The possible applications of artificial intelligence (AI) are currently being discussed very intensively in various circles. As a revision of the Machinery Directive 2006/42/EC is currently being discussed in Brussels, the inclusion of AI in Annex I of the essential health and safety requirements is also a topic here. Contrary to the original planning, the EU Commission will not publish its draft revision until the 2nd quarter of 2021 (see here).
There is no generally accepted definition for artificial intelligence. In a joint project with the German Federal Ministry for Economic Affairs and Energy (BMWi), DIN and DKE have developed a standardization roadmap for artificial intelligence in order to provide a comprehensive overview here of the status quo, requirements and challenges, as well as standardization and standardization needs for various key topics related to artificial intelligence.
An important aspect of AI is under which conditions safety-relevant decisions of a system may be influenced or automated by artificial intelligence methods. Concrete legal requirements are currently still lacking. The internal market regulations to date make no reference to the aspect of AI. AI systems would be conceivable, particularly in mechanical and plant engineering, and therefore reference will certainly be made to this in the revision of the Machinery Directive in order to have an evaluation standard for the legally required risk assessment.
Based on the hierarchy of protective measures (the so-called product liability pyramid), a product must be designed in such a way that hazards cannot arise in the first place. Where this is not possible, protective devices must reduce the risks until only acceptable residual risks remain. Finally, the users must be informed about these risks (obligation to instruct). Control systems play a significant role in this concept if they are used to execute the safety functions of a product (cf. Annex I, Section 1.2.1, 2006/42/EC). It is crucial that manufacturers are able to assess the risks posed by their products.
This is where the problem lies: designers of systems based on the more complex methods of artificial intelligence (such as machine learning with neural networks) have so far been unable to satisfactorily explain, even in retrospect, why their system behaved in a certain way.
The aim should be to use such approaches, which come more from risk management, to establish catalogs of criteria for an acceptable level of risk for artificial intelligence methods as well. These criteria catalogs can contain specifications for specification and modeling, explainability and comprehensibility of decisions, transferability to different situations, verification and validation of the system, monitoring during runtime, human-machine interaction, process assurance and certification, as well as safety-related ethics and data security.
Therefore, the European Parliament calls for a Regulation (EU) on ethical principles for the development, deployment and use of artificial intelligence, robotics and related technologies. This proposes such criteria for assessing compliance.
Initial approaches can be found in the recently published technical report ISO/TR 22100-5 3 "Safety of machinery - Relationship with ISO 12100 - Part 5: Implications of artificial intelligence machine learning" (January 2021). In addition to the revision of the Machinery Directive, the EU Commission will also present a proposal for a regulation on artificial intelligence that contains legally binding framework conditions for the use of artificial intelligence. These framework conditions must contain complete, clear and verifiable requirements for the cases in which and the conditions under which safety-relevant decisions of a system may be influenced or automated by artificial intelligence methods.