Transparency Obligations Under Article 50 EU AI Act Part 1

On July 20, 2026, the European Commission released detailed guidelines clarifying the transparency obligations imposed on providers and deployers of artificial intelligence systems under Article 50 of the AI Act. Although the AI Act formally entered into force on August 1, 2024, its specific transparency framework under Article 50 becomes enforceable on August 2, 2026.

This article forms Part 1 of a four-part breakdown of Article 50, focusing on Article 50(1)—which regulates AI systems designed to interact directly with natural persons.

Who Is Responsible for Providing Information?

To understand compliance under Article 50, one must first grasp the distinct responsibilities assigned to providers and deployers.

Providers

A provider is any natural or legal person, public authority, or agency that develops an AI system (or has one developed) and places it on the market or puts it into service under its own name or trademark, whether for payment or free of charge.

Providers are directly responsible for ensuring compliance with Article 50(1) (interactive AI disclosure) and Article 50(2) (machine-readable marking of synthetic content). Providers must embed transparency features directly into the system’s design and architecture before the AI is placed on the market or put into service. Global reach is a key nuance: providers established outside the European Union are fully subject to these obligations if the output of their AI system is used within the Union. However, downstream, unauthorised, or purely incidental use in the EU does not automatically penalise third-country providers.

Deployers

A deployer is any natural or legal person, public authority, or agency using an AI system under its authority, except when the system is used for purely personal, non-professional activities.

Deployers bear the responsibility for compliance under Article 50(3) (emotion recognition and biometric categorisation disclosures) and Article 50(4) (labelling deepfakes and public-interest text). Exercising authority over an AI system does not necessarily mean having technical control over its internal software. Instead, it means making the decision to deploy the system and directing how and for what purpose it operates. Individual employees working under the instruction and control of an enterprise are not individual deployers, the employer (legal entity) holds deployer status. Furthermore, non-EU deployers fall under the AI Act if they direct or foresee that the AI’s output will be disseminated within the Union.

Dual Roles and Third-Party Intermediaries

An organisation can hold dual roles simultaneously. For example, a firm that develops a generative AI tool in-house and uses it to publish media acts as both a provider and a deployer, requiring adherence to both sets of rules.

Moreover, actors whose role is strictly limited to disseminating third-party content (such as internet hosting services, social platforms, or broadcasters) are not deployers. While they are strongly encouraged to maintain AI markings to preserve information integrity, the AI Act does not assign direct deployer liability to them unless they exercise operational authority over the AI system itself.

Exclusions From the Scope of the AI Act

The AI Act outlines targeted exclusions to prevent over-regulation, though each comes with important legal nuances.

2.1. Deployers Acting in the Context of Purely Personal Non-Professional Activity

Under Article 2(10), the obligations assigned to deployers do not apply to individuals using AI in the course of a purely personal, non-professional activity.

To qualify for this exemption, an activity must meet both criteria: personal and non-professional. If an individual generates content or uses AI in a way that brings regular economic benefits, or uses it within a business, trade, or freelance capacity, the activity is considered professional and falls strictly within scope.

Examples: An individual generating AI-crafted holiday greeting cards or a student using AI to structure homework operates outside deployer obligations.

Important Exception: This exclusion applies only to the deployer. The commercial provider who built and commercialised the underlying tool remains fully obligated to satisfy all provider-side transparency obligations (such as embedding machine-readable markers under Article 50(2)).

2.2. Research & Development

Scientific research and development receive broad latitude under the AI Act, but within clearly defined boundaries.

• Article 2(6) excludes systems or models developed and put into service solely for scientific R&D.
• Article 2(8) excludes research, development, and testing activities conducted prior to placing an AI system on the market.

Crucial Limitations: The moment an R&D system is repurposed or placed on the market for non-research uses, full transparency rules apply. Furthermore, real-world testing (even when conducted inside regulatory sandboxes) is not-exempt from transparency mandates.

2.3. AI Released Under Free and Open-Source Licences

While free and open-source software models enjoy general carve-outs across product legislation, Article 2(12) explicitly establishes that open-source AI systems are not exempt from Article 50 transparency obligations.

A fully functional open-source AI system must satisfy transparency requirements. However, providers of standalone open-source components (such as individual software libraries, datasets, or development toolkits) that do not constitute a complete AI system on their own are not directly bound by Article 50. Still, creators of such tools are encouraged to assist downstream integrators in building compliant systems.

ARTICLE 50(1) AI ACT. TRANSPARENCY FOR INTERACTIVE AI SYSTEMS.

3.1. Main Components: AI Systems Intended to Interact Directly With Natural Persons

For an interactive software tool to trigger the disclosure requirements under Article 50(1), four cumulative criteria must be present:

1. An AI System: The application must satisfy the legal definition of an AI system. Standard, non-AI automated mechanisms, such as traditional out-of-office email replies or basic rule-based phone tree menus, are excluded.

2. Intended to Interact: The system must be designed for bidirectional exchange of information or actions, exhibiting a genuine conversational or responsive character across text, audio, visual, or physical mediums. Passive systems that merely collect data (such as automated facial-recognition entry gates) do not meet this definition.

3. Direct Interaction: The interaction must occur directly in real-time or near real-time without an intermediary human vetting or altering the content. If a human reviews, approves, and transmits the output as the primary interlocutor, the direct interaction chain is broken. However, if automated systems blend AI responses with light human curation, the AI-generated elements must still be disclosed.

4. With Natural Persons: The interaction must directly reach a human being (whether an end-consumer or a professional user). Closed industrial setups, machine-to-machine backend workflows, and factory automation in human-excluded zones fall outside this mandate.

Special Rules for AI Agents: Autonomous AI agents capable of executing multi-turn actions (such as managing bookings, processing payments, or negotiating contracts) are fully covered by Article 50(1). They must be designed to disclose not only their artificial nature but also the identity of the person or entity on whose behalf they operate. If a provider cannot predict every deployment context, the agent must be architectural programmed to declare its AI origin whenever human interaction is reasonably foreseeable.

Examples:

• In Scope: Customer service chatbots, AI voice assistants, virtual support avatars, interactive coding agents, and conversational social media bots.

• Out of Scope: Industrial assembly line robots, content recommender algorithms, email spam filters, single-turn navigation tools, and predictive maintenance tools.

3.2. Information Obligation Under Article 50(1) AI Act

Providers of interactive AI systems must design their systems so that users are clearly informed that they are communicating with an AI.

• Timing: Notification must occur before or during the interaction, and at the latest at the moment of first interaction.

• Delivery Formats: Disclosures can take several forms depending on the medium:

  • Textual: Clear UI banners, persistent badges, or conversational greetings (e.g., “Hello, I am an AI assistant”).

  • Auditory: Spoken opening statements in voice calls or clear audio cues.

  • Visual: Recognizable icons, watermarks, or distinct interface framing.

  • Multimodal: Combining visual and auditory cues to reinforce user clarity.

• Vulnerable Audiences: If an AI system is intended or reasonably likely to interact with vulnerable groups, such as children, the elderly, or persons with disabilities, the disclosure must be tailored accordingly. Disclosures aimed at children must be age-appropriate, child-friendly, and easy to grasp.

• Inadequate Disclosures: Burying disclosures deep inside Terms & Conditions, using hidden metadata, relying on ambiguous terms like “digital assistant,” issuing site-wide generic notices (“This website uses AI”), or using complex jargon (“Powered by LLM architecture”) do not meet legal standards.

3.3. Exceptions to the Information Obligation Under Article 50(1)

3.3.1 Exception for Obvious Interaction With an AI System

Providers are exempt from issuing an explicit AI disclosure if the artificial nature of the interaction is obvious to a natural person who is reasonably well-informed, observant, and circumspect, taking into account the context and target audience.

The “average consumer” benchmark: is borrowed from EU consumer protection law, this standard evaluates the average recipient within the intended or foreseeable audience. Because this exception deprives individuals of an explicit warning, it is interpreted restrictively. General public awareness that AI chatbots exist does not mean every specific chatbot interaction is “obvious”. If an interactive system is restricted exclusively to specialised professionals (such as software engineers using coding assistants or medical staff using diagnostic software), the threshold for obviousness is easily met. However, if the system interacts with the general public or vulnerable consumers, the exception can rarely be claimed.

3.3.2 Exception for AI Systems Authorised by Law for Law Enforcement Purposes

An exemption applies if an interactive AI system is authorised by law to detect, prevent, investigate, or prosecute criminal offences (e.g., automated AI tools used by law enforcement investigators).

There is a major exception to this exemption: if the AI system is public-facing and provides a mechanism for citizens to report crimes or submit evidence (such as an automated police reporting chatbot or fraud hotline), the system must still declare its AI nature. Citizens interacting with public law enforcement reporting tools must always be aware they are dealing with an AI system.

3.4. Interplay With Other Union Legal Acts

Article 50(1) complements existing European Union legislation:

• Consumer Protection Law: Under the Unfair Commercial Practices Directive (UCPD) and Consumer Rights Directive (CRD), traders must provide clear pre-contractual information regarding essential product characteristics. Disclosing that a service is AI-driven remains mandatory under consumer law even if the interaction might be considered “obvious” under the AI Act.

• Digital Services Act (DSA): Rules under the DSA requiring online platforms to disclose recommender system parameters complement Article 50(1) by covering algorithmic content ranking.

• Data Protection (GDPR): Mandatory privacy notices under the GDPR serve distinct legal purposes and must be provided alongside AI origin disclosures without replacing one another.

In the next part we will discuss requirements to the AI systems generating or manipulating synthetic image, video, audio or text content.

Link to the Guidelines