Limited Risk AI · Chapter IV

Article 50: Transparency Obligations

Article 50 establishes transparency obligations for AI systems that interact with people, generate content, or process biometric data — regardless of whether those systems are high-risk. Unlike Articles 9–15, Article 50 applies across a broad range of AI products and is already in force. If you build a chatbot, a content generation tool, or a system that processes emotion or biometric data, Article 50 almost certainly applies to you.

Applies to: Providers and deployers of conversational AI, synthetic content generators, emotion recognition systems, biometric categorisation systems, and AI systems that generate text about public interest matters. Article 50 applies regardless of risk tier — it is not limited to high-risk AI systems.

Five Distinct Obligations in Article 50

Article 50 is not a single obligation — it contains five subsections, each with a different scope, applying to different actors (providers vs. deployers) and different types of AI system. The distinctions matter: a provider obligation cannot be satisfied by a deployer's disclosure, and vice versa.

Art. 50(1)Providers

Chatbot and conversational AI disclosure

Providers of AI systems intended to interact directly with natural persons must ensure that systems are designed and developed so that the natural persons concerned are informed that they are interacting with an AI system. This obligation applies unless the AI nature is obvious from the circumstances and context of use.

Practical note: The disclosure must come from the system design — providers cannot simply rely on deployers to add a disclosure in their UI. The Act requires the AI system itself to be capable of informing users.
Art. 50(2)Providers

Machine-readable marking of synthetic content

Providers of AI systems that generate synthetic audio, image, video, or text content must ensure the outputs are marked in a machine-readable format and are detectable as artificially generated or manipulated. This applies to content generation systems broadly — not only deepfake-specific tools.

Practical note: Technical standards for machine-readable marking (watermarking, metadata embedding) are still being developed by ETSI and CEN/CENELEC. Until harmonised standards are published, providers must implement robust marking systems using available techniques.
Art. 50(3)Deployers

Emotion recognition and biometric categorisation disclosure

Deployers of emotion recognition systems or biometric categorisation systems must inform the natural persons exposed to those systems of their operation. This is a deployer obligation — the company using the system bears the disclosure responsibility, not necessarily the system provider.

Practical note: Note the distinction: providers ensure the system can make disclosures (50(1)); deployers ensure disclosures are actually made in their specific deployment contexts (50(3)). Both obligations can apply simultaneously.
Art. 50(4)Deployers

Deepfake and synthetic media disclosure

Deployers who use an AI system to generate or manipulate image, audio, or video content constituting a deepfake must disclose that the content has been artificially generated or manipulated. The disclosure must be clearly visible in or accompanying the content.

Practical note: This obligation applies even when the deepfake was created using a provider's general-purpose image or video generation tool. The deployer's use for deepfake purposes triggers the disclosure obligation regardless of the original tool's intended purpose.
Art. 50(5)Deployers

AI-generated text about public interest matters

Deployers of AI systems that generate text published for the purpose of informing the public on matters of public interest must disclose that the text has been artificially generated. This targets AI-generated news articles, political commentary, and public information content.

Practical note: Article 50(5) includes an exception for content that has undergone substantial human review or editing, where the human contribution is sufficient to warrant the publication being attributed to a human author.

Does Article 50 Apply When Users Are Businesses?

Article 50 refers to "natural persons" — individual people, not corporate entities. The obligations are triggered by interactions with or exposure of individual people to the AI system, not by the commercial relationship through which the product is sold. A B2B SaaS product is still subject to Article 50 if the end users of that product are individual people, even if the customer relationship is with a company.

For example: a company that sells an AI recruiting tool to enterprise customers (B2B) is building a product whose end users are individual job applicants (natural persons). Article 50(1) applies to the chatbot interface if it's used in candidate screening. The B2B commercial structure does not create an exemption.

Interaction With GDPR and the Digital Services Act

Article 50 operates alongside — and does not replace — transparency obligations from other EU regulations. GDPR Article 22 requires transparent, meaningful information when automated decision-making has legal or significant effects on individuals. The Digital Services Act imposes transparency requirements on algorithmic recommendation systems operated by Very Large Online Platforms. The EU AI Act's Article 50 is narrower in some ways (it applies at the system level) and broader in others (it applies to all providers regardless of platform size).

Compliance with one regulation does not automatically satisfy the others. A GDPR privacy notice that mentions AI processing does not satisfy Article 50(1)'s requirement that the AI system itself inform users of its AI nature at the point of interaction. Treat each regulatory requirement independently and design disclosure systems that can satisfy all simultaneously.

Practical Implementation: UI Patterns That Satisfy Article 50

Article 50 does not prescribe specific UI patterns, but regulators have indicated what constitutes adequate disclosure through guidance and enforcement actions in adjacent areas. The following approaches are consistent with the requirement:

Chatbot opening disclosure

Display a clear, visible message at the start of every conversation session stating that the user is interacting with an AI. Repeat the disclosure if the conversation resumes after a break. Do not rely solely on a product name or icon to imply AI nature.

Content generation labels

Label AI-generated content with a visible, persistent indicator adjacent to the content — not in a settings page or footnote. For published text, include the label in the byline or header. For images and video, include both visible labelling and machine-readable metadata (C2PA or equivalent).

Deepfake disclosure placement

The disclosure must be clearly visible in or accompanying the content itself — not only in the platform's terms of service. For video, an overlay or caption at the opening frame is the minimum standard. For audio, a verbal statement at the beginning of the recording.

Emotion recognition notices

If your system processes facial expressions, tone of voice, or other behavioural indicators to infer emotional state, inform individuals before exposure — ideally through a specific notice rather than general product documentation. In employment contexts, this notice must be given to employees before the system is used in their performance evaluation context.

Common Mistakes

Relying on terms-of-service disclosure

A terms-of-service clause informing users that the product uses AI does not satisfy Article 50(1). The Act requires disclosure at the point of interaction, not in a document users accept once and never read again.

Labelling only in settings menus

Placing AI disclosure in account settings, product documentation, or an 'about this product' page does not meet the 'clearly visible' standard for content labelling. The disclosure must be proximate to the content or interaction it concerns.

Treating the obligation as applying only to consumer products

Article 50 applies to professional contexts as well as consumer contexts. B2B tools used by employees, contractors, or customers in professional settings are subject to the same obligations.

Confusing provider and deployer obligations

Article 50(1) and 50(2) impose obligations on providers. Article 50(3), 50(4), and 50(5) impose obligations on deployers. A provider building a general-purpose content generation API is responsible for machine-readable marking (50(2)); the deployer using that API to publish deepfakes is responsible for visible disclosure to viewers (50(4)). Both must act — one does not substitute for the other.

Understand your full compliance scope

Nytivo's risk classification wizard identifies which articles apply to your specific AI system — including Article 50 obligations — and generates the documentation you need to demonstrate compliance.

Start free trial
FAQ

Article 50 — Frequently Asked Questions

Does Article 50 apply to internal tools used only by employees?

Article 50 applies to AI systems used in professional contexts. For internal tools, Article 50(3) (emotion recognition / biometric categorisation) explicitly applies — deployers must inform employees that such systems are operating. For general chatbot tools used internally, Article 50(1) still applies technically, but the practical enforcement risk is lower because the context makes the AI nature more readily apparent. The safest approach is to treat all AI systems interacting with individuals — including employees — as subject to Article 50.

What does 'clearly visible' mean for disclosure requirements?

The Act does not define 'clearly visible' precisely. The standard implied by the regulation is that a reasonable person in the relevant context would encounter and understand the disclosure without having to actively seek it out. A disclosure buried in terms of service, a settings menu, or a small-print footnote would not satisfy this standard. The disclosure should be presented at or before the point where the user first encounters the AI-generated content or interacts with the AI system.

Are there technical standards for AI watermarking and content marking?

As of 2026, harmonised technical standards for Article 50(2) machine-readable marking are still under development by ETSI and CEN/CENELEC. The Coalition for Content Provenance and Authenticity (C2PA) standard provides a practical implementation framework that several major AI providers have adopted. In the absence of published harmonised standards, using a robust, widely-adopted technical approach with documented methodology is the appropriate course.

Does Article 50 apply to AI that assists humans but doesn't generate content directly?

Article 50 is specifically targeted at AI systems that directly generate outputs or directly interact with individuals. An AI system that provides recommendations to a human author who then writes content independently — where the human makes substantive editorial decisions — is less clearly in scope for 50(2) and 50(5). However, the line is not bright. If the AI system generates text that is published with minimal human modification, the deployer's Article 50(5) obligations are likely engaged.

What is the penalty for Article 50 non-compliance?

Article 50 violations fall within the broader penalty framework of Article 99. Non-compliance with transparency obligations (as opposed to prohibited AI or high-risk non-compliance) can attract fines of up to €15M or 3% of global annual turnover — the same tier as high-risk AI non-compliance. The proportionality principle applies, and first violations with remediation cooperation typically result in lower penalties, but the maximum is substantial.

Transparency obligations by industry

Article 50 applies across high-risk categories and GPAI. See how disclosure requirements interact with your product's use case.