Paper accepted in AI & Society journal

14 April 2026

Excited to share our new paper: “Not a statistical accident: framing bias as a semiotic property of image datasets,” now published in AI & Society.

This work is led by PhD researcher Simone Fabrizi, in collaboration with Symeon Papadopoulos and Yiannis Kompatsiaris from CERTH, Greece.

In this paper, we argue that image datasets are not neutral samples of reality but semiotic systems that frame concepts through recurring visual associations, making framing bias an inherent property of datasets.

 

Key takeaways:

  • Bias is not just statistical. Bias is often attributed to sampling errors, imbalance, or annotation issues. We argue that bias is structural and interpretive: datasets function as meaning-making systems that inevitably encode perspectives.
  • Meaning emerges through co-occurrence. Concepts are framed through recurring visual associations (e.g., women + kitchens, Africa + poverty) which create dataset-level narratives that models can learn and reproduce.
  • Bias also appears through omission. In our analysis of Visual Genome, leisure activities are heavily represented while everyday labour is largely absent. Despite aiming to represent the “visual world”, the dataset portrays only a narrow slice of reality.
  • Framing works in both directions. Individual images contribute to dataset-level narratives, while the dataset context shapes how individual images are interpreted.

 

Implications: Instead of asking only: "Is my dataset balanced?" we should also ask:

  • What worldview does my dataset encode?
  • What stories does my dataset tell?
  • Which contexts dominate, and which ones are missing?

 

No dataset can be completely unbiased. What we need instead is better bias understanding, dataset documentation, and context-aware deployment.

 

Aktuelles

MMM-Fair presented in the Demo Track at CIKM 2025

“MMM-Fair: An Interactive Toolkit for Exploring and Operationalizing Multi-Fairness Tradeoffs” was presented in the Demo Track at CIKM 2025, held in Seoul, Republic of Korea.

Kick-off Workshop of the DZdA Project – German Center for Digital Tasks in Higher Education Teaching

The kick-off workshop of the project German Center for Digital Tasks in Higher Education Teaching (DZdA) took place as a two-day intensive event with all nine project partners from universities across Germany. The host was OTH Amberg-Weiden, which, together with the strategic consulting firm pictomind GmbH, facilitated the event. The goal of the workshop was to establish a shared understanding of the project, define strategic objectives, and lay the foundation for successful collaboration.

Fair & Bias-Aware AI: MAMMOth Concludes

The Horizon Europe-funded project MAMMOth (Multi-Attribute, Multimodal Bias Mitigation in AI Systems) officially concluded on 31 October 2025 after three years of pioneering work to make Artificial Intelligence fairer, more inclusive, and more accountable. The project brought together academic, industrial, and societal partners to address bias in AI systems and to deliver practical tools and knowledge for fairness-aware AI development. Its outcomes provide policymakers, technology developers, and society with concrete strategies for embedding fairness at the centre of AI innovation.

Paper accepted: IEEE Big Data 2025

Our Paper "A Deep Latent Factor Graph Clustering with Fairness-Utility Trade-off Perspective" was accepted at IEEE Big Data 2025. Congratz, Siamak!

The Multifaced Nature of Bias in AI: Impact on Model Generalization, Robustness, and Fairness

At last week’s EU AI Fairness Cluster meeting in Brussels, Eirini presented insights on the different forms of bias in AI in a talk titled "The Multifaced Nature of Bias in AI: Impact on Model Generalization, Robustness, and Fairness".

EU AI Fairness Cluster Brussels

Eirini participated in the final event of the EU AI Fairness Cluster in Brussels. It was a great opportunity to explore different aspects of fairness in AI.