TÜV test for AI applications – com! professional

In the “TÜV AI Lab”, AI experts from TÜV organizations will work on practical test scenarios. For example, the AI ​​inspectors could determine how safely automated vehicles with AI systems recognize people, traffic signs or certain obstacles and react to them. Both performance and aspects of cybersecurity and the robustness of AI play an important role. A corresponding test should be a prerequisite for the approval of new vehicles in the future. “Artificial intelligence is complex and can behave dynamically,” said Dr. Dirk Schlesinger, head of the TÜV AI Lab. “So we cannot simply transfer existing test approaches, we have to develop new test methods.” The data with which algorithms are trained are also taken into account. For example, the training data of an AI system for personnel selection can lead to a disadvantage for certain groups of people if the data is not balanced. Schlesinger: “The TÜV AI Lab will develop criteria for assessing the suitability of training data for certain AI applications.”

Another goal of the TÜV AI Lab is to develop solutions for classifying AI applications into risk classes. “Not all AI systems have to meet the same requirements,” said Schlesinger. Therefore, AI-supported systems should be regulated according to their risk potential. The requirements could range from a complete waiver of regulation, transparency obligations and approval procedures before the market launch to a ban on particularly dangerous or ethically questionable AI applications. The EU Commission has proposed two risk classes (low and high risk), while the Federal Government’s Data Ethics Commission advocates five levels. Schlesinger: “The question is how many risk classes actually make sense and according to which criteria AI systems can be assigned to the individual classes in practice.” The TÜV AI Lab, which will start work in the first quarter of 2020, will also deal with this.