On November 26, 2024, a Diner Pensant was held in cooperation with the AI Hub South Holland, a networking event that this time revolved around a substantive discussion on the AI regulation. With an interesting company consisting of guests from the business community, from corporates to SMEs, from government agencies and interest groups, and virtually all scientific institutions in the region.
Foundation for reliable AI systems
Initiated by Vandana Dwarka and Marieke Kootte, both associate professors of mathematics at TU Delft, this evening focused on the importance of mathematical knowledge as a foundation for reliable AI systems. The AI regulation, which sets strict requirements for the testing and robustness of these systems, underscores the urgency of this topic.
In their pitch prior to the discussion, Dwarka and Kootte used a case to show how problems such as discrimination and bias could have been avoided here by correctly applying mathematical tests. They emphasized that mathematical principles should be better integrated into existing practices, such as impact assessments and ethical guidelines. Kootte and Dwarka warn that "without a solid mathematical foundation, compliance with AI regulation becomes a guessing game." As a possible solution, they suggest post-graduate professional training, ideally suited to combine technical knowledge and practical experience.
These insights sparked a wide-ranging discussion, spurred by moderator Tom Jessen, in which participants from a variety of disciplines reflected on the key question: how can we bring mathematical knowledge and practice together in a way that meets the requirements of the AI Act?
AI regulation compliance
A representative from a large Dutch corporate indicated that they are reasonably comfortable with how they are organized to be compliant with the AI act. He noted that there is sufficient expertise in the teams of corporates. In SMBs, the challenge is greater, because although they are likely to work less with high-risk AI systems, the need for AI compliance is no less urgent. There is a lot of potential in making the AI law tangible for smaller companies. One suggestion was that customized, off-the-shelf tools can bridge this gap. Developing certified "AI law-compliant programs" that SMBs can use can lower the barrier for SMBs and give them confidence in their AI systems.
Interdisciplinary collaboration
An associate professor of engineering computer science emphasizes breaking disciplinary silos. "Even within our own research field, there are different assumptions about what constitutes good data. Interdisciplinary teams ensure that we cross-check these assumptions so that algorithms are fair and robust." An AI consultant with the central government uses a recent example to illustrate how this can have major consequences even in seemingly simple algorithms. A visa application system disproportionately flagged certain nationalities, despite rigorous testing. Unintended, but "this example underscores the need for broader oversight and interdisciplinary teams to recognize biases early."
An associate professor of Responsible AI and digital ethics knows from experience that this approach works: a project at a medical center linked philosophical and ethical considerations directly to the design of mathematical models. Together with clinicians, sociologists, philosophers, and others, they evaluated the impact of early discharge algorithms on patients. This kind of collaboration is essential to understanding the broader implications of AI.
Reliability is crucial
If AI is to make meaningful impact, reliability is crucial. This starts with transparency and extends to the broader ecosystem of developers, users and regulators. Sabine Herbrink, coalition secretary NL AIC, states, "Our role as an umbrella organization is to facilitate these discussions and ensure that all stakeholders have a voice."
The road to AI regulation compliance may be complex, but it is also an opportunity to build better more responsible AI systems. By investing in education, fostering collaboration and developing practical tools, organizations can not only comply with regulatory requirements but also lead the way in ethical AI innovation. At the end of the discussion, the intention was expressed to share this insight with regulators as well.
Collaboration for the future
This dinner pensant is the third in a series of meetings. It was organized by the AI Hub South Holland and the Dutch AI Coalition. This meeting highlights the need for collaboration and innovation in the AI domain and the importance of a collaborative approach to reap the benefits of AI and meet the associated challenges.
