Trustworthy AI I
Sat 17 Oct · 11:00–12:30 · Room A (First floor)
Paper session · Trustworthy and explainable AI · Chair: Mateo Silva
Papers
Sign in to ask or upvote questions. Sign in
- Full paperPresented
Explanations that clinicians use: a field study of feature attributions in triage
Theo Laurent (Meridian Research Centre) (presenting), Iris Andreou (Thalassa University)
Abstract
Feature attribution methods are often evaluated with proxy metrics rather than with the people who rely on them. Over eight weeks, 23 triage nurses used a sepsis risk tool with and without SHAP explanations. Explanations did not change accuracy overall but halved the time to override wrong alerts. Nurses asked for explanations in terms of recent changes rather than absolute values, which we turn into design guidance.
explainability · healthcare · human factors · SHAP
Questions from the audience 0
No questions yet.
- Full paperPresented
Auditing hiring classifiers for intersectional bias with limited demographic data
Viktor Ivanov (Meridian Research Centre)
Abstract
Bias audits need demographic labels that companies often do not hold. We propose an audit protocol that combines a small consented sample with probabilistic inference of group membership and reports confidence intervals rather than point estimates. Applied to three anonymised hiring models, the protocol detects a 12-point selection gap for one intersectional group that a single-attribute audit misses.
fairness · bias audit · hiring · intersectionality
Questions from the audience 0
No questions yet.
- Short paperNot presented
Mapping the EU AI Act's human oversight requirement to concrete interface patterns
Irene Galanis (University of Eastvale) (presenting), Pablo Moreno (Polytechnic of the North)
Abstract
Article 14 of the EU AI Act asks for effective human oversight of high-risk systems but does not say what that looks like. We collect 31 interface patterns from deployed decision-support tools and map each to the oversight capabilities listed in the Act, identifying four capabilities with no established pattern.
AI Act · human oversight · interface design · regulation
Questions from the audience 0
No questions yet.