Newly published: DIN SPEC 92001-1 on the quality of AI modules

Just in time for the Hanover Fair, DIN presented the new DIN SPEC 92001-1, Artificial Intelligence – Life Cycle Processes and Quality Requirements – Part 1: Quality Metamodel. This DIN SPEC aims to ensure the quality of AI throughout its entire life cycle using a uniform concept. To this end, it defines a quality metamodel that comprises and combines all key aspects of AI quality. Special focus is placed on the individual phases in the life cycle of AI modules; these can have certain quality requirements assigned to them. Particularly in the conception and development stages, for example, it is important to prevent bias when information is processed. In addition, the metamodel identifies the three quality pillars that are the goals central to quality assurance: functionality and performance, robustness and comprehensibility. It also establishes the components of an AI module that are relevant to its quality, such as environment or data.

Universally valid model

“In contrast to conventional software, it is difficult if not impossible to evaluate assess the decision-making rules by analysing the source text of the AI algorithm because AI reactions are based on highly complex mathematical models,” according to Stephan Hinze, Managing Director of neurocat GmbH and the initiator of DIN SPEC 92001-1. “This can make the quality management of AI challenging – an aspect that developers and users must bear in mind for the entire life cycle.”  To make this easier, DIN SPEC 92001-1 provides a structural basis to ensure the quality of KI in all imaginable application situations. “The metamodel we developed is universally valid. It is not limited to specific applications of AI, but rather covers the topic of quality as broadly as possible,” says Stephan Hinze. In doing so, DIN SPEC 92001-1 distinguishes between high-risk and low-risk AI modules – depending on whether the application involves relevant safety, data protection or ethical issues.  In the case of high-risk modules, any deviations from defined quality requirements are not permissible or must be justified accordingly. In AI applications this could be the case, for example, if human life were at risk. The second part of DIN SPEC 92001-1, Artificial Intelligence – Life Cycle Processes and Quality Requirements – Part 2: Quality Requirements defines specific quality requirements for AI.

Artificial intelligence (AI) is based on neural networks and IT architectures that emulate those of the human brain. Instead of storing information unquestioningly like conventional storage elements, these networks and architectures process it autonomously and react to it. AI is one of the most significant topics of the future, impacting not only on industry and business but other areas too, including our private lives. Whatever the area, the quality aspect plays a central role. DIN SPEC 92001-1 was developed within a period of 8 months by Acsioma GmbH, DFKI GmbH, Ernst & Young AG, EMEIA-GSA Automation, Fraunhofer Institute for Open Communication Systems (FOKUS), Fraunhofer Institute for Molecular Biology and Applied Ecology (IME), GESTALT Robotics GmbH, Hamburg University of Applied Sciences (HAW), HTW University of Applied Sciences in Berlin (HTW) FB04 School of Business and Economics, Micropsi industries GmbH, Microsoft Germany GmbH, neurocat GmbH, Otto von Guericke University Magdeburg, Institute III: Philosophy, Robert Bosch GmbH, Stiftung neue Verantwortung e. V., STILL GmbH, TÜV Süd Auto Service GmbH, the University of Osnabrück and the University of Tübingen.  DIN SPEC 92001-1 has been published in English and is available for download free of charge at www.beuth.de.

In November 2018 the German Federal Government released its “Federal Strategy on Artificial Intelligence”. One of the twelve action points is “To set standards”. This demonstrates how important standardization is for this innovative topic of the future.