputs contain machine-readable markings.
These markings should allow detection systems to identify whether content was generated or modified using AI.
This obligation primarily applies to AI system providers. It does not mean that every piece of AI-generated content must include a visible watermark.
Machine-readable markings support technical detection. Text, icons or audio messages that ordinary users can perceive belong to a separate layer of disclosure obligations.
If an AI system performs only spelling correction, formatting or another standard editing task without substantially altering the input data or its original meaning, machine-readable markings may not be required.
Emotion Recognition and Biometric Categorization
Deployers of emotion-recognition or biometric-categorization systems must inform the people whose data is being analyzed.
For example, a customer service center may need to provide a notice if it uses AI to analyze a caller’s emotions. The same may apply when cameras, audio recordings or voice data are used to infer a person’s emotional state.
This obligation may apply not only to real-time analysis but also to analysis performed after the interaction has taken place.
A disclosure obligation does not automatically mean that the system is lawful.
Certain uses of emotion recognition in workplaces or educational institutions may still be prohibited or restricted under other provisions of the EU AI Act.
Deepfakes and AI-Generated Public-Interest Text
Images, audio recordings or videos may qualify as deepfakes when AI is used to generate or manipulate content that resembles a real or potentially real person, object, location, organization or event and could cause an ordinary person to believe that the content is authentic.
Deployers must clearly disclose that the content was generated or manipulated using AI when the public first encounters it.
A machine-readable marker hidden inside the file is not sufficient. The notice must also be visible or audible to ordinary users.
AI-generated or AI-modified text published to inform the public about matters of public interest must also be clearly disclosed when it has not undergone meaningful human review or editorial control.
Matters of public interest may include:
- Politics and democratic processes
- Judicial and law-enforcement matters
- Public safety
- Public health
- Environmental protection
- Consumer safety
- Finance
- Science
- Culture
- Other issues of significant public debate
EU AI Act Implementation Timeline: From Legislation to Transparency Guidance
In November 2025, the European Commission began developing a Code of Practice for the transparency of AI-generated content.
Following participation from industry representatives, civil society organizations and technical experts, the final version was published on June 10, 2026.
On July 20, 2026, the European Commission issued implementation guidance for Article 50 and released AI content-labelling icons.
The guidance explains how providers, deployers, deepfake content and public-interest text should be assessed.
Generative AI systems that were already placed on the market before August 2, 2026 receive a limited transitional period only for the machine-readable marking and detectability requirements under Article 50(2).
Those systems may delay compliance with that specific requirement until December 2, 2026.
There is no equivalent general grace period for chatbot disclosures, emotion-recognition notices, deepfake labelling or disclosures for public-interest text.
Content generated before August 2, 2026 does not need to be labelled retroactively. However, the EU still encourages deployers to provide voluntary disclosures where technically and practically feasible.
What Is AI Disclosure Fatigue?
Disclosure fatigue is not a legal term used in the EU AI Act. It is a user-experience issue that has emerged as Article 50 begins to apply.
When customer service channels, advertisements, videos, social media posts and voice services all begin displaying AI notices, users may quickly become accustomed to closing or ignoring them.
WIRED has compared this possibility with the cookie banners that became widespread following the introduction of the GDPR. The notices remain visible, but fewer people may actually read them.
The problem is not only the number of notices. It is also whether situations involving very different levels of risk are presented using the same warning.
Correcting the spelling in a paragraph, using an AI chatbot to answer a return question and generating a fake video of a political figure clearly involve different levels of risk.
If all three situations are presented with the same icon, the same text and the same visual emphasis, users will have difficulty understanding what AI actually did.
Over time, the label may simply become part of the interface and lose its ability to attract attention.
A more reasonable approach is to use a short and direct first-layer notice, such as:
- You are interacting with an AI customer service system.
- Parts of this video were modified using AI.
Users who want more information could then access a second layer explaining the technology used, the scope of the modification and the organization responsible.
How Should the EU AI Labelling Icons and Code of Practice Be Used?
The EU currently provides three AI content icons:
- A general AI icon
- Fully generated by AI
- Partially modified using AI
Each icon is available in black, white and semi-transparent black-and-white versions.
The icons may be used free of charge, and organizations do not need to display the name of the European Commission or the AI Office.
These icons are not the only legally acceptable format.
The disclosure obligations under Article 50 are legally binding, but companies may use the EU icons or create alternative labels that are equally clear, recognizable and accessible.
In principle, the EU icons should be visible when an ordinary person first encounters the content. They should not be hidden behind other interface elements.
Where possible, the label should remain attached when the content is downloaded, copied or forwarded.
Text accompanying the icon should use language that ordinary people can understand. Companies should avoid relying only on technical abbreviations or vague terminology.
The Code of Practice for the transparency of AI-generated content is a voluntary compliance tool.
Signatories may use the marking, detection and disclosure measures described in the Code to build evidence that they comply with Article 50(2), (4) and (5).
Companies that do not join the Code may design their own alternative systems, but they must be able to demonstrate that those measures provide an equivalent level of effectiveness.
Signing the Code or displaying an EU icon does not automatically mean that every legal requirement has been satisfied.
Which AI Content May Not Require Labelling?
Article 50 does not require every use of AI to receive the same label. Different exceptions apply to different obligations.
If AI is used only for spelling correction, formatting or another standard editing process without substantially altering the input data or its original meaning, the provider may not be required to add a machine-readable marker to the output.
Using AI to identify typographical errors is legally different from asking AI to generate an entire article.
The length of the text is not the only factor. The assessment must also consider whether the original information, viewpoints and meaning were substantially changed.
AI-generated public-interest text may not require an additional visible AI disclosure when it has undergone meaningful human review or editorial control and a natural or legal person assumes final editorial responsibility.
Human review cannot be limited to spelling, grammar or layout.
The reviewer should possess relevant knowledge of the subject, be able to verify facts, arguments and sources, and have the authority to modify, return or reject the content.
Therefore, using AI to assist with a blog post or media article does not automatically mean that every publication requires an AI label.
If the content concerns a matter of public interest but has been meaningfully reviewed by a person with relevant knowledge, and an identifiable individual or organization accepts final editorial responsibility, the exception may apply.
Deepfake content that is clearly part of an artistic, creative, satirical, fictional or similar work must still disclose that AI was used to generate or modify it.
However, the disclosure may be presented in a way that does not interfere with the display or enjoyment of the work.
This is a relaxation of the disclosure format, not a complete exemption.
How Can Small and Medium-Sized Businesses Comply with EU AI Act Article 50?
Small and medium-sized businesses do not need to begin by attaching the same icon to every item that involves AI.
A more practical approach is to identify how the company uses AI and then determine its obligations according to its role and the type of content involved.
Step 1: Determine Whether the Company Is a Provider or a Deployer
A company that develops an AI product or places an AI system on the EU market under its own name may be considered a provider.
A company that uses ChatGPT, Claude, Gemini or another third-party tool to produce customer service responses, images, videos or articles is generally more likely to be a deployer.
The same company may have different roles across different services. It is not sufficient to classify the entire organization only once.
Step 2: Identify AI Situations Encountered by the Public
Priority areas for review include:
- Website chatbots and telephone customer service
- AI virtual characters or digital replicas
- Emotion recognition and biometric categorization
- AI-generated or AI-modified images, audio and video
- Text concerning politics, public safety, health, finance or other matters of public interest
- Public advertisements, social media posts and marketing materials
Each situation may require a different form of disclosure.
A general statement in the website footer saying that the company may use AI will usually not be sufficient to meet every obligation.
Step 3: Separate Technical Markings from User-Facing Notices
Machine-readable markings allow systems and detection tools to identify AI-generated content.
Clear labels allow ordinary users to understand that content was generated or modified using AI.
A company cannot complete one requirement and assume that the other has also been addressed.
Deepfake content in particular cannot rely only on technical markers hidden within a file.
Step 4: Keep Records of Human Review and Publication Decisions
Companies should document:
- Which AI system was used
- Who reviewed the content
- Which aspects were reviewed
- Whether the reviewer had authority to modify or reject publication
- Who accepted final editorial responsibility
If a company claims that public-interest text received meaningful human review, a simple checkbox marked “reviewed” is unlikely to be sufficient.
The review process should be able to show how facts were verified, how sources were checked and who assumed final editorial responsibility.
Step 5: Review Label Placement and Accessibility
The notice should appear at the first interaction or exposure to the content.
It should not be hidden in terms of service, a website footer or a page that requires several clicks to locate.
Text beside the icon should directly explain what the AI did, for example:
- You are interacting with an AI customer service system.
- This image was fully generated by AI.
- Parts of this video were modified using AI.
- This article was generated by AI and has not undergone human editorial review.
Companies serving users in the European Union should assess whether the EU AI Act applies to those services even when the company itself is established outside the EU.
What Are the Penalties for Violating EU AI Act Article 50?
Article 50 is primarily enforced by market-surveillance authorities in the individual EU member states.
The AI Office may supervise certain AI systems directly supplied by providers of general-purpose AI models, as well as specific systems integrated into very large online platforms or search engines.
A company that breaches the relevant obligations may face an administrative fine of up to €15 million or 3% of its total worldwide annual turnover for the preceding financial year.
The actual penalty will take into account the size of the company, the circumstances of the infringement and the principle of proportionality.
Specific considerations may apply to small and medium-sized enterprises and small mid-cap companies.
The changes introduced by Article 50 involve more than adding an AI icon to the corner of an image.
Companies must first determine who created the content, what the AI changed, whether meaningful human review took place and who accepts responsibility for publication.
Whether ordinary users receive useful information also depends on whether the notice clearly explains what the AI actually did.
Too little disclosure can be misleading. Using the same warning for every situation may create disclosure fatigue.
The next stage of implementation will depend on whether regulators and businesses can achieve a reasonable balance between risk, the intensity of the disclosure and the user experience.
Frequently Asked Questions
Article 50 of the EU AI Act became applicable on August 2, 2026.
Although the EU AI Act entered into force on August 1, 2024, its provisions apply according to a phased implementation schedule.
No.
The provider’s obligation to include machine-readable markings is separate from the deployer’s obligation to provide a visible or audible notice to ordinary users.
The main categories requiring clear user-facing disclosures include deepfake content and public-interest text that has not undergone meaningful human review.
Not necessarily.
If an article concerns a matter of public interest but has undergone meaningful human review by a person with relevant knowledge, and a natural or legal person accepts final editorial responsibility, an additional disclosure under Article 50 may not be required.
Spelling, grammar or layout checks alone do not constitute meaningful review.
Content generated before August 2, 2026 does not need to be labelled retroactively.
However, the European Commission still encourages deployers to provide voluntary disclosures where technically and practically feasible.
Only a limited transitional period applies.
For generative AI systems placed on the market before August 2, 2026, the machine-readable marking and detectability obligations under Article 50(2) may be delayed until December 2, 2026.
The other transparency obligations generally apply from August 2, 2026.
Yes. AI systems that interact directly with people must generally disclose that the user is communicating with AI no later than the first interaction, unless this is already obvious from the context to a reasonably informed person.
Violations may result in administrative fines of up to €15 million or 3% of the organization’s total worldwide annual turnover for the preceding financial year, subject to proportionality and other enforcement considerations.