Keuper Labs
← All publications

exCV Workshop at ECCV 26 · 2026

Position: We need theory-grounded explainability for abstract classification to quantify model bias

Katharina Prasse, Margret Keuper

Explainable AIModel BiasTheory-Grounded Evaluation

Why this publication matters

People do not all mean the same thing when they call an image wealthy or affluent. This paper shows that models overlook some human wealth cues and that human ratings of generated scenes vary across countries. It argues for defining the social context and using validated measures, so that explanations can support a meaningful assessment of bias.

Abstract

Explainability for denotative and connotative tasks differs fundamentally. While denotative features (e.g. of a bird species or of a melanoma) can be universally agreed upon, representations of connotative classes (e.g. wealth of a person or affluence) are subject to cultural, geographic location, and personal beliefs. Because images offer a richer contextual signal compared to text, they may amplify model bias. We advocate for explainability in abstract classification tasks on the basis of two case studies: (1) Comparing seven foundation models in their perception of personal wealth against a psychologically validated survey on wealth cues shows that models collectively omit the character dimensions of the taxonomy. (2) Human annotators reach satisfactory levels of agreement (Krippendorff’s α ≥0.67) on affluence in generated images in 3 / 14 assessed countries. Current XAI methods lack a validated ontology to evaluate abstract classes against and thus seemingly plausible explanations may be inadequate for the given task. We argue that abstract classification requires an explicit declaration of context - population, geographic location, cultural norm, time frame - and an externally validated measure suitable for the specific context. This renders classification bias quantifiable and models steerable.

Abstract source ↗

Figures

A heatmap compares wealth cues used by psychologists and seven foundation models.
Figure 1. Psychologists [25] and foundation models show different foci in wealth cues. While all observations contain possessions as an indicator, we observe divergence in high-spending willingness, character, and self-presentation. View in source ↗
Human affluence ratings disagree differently for generated images of Indonesia and the United Arab Emirates.
Figure 2. GeoDiv country-specific human annotation statistics for affluence in generated images show divergence between annotators, especially for UAE, where annotator 2’s predictions are more centered around the middle score, whereas annotators 1 and 3 reported right-skewed affluence levels. [Note: y-axis scales differ between (a) and (b).] View in source ↗

Cite this paper

Download .bib 
@inproceedings{prasse2026weneedtheory95,
  title = {{Position: We need theory-grounded explainability for abstract classification to quantify model bias}},
  author = {Katharina Prasse and Margret Keuper},
  booktitle = {exCV Workshop at ECCV 26},
  year = {2026},
  url = {https://openreview.net/forum?id=Oz2id4l0SX}
}

Figures and abstract are reproduced from the linked research sources. Credit remains with the authors and publishers.