Seminars
Previous Seminars
- Abstract:
Trustworthy AI has moved from a statement of principles to a condition of doing and teaching AI research in Europe. Ethics appraisal, risk classification and documentation requirements now shape how projects are proposed, reviewed and funded, and increasingly, what counts as a finished result. This talk asks what the shift actually demands of researchers and lecturers and whether trustworthy AI can function as a precondition at all, when the tensions it concerns only become visible in use. We begin by defining trustworthy AI as the European framework uses the term, and set out its concrete requirements. We then turn to the experience of applying the Z-Inspection process in EU projects to real AI systems in healthcare and other high-stakes settings. Z-Inspection is a holistic, interdisciplinary assessment process built around socio-technical scenarios. Perhaps surprisingly, what it surfaces is rarely a violated principle but an ethical tension, such as between accuracy and explainability, or between clinical usefulness and data minimisation that no checklist resolves. We close with what this means for research practice, and for what we should be teaching our students.Speaker: Magnus Westerlund, Arcada (Finland)
Date: 04.11.2026
- Abstract:
The seminar will concentrate on a critical discussion of two publications that address AI technologies in the media-epistemic sense of the Kristiania AI Evenings lecture (as delivered by Wolfgang Ernst). A close reading (at least of the "Abstracts" ...) of the following texts is therefore required:
a) Laura Tripaldi, Softness: An Ecological Paradigm for Embodied Technological Intelligence, in: Social Epistemology. A Journal of Knowledge, Culture and Policy (published online 13 January 2025); https://www.tandfonline.com/doi/full/10.1080/02691728.2025.2449607
b) David M. Berry, Vector Theory: Epistemology, Political Economy, and Probabilistic Computation, in: Philosophy & Technology, vol. 39 (2026), article no. 149; https://doi.org/10.1007/s13347-026-01162-w
Both text discussions will not be summarized but initiated by short impulses from Ernst's side:
a) "Smart Matter vs. Technological Intelligence?"
b) "Asemic: The Subsymbolic Approach"
Further recommended reading: Leonardo Impett / Fabian Offert, Vector Media, Harrogate (MNG University Press) 2026
Speaker: Wolfgang Ernst, Humboldt University (Germany)Date: 14.10.2026
- Abstract:
This seminar examines AI swarms as social systems, connecting our early research on information diXusion with recent work on synthetic consensus, cyborg propaganda, and coordinated influence. I present IO Factory as a simulation framework for studying influence campaigns, then outline current research on narrative amplification, tipping points, adaptive coordination, and collective safety. The central question is what coordination adds beyond individual agent capabilities, and which findings from artificial populations can transfer to human settings.
Speaker: Daniel Schroeder, SINTEF
Date: 30.08.2026 - Abstract:
Financial-distress prediction support lending, screening and early intervention, but their usefulness depends on the event predicted and the decision they inform. This design-oriented systematic review synthesizes 58 primary studies selected through citation searching and a database search. It examines how studies define distress, construct and explain estimates, and evaluate predictive performance and decision usefulness. Three findings emerge. First, legal failure, regulatory designation, payment default and accounting-based deterioration represent diVerent monitoring tasks, with diVerent horizons and opportunities for response. Second, richer data and modelling approaches provide several distinct capabilities such as incorporating external conditions, modelling firm-specific sensitivity and explaining a prediction should not be treated as equivalent. Third, reporting is more complete for conventional validation than for several aspects of use. 50 studies report out-of-sample or cross-validation evidence, eight quantify economic cost, loss or usefulness, and seven assess calibration.Studies of lending loss, profitability and bank based testing provide concrete decision-oriented contributions, although they do not establish every aspect of operational use. The synthesis proposes eight purpose-dependent design requirements and corresponding evaluations. These are propositions for future system development, rather than validated requirements for everydistress model.
Speaker: Tayyab Ahmad
Date: 19.08.2026 - Abstract:
Eye-tracking is becoming cheap enough, in VR headsets and ordinary cameras, to work as an early screening tool for conditions such as autism and ADHD. But eye-movement research is fragmented: physics describes gaze as one random process governed by a single scaling parameter; clinical practice assumes two phases, fixations and saccades; machine learning assumes nothing and predicts well, but resists interpretation. In this talk we bring these into contact. Simple models with a few discrete states reproduce the heavy tails and extreme events of gaze better than the generative networks we tested, and a process that explicitly alternates between phases fits individual trajectories better than a Lévy walk: the clinical two-phase picture has a mathematical basis it previously lacked. Paired with standard classifiers, these models detect ADHD-related gaze signatures efficiently. We then extend the same treatment past the binary. Microtremors, faint rhythmic movements during fixations, relate to individual limits in temporal perception; and saccades separate into distinct subtypes rather than forming one class. Both improve screening for neurodevelopmental conditions, and both suggest gaze needs a more granular taxonomy than two states allow.
Speaker: Pedro Lencastre, OsloMet
Date: 19.08.2026 - Abstract:
We introduce FærdXel, an expert system for reasoning within the domain of Danish traffic law. FærdXel can analyse concrete legal cases and explain its conclusions by presenting a chain of reasoning based on concrete legislation, case law, and the specifics of each case.
While the core system can only be developed and used by trained experts, we are in the process of integrating generative AI following human-in-the-loop principles to ensure correctness. The goal is that this combination will result in a system that is accessible to all, and can correctly update itself as legislation changes.Speaker: Luís Cruz-Filipe (IT University, Denmark)
Bio: Luís Cruz-Filipe is a researcher fascinated by the potential in the interaction between humans and computers in the area of mathematical proofs. His research over the last twenty years has focused on understanding how humans reason in specific domains, and how to develop computer systems that can "reason" in similar ways. He is currently Head of Education at the IT University in Copenhagen.
Date: 26.05.2026