Accessibility Obligations in the AI Act

AI Act: Accessibility Obligations for AI Systems

The AI Act establishes a harmonized legal framework for the development, marketing, and use of artificial intelligence systems within the European Union. A significant aspect of this regulation is the emphasis on the accessibility of information and interfaces associated with AI systems.

Key Situations for Accessibility Obligations

These obligations apply primarily in two scenarios:

When AI systems are subject to transparency requirements for users;
When AI systems are classified as high-risk.

In these cases, the information provided to users or operators must be accessible, including to individuals with disabilities.

Existing Accessibility Frameworks

While the regulation does not define specific technical standards to ensure accessibility, it refers to the existing European framework on digital accessibility, mainly composed of:

Directive (EU) 2016/2102 regarding accessibility of public sector websites and mobile applications;
Directive (EU) 2019/882, known as the European Accessibility Act.

These directives are based on European and international technical standards, including the EN 301 549 standard for accessibility requirements for ICT products and services, and the Web Content Accessibility Guidelines (WCAG).

European Accessibility Directives

Directive (EU) 2016/2102

This directive aims to harmonize accessibility requirements for digital services provided by public sector bodies. It applies to:

Textual and non-textual content;
Downloadable documents;
Online forms and services;
Multimedia content.

Article 4 mandates that member states ensure public sector websites and mobile applications are designed to be perceptible, usable, understandable, and robust.

Directive (EU) 2019/882

This directive extends accessibility requirements to a range of digital products and services in the private sector, including:

Consumer terminal equipment with interactive computing capabilities;
Self-service terminals;
Digital e-readers;
Equipment facilitating access to electronic communication services.

Annex I stipulates that products and services must be designed to ensure optimal use by persons with disabilities, emphasizing accessibility principles.

Technical Accessibility Standards

EN 301 549

The EN 301 549 standard defines technical requirements to verify the accessibility of Information and Communication Technology (ICT). It encompasses websites, mobile applications, software, electronic documents, and user interfaces.

Compliance with this standard typically demonstrates adherence to the accessibility requirements set by European directives.

Web Content Accessibility Guidelines (WCAG)

The WCAG, established by the World Wide Web Consortium (W3C), provides international guidelines for making digital content accessible. These guidelines are based on four fundamental principles, often summarized by the acronym POUR:

Perceptible: Information must be presented in ways that users can perceive, such as providing text alternatives for images.
Operable: Interfaces must be usable through multiple interaction modes, ensuring compatibility with assistive technologies.
Understandable: Information and interfaces should be easy to comprehend, with clear language and predictable behavior.
Robust: Content must be robust enough to be reliably interpreted by various browsers and assistive technologies.

WCAG establishes testable success criteria organized into three levels of compliance: A, AA, and AAA, with level AA being the standard typically targeted by European regulations.

Focus on AI: Accessibility Obligations in the AI Act

The AI Act incorporates accessibility as a transversal requirement applicable to certain AI systems, focusing primarily on:

High-risk AI systems;
Systems subject to transparency obligations.

Accessibility for High-Risk AI Systems

Article 16 outlines the obligations of providers of high-risk AI systems. It mandates that:

Providers ensure compliance with accessibility requirements as per directives (EU) 2016/2102 and (EU) 2019/882;
Accessibility must be integrated from the design phase of the AI system;
User interfaces and interactions with the system must be accessible;
Compatibility with assistive technologies must be ensured.

This approach is reinforced by Recital 80, which emphasizes that accessibility must be integrated from the design phase of the system, promoting the concept of accessibility by design.

Accessibility of Transparency Obligations

Article 50 of the AI Act imposes transparency obligations for certain AI systems, particularly those that:

Interact with individuals (e.g., chatbots);
Generate or manipulate content (text, images, audio, or video);
Employ emotion recognition or biometric categorization;
Create deepfakes.

Paragraph 5 of Article 50 stipulates that:

Information provided to individuals must be clear and recognizable upon first interaction, adhering to accessibility requirements;
Notifications of transparency must be accessible to individuals with disabilities and users of assistive technologies.

In practice, this means that information must be:

Perceptible (e.g., text alternatives, subtitles, sufficient visual contrast);
Operable (accessible navigation compatible with assistive technologies);
Understandable (clear language, coherent interface);
Robust (interpreted accurately by assistive technologies).

Accessibility: An AI Governance Challenge

With the AI Act, accessibility becomes a compliance requirement for artificial intelligence systems. Organizations developing or deploying AI systems must now incorporate these obligations into their regulatory compliance processes, which include:

Designing accessible interfaces;
Ensuring transparency notifications are accessible;
Documenting system compliance.

In essence, accessibility is no longer just about design or user experience; it has become a governance issue in AI.

Implementing Accessibility in the Long Term

Integrating accessibility goes beyond one-time adjustments; it requires the ability to identify, track, and demonstrate compliance throughout the lifecycle of AI systems. This involves: