The goal of the project presented here was to identify necessary changes to existing legal regulations such as the product safety law and the industrial safety law due to the use of AI-based algorithms and other software with autonomy features in physical systems. The overview is limited to industrial systems for which a safety assessment is required. In order to be able to answer these central questions of the research project, a system of categories (i.e. taxonomy) was developed, which summarizes safety-related factors in recently developed software-physical (AI-based) systems. The taxonomy was based on extensive expert interviews and supplemented by literature analysis.
Complex data-driven algorithms as found in machine learning, a subfield of AI, differ from conventional software. They are characterized by aspects such as understandability, predictability, ability for specification, robustness or transparency. Particularly relevant for the legal considerations is the fact that data-driven system behavior can be tailored it to its specific application by extending the data-based training process into its operational use. Systems that continue to learn - i.e. systems that use data collected in operational use to modify their behavior - are not considered in product safety law, since changes to the system after commissioning are not taken account of in the risk assessment, which is carried out in the process of commissioning (and not afterwards).
Within this project, various proposals for the enhancement of product safety law were developed. In the proposal "modification of product definition", the product definition is extended by stating that changes to the product do not create a new product as long as the changeability is intended. It is important that the functionality intended is precisely defined for the product in question. During the course of drafting this alternative legislation, the term "mutability" is defined as a combination of several dimensions of the developed taxonomy.
As a legal consequence of a "mutable product", an adapted "product support concept" is developed. According to this approach, the manufacturer is obliged to ensure safety over the entire intended product life cycle. The product support concept includes both the collection and evaluation of information during the operation of the product as well as taking appropriate measures. In contrast to conventional product monitoring, this approach is based on both observing and reacting.
Please download the complete report "Legal framework for making available autonomous and AI systems" (in German only).
T. Jürgensohn, C. Platho, D. Stegmaier, M. Hartwig, M. Krampitz, L. Funk, T. Plass, H. Ehrlich:
Rechtliche Rahmenbedingungen für die Bereitstellung autonomer und KI-Systeme.
1. edition. Dortmund: Bundesanstalt für Arbeitsschutz und Arbeitsmedizin 2021. pages 243, Project number: F 2432, PDF file, DOI: 10.21934/baua:bericht20210423
© Federal Institute for Occupational Safety and Health