Raumanns et al. (2026) — Fontys University of Applied Sciences & IT University of Copenhagen. arXiv preprint (cs.CY).
Machine learning courses typically hand students pre-labeled datasets, hiding the subjectivity baked into human annotation and cultivating an overly trusting view of AI data pipelines. This two-university study (Fontys, Netherlands and IT University Copenhagen; N=43) had students annotate skin-lesion images for hair coverage on a 3-point scale, then surveyed their understanding of annotation ambiguity, data quality, bias, and fairness. Self-reported familiarity with these concepts rose substantially across the board, and most students came to recognise that personal interpretation shapes labels \u2014 positioning hands-on annotation as a lightweight, transferable exercise for building ai-literacy and the kind of data skepticism central to critical-thinking-genai-scaffolding.\n\nThe pedagogical claim is that interpretive diversity in labeling is itself a teachable object: rather than treating disagreement as noise to be resolved, instructors can surface it to teach bias-mitigation and fairness reasoning within cs-education. This complements wiki threads on student-ai-interaction and epistemic vigilance by targeting the data layer \u2014 students who have personally wrestled with ambiguous labels are less likely to treat model outputs as ground truth.
Related Pages
- ai-literacy — related wiki thread
- critical-thinking-genai-scaffolding — related wiki thread
- cs-education — related wiki thread
- bias-mitigation — related wiki thread
- student-ai-interaction — related wiki thread