AI Act Article 50: AI Content Labeling and Code of Practice from August 2, 2026 

On August 2, 2026, key transparency obligations under Article 50 of the AI Act regulation will enter into full force. These rules aim to counter the risks of disinformation and manipulation by requiring businesses to clearly label content generated or manipulated by AI.

To assist the market with this adaptation, the European AI Office has published a Code of Practice prepared by independent experts.

What Does Section 50 of the AI Act Cover? AI Transparency Requirements in Practice

Article 50 of the AI Act regulation, titled “Transparency obligations for providers and deployers of certain AI systems,” serves as the legal foundation for informing the public when they are interacting with artificial intelligence.

This article regulates the following key areas:

• Obligation to inform about human-machine interaction (for providers): AI systems intended to interact directly with natural persons (e.g., chatbots, voice assistants) must be designed and developed to clearly communicate to the user that they are dealing with AI. Situations where this is completely obvious from the context of use represent an exception.
• Machine-readable marking (for providers): Creators of generative AI that produces synthetic audio, images, video, or text (including general-purpose AI models) must ensure that the outputs of their systems are marked in a machine-readable format. This allows other software to easily detect that the content has been artificially generated or manipulated. These solutions should be effective, interoperable, and robust against attempts to remove them.
• Transparency of biometric and emotion recognition systems (for deployers): Entities utilizing emotion recognition or biometric categorization systems are under a strict obligation to inform natural persons that they are being subjected to such technologies.
• Labeling of deepfakes and texts (for deployers):
  • Entities publishing images, video, or audio that look or sound strikingly similar to real people, places, or events (deepfakes) must clearly disclose that the content was generated or manipulated by an AI system. Certain allowances are provided for artistic, satirical, or cinematographic works – in such cases, the label can be placed in a way that does not interfere with the artistic expression, but the presence of AI must still be disclosed.
  • Texts generated or manipulated by an AI system that are published to inform the public on matters of public interest must also be labeled. The only way to be exempted from this obligation is if the text undergoes verification by a human (editorial review) and a specific person or company assumes full editorial responsibility for the publication.

The above information must reach the user in a clear, distinct manner, and at the latest during the first interaction or first contact with the generated content. It is also required that these labels meet accessibility standards for individuals with disabilities.

The AI Act Code of Practice – What Is It and How Does It Help Fulfill the Obligations Under Article 50?

The direct basis for drafting the Code of Practice is Article 50 of the AI Act itself. In this provision, the EU legislator explicitly encourages market participants to collaboratively create and adhere to codes of conduct that help businesses smoothly and effectively implement these new obligations into practice.

In response to this call, the EU Code of Practice was developed through a multistakeholder process by independent experts. This practical tool supports companies in meeting strict legal requirements by providing standardized operating methods. Among these, it offers free, ready-made icons (e.g., featuring text like “AI GENERATED” or “AI MODIFIED”) that organizations can directly overlay onto generated materials.

The Code of Practice itself holds the status of an entirely voluntary technical manual. While companies are under no legal compulsion to sign up for it, adopting its rules is currently considered the safest business path.

Following the Code makes it easier to prove compliance with the AI Act to supervisory authorities, removing the risky and costly burden of potentially independently convincing regulators that proprietary content-marking methods are sufficient and meet EU standards.

Who is subject to the obligations under Article 50 of the AI Act? Providers and entities that use AI systems

The new obligations impact a broad market. Two main groups of entities can join the Code as signatories:

• Providers of AI systems – in simple words, these are technology companies that build and develop AI models and systems from the ground up.
• Deployers of AI systems – this is by far the largest and most diverse business segment. Simply put, this group includes companies and institutions that do not build AI tools from scratch but deploy ready-made, external solutions within their daily operations. This category covers publishers, editorial newsrooms, marketing agencies, HR departments, e-commerce stores, and more.

AI Act Code of Practice: Requirements for AI Providers and Users

The guidelines are structured around three distinct compliance levels. By using the keywords “will” (mandatory), “encouraged” (recommended), and “may” (strictly voluntary), the document clearly signals what is required versus what is rather optional.

Obligations of Providers of AI Systems:

• Machine-Readable Marking (Commitment 1):
  • Must (Will): Implement multilayered technical solutions so that system outputs are marked in a machine-readable format (machine-readable marking techniques). This includes cryptographically secured metadata and imperceptible watermarks embedded directly within the file. Providers must also legally prohibit their users from removing or modifying these marks in their terms of service.
  • Recommended (Encouraged): Enrich metadata with additional provenance information (provenance), such as the name of the AI tool, the name of the company (provider), and a precise description of the modifications made to the content.
• Detection Mechanisms (Commitment 2):
  • Must (Will): Publicly disclose and provide, as a rule free of charge, tools or services that allow users, authorities, and researchers to verify (detect) whether a given piece of content originated from their system. These tools must adhere to strict privacy requirements, including a “zero retention” policy (immediate deletion of the uploaded file after analysis).
  • Optional (May): Given the potentially lower reliability of text watermarking compared to images or video, providers may restrict access to text detection tools exclusively to verified experts (e.g., law enforcement authorities or researchers).
• Quality and Compliance Requirements (Commitment 3 & 4):
  • Must (Will): Ensure that their marking and detection systems are effective, reliable, interoperable, and resilient against cyberattacks or attempts to remove watermarks. Providers must also document compliance processes, rigorously test solutions before market release, and train their personnel.

Obligations of Deployers Utilizing AI Systems

• Disclosure of Deepfakes and Texts (Commitment 1):
  • Must (Will): Clearly and prominently disclose AI-generated content at the latest during the user’s first contact with the material. For images and video, companies must display a visible, unobstructed icon or label featuring a large, distinct “AI” text. For audio materials, a clear spoken warning (audible disclaimer) must be played at the very beginning of the track.
  • Recommended (Encouraged): Expand the primary “AI” icon with an additional layer of information specifying whether the content was entirely generated (“AI GENERATED”) or merely modified (“AI MODIFIED”). For long-form video or audio, it is recommended to repeat the AI disclosure at regular intervals (e.g., after commercial breaks).
• Internal Processes (Commitment 2):
  • Must (Will): Implement or update internal compliance procedures (proportionate to the company’s size) that clearly document how the company executes its content-marking duties. They must also sensitize and train employees regarding these new legal obligations.
  • Recommended (Encouraged): Provide accessible feedback channels (such as contact forms) through which audiences can report missing or incorrect AI labels to the company.
• Exceptions for Artistic Creation (Commitment 3):
  • Must (Will): For deepfakes that form part of an artistic, satirical, fictional, or cinematographic work, the labeling can be deployed in a way that is consistent with the author’s artistic or creative intent. The information about the use of AI must remain fully noticeable to the viewer, but the presentation format can be adapted so that it does not disrupt the aesthetics or the narrative delivery of the work.
• Editorial Oversight (Commitment 4):
  • Must (Will): Signatories who publish AI-generated text regarding public affairs without labeling it must prove that it has undergone substantive human control. It is mandatory to implement proper human review and editorial control procedures, explicitly identifying a specific natural or legal person who carries full editorial responsibility for the publication.

Penalties for Violating Article 50 of the AI Act – What Are the Consequences of Failing to Label AI Content?

Failure to prepare for the new obligations under Article 50 of the AI Act may result not only in the risk of an audit, but also in serious financial penalties and a loss of customer trust. Therefore, companies using AI systems should verify now whether their content, processes, and procedures meet transparency requirements.

We will soon publish a comprehensive AI Guide, in which we will discuss in greater detail the impact of technology regulations on doing business in Poland. If your organization uses artificial intelligence in its day-to-day operations, please contact us – we can help you assess risks, train your team, and prepare documentation compliant with the AI Act.