Abstract

This document defines overarching categories (e.g., data, algorithm, process, actors) and establishes criteria (e.g., data origin, data quantity, data labeling, population consistency, disjunction of training/validation/test data) for a unified understanding of transparency requirements concerning AI systems. This includes the following dimensions to be considered when setting transparency requirements: - Phases of the development cycle during which the transparency requirements can be implemented. - Types and use cases of AI systems for which these transparency requirements are relevant and practically applicable. - Types and detail levels of information to be provided for transparency purposes. - Actors to whom information must be provided for transparency reasons - including the question of which information is relevant for each. - Actors responsible for providing the information. - Actors about whom information must be provided for transparency reasons (e.g., qualifications, procedures and software toolsets used, all in compliance with the EU GDPR). - Types of disclosure (e.g., through user manuals or targeted online queries). - Combinations of the above aspects, e.g., in terms of which information should be disclosed to which actors, at which phase of the development cycle, and in what manner. The work should consider transparency in the sense of information provision, traceability, and explainability, among other things. The planned DIN SPEC is based on the definition of the term AI system from the EU AI Regulation ("AI Act," Art. 3.1), which is based on an OECD definition: "[An AI system is] a machine-based system that is designed to operate with varying degrees of autonomy, and that can learn and adapt post-deployment, and that generates outputs such as predictions, recommendations, or decisions influencing physical or virtual environments towards achieving specific or derived objectives." This definition describes the scope of AI systems addressed in the planned DIN SPEC, which includes data-driven AI, knowledge-based AI, generative AI, as well as hybrid combinations of these approaches (hybrid systems). It also considers both application-specific (narrow AI) and general-purpose AI systems with a broad application scope (general purpose AI). This document targets all actors involved in the development or use of AI systems, including but not limited to: - Developers, - Regulators, - Operators, - Users.

Begin

2024-12-04

Planned document number

DIN SPEC 91528

Project number

62022051