Introduction: With the growing prevalence of AI-based systems and the development of specific regulations and standardizations in response, accountability for consequences resulting from the development or use of these technologies becomes increasingly important. However, concrete strategies and approaches of solving related challenges seem to not have been suitably developed for or communicated with AI practitioners.Methods: Studying how risk governance methods can be (re)used to administer AI accountability, we aim at contributing to closing this gap. We chose an exploratory workshop-based methodology to investigate current challenges for accountability and risk management approaches raised by AI practitioners from academia and industry.Results and Discussion: Our interactive study design revealed various insights on which aspects do or do not work for handling risks of AI in practice. From the gathered perspectives, we derived 5 required characteristics for AI risk management methodologies (balance, extendability, representation, transparency and long-term orientation) and determined demands for clarification and action (e.g., for the definition of risk and accountabilities or standardization of risk governance and management) in the effort to move AI accountability from a conceptual stage to industry practice.

The current endeavor of moving AI ethics from theory to practice can frequently be observed in academia and industry and indicates a major achievement in the theoretical understanding of responsible AI. Its practical application, however, currently poses challenges, as mechanisms for translating the proposed principles into easily feasible actions are often considered unclear and not ready for practice. In particular, a lack of uniform, standardized approaches that are aligned with regulatory provisions is often highlighted by practitioners as a major drawback to the practical realization of AI governance. To address these challenges, we propose a stronger shift in focus from solely the trustworthiness of AI products to the perceived trustworthiness of the development process by introducing a concept for a trustworthy development process for AI systems. We derive this process from a semi-systematic literature analysis of common AI governance documents to identify the most prominent measures for operationalizing responsible AI and compare them to implications for AI providers from EU-centered regulatory frameworks. Assessing the resulting process along derived characteristics of trustworthy processes shows that, while clarity is often mentioned as a major drawback, and many AI providers tend to wait for finalized regulations before reacting, the summarized landscape of proposed AI governance mechanisms can already cover many of the binding and non-binding demands circulating similar activities to address fundamental risks. Furthermore, while many factors of procedural trustworthiness are already fulfilled, limitations are seen particularly due to the vagueness of currently proposed measures, calling for a detailing of measures based on use cases and the system’s context.

Despite the popularity of density-based clustering, its procedural definition makes it difficult to analyze compared to clustering methods that minimize a loss function. In this paper, we reformulate DBSCAN through a clean objective function by introducing the density-connectivity distance (dc-dist), which captures the essence of density-based clusters by endowing the minimax distance with the concept of density. This novel ultrametric allows us to show that DBSCAN, k-center, and spectral clustering are equivalent in the space given by the dc-dist, despite these algorithms being perceived as fundamentally different in their respective literatures. We also verify that finding the pairwise dc-dists gives DBSCAN clusterings across all epsilon-values, simplifying the problem of parameterizing density-based clustering. We conclude by thoroughly analyzing density-connectivity and its properties -- a task that has been elusive thus far in the literature due to the lack of formal tools. Our code recreates every experiment below: https://github.com/Andrew-Draganov/dc_dist

Machine learning algorithms are playing an increasingly important role in public health measures, accelerated by the Covid-19 pandemic. It is therefore vital that machine learning algorithms are applied in ways that are generally considered fair. However, the question of how to define fairness in a public health context is still an open one. In this study, we investigated people’s attitudes towards two ways of defining fairness in the context of Covid-19 contact tracing apps. In the first‘, high-individuality’ approach, the likelihood of an algorithm asking a person to self-isolate would depend on the person’s individual characteristics, such as their risk of spreading the virus through regular contacts. In the second ‘low individuality’ approach, these individual characteristics would not be used to come to a decision. For each approach, participants rated its fairness, overall quality, and their privacy concerns, and answered questions about basic psychological need satisfaction. Participants rated the high-individuality approach as fairer and better overall compared to the low-individuality approach, despite having greater privacy concerns. Further, we found a strong correlation between the participants’ fairness perceptions and their overall impression of the tracking tool. Together, these findings suggest that people prefer individualised approaches in some contexts and perceive them as fairer. However, policy makers should consider the privacy trade-off of employing such measures.