EU AI Act: Closing regulatory gaps quickly

TÜV AUSTRIA presents requirements and proposed solutions for the assessment of AI systems, recommends a corresponding clarification for the EU AI Act and published a position paper.

After years of negotiations, Europe’s new AI law, the EU AI Act, was recently adopted by the European Parliament. The final text will not be available until 2024, with the new law to be fully applied from 2026.

The foundations for future conformity assessments with regard to the functional requirements for these systems – such as robustness, transparency and reliability – are not sufficiently taken into account in the standardization bodies and in the EU AI law, Doms emphasizes.

There is also a lack of concrete specifications for uniform risk assessment procedures and possible opt-out procedures for stand-alone systems. This poses the risk of unequal treatment of similar products.

Doms calls for so-called low-profile certifications or purely paper and document-based reviews of risky AI applications to be avoided. “Our experience from certification projects that have already been carried out shows that the frequently expressed fear of some stakeholders that testing procedures with sufficient testing depth would lead to overregulation is a bogus argument. After all, even tests for complex models and applications have never exceeded an economically justifiable scope when taking functional safety requirements into account.”

Spot checks instead of paper and document-based checks

Thomas Doms: “The complexity, as we know it from deep neural networks, for example, also makes it difficult to detect and rectify malfunctions or manipulations from outside.

In order to ensure the safe, robust and transparent use of AI, especially in safety-critical applications, sufficient functional requirements must be specified and appropriate concepts, methods and tools for corresponding tests are required.”

The quality of a trustworthy AI decision system can be determined above all by the correct statistical testing of samples and the precision of the definition of the field of application, which makes it possible to draw representative samples in the first place. Thomas Doms: “We therefore demand that a reliable assessment of the

statistical functional properties of an AI system according to the known scientific state of the art must form the indispensable and binding core of a conformity assessment.”

This requires

 “The TÜV AUSTRIA TRUSTED AI test method can be applied to both supervised and unsupervised learning applications. We are happy to share our criteria, experience and evaluation results with the EU Commission and its standardization bodies. So that existing regulatory gaps in EU law can be closed quickly,” concludes Thomas Doms.

Download the TÜV AUSTRIA position paper

Podcast with Thomas Doms (Managing Director TRUSTIFAI) and Bernhard Nessler (Research Manager Deep Learning & Certification, SCCH)