Introduction
As Artificial Intelligence (AI) systems become deeply embedded in societal infrastructure, the biases they inherit and perpetuate pose significant threats to individual rights and equity. While the field of "Fair-AI" has expanded rapidly, much of the existing research adopts a "hegemonic view," treating fairness as a purely technical optimization problem solvable through mathematical metrics. This narrow focus often neglects the socio-technical roots of bias and the complex legal requirements of frameworks like the GDPR and the EU AI Act.
This paper synthesizes findings from the four-year European NoBIAS project to provide a multidisciplinary roadmap. It argues for a "techno-legal" approach that bridges the gap between technical methods, historical context, and evolving policy requirements.
The NoBIAS Framework
The authors organize their findings through the NoBIAS architecture, a structured framework designed to manage bias across the entire AI lifecycle. This architecture consists of two primary components: the Bias Management Layer and the Legal Layer.
Bias Management Layer
This layer is subdivided into three critical stages:
- Understanding Bias: Investigating socio-technical entanglements, such as how historical inequalities are encoded into data, and ensuring representational adequacy.
- Mitigating Bias: Implementing technical interventions at both the data level (e.g., enrichment) and the algorithmic level (e.g., classification and ranking methods).
- Accounting for Bias: Establishing mechanisms for long-term monitoring and providing explainability (XAI) to make "black-box" decisions transparent.
Legal Layer and Techno-Legal Synergies
The Legal Layer acts as an orthogonal component, providing the necessary EU legal grounds to ensure technical solutions align with anti-discrimination and data protection laws.
A central challenge identified is the Causal Knowledge Gap. Current AI models often lack the "auxiliary causal knowledge" required by law to prove discrimination. Legally, one must demonstrate that a decision was made because of a protected attribute; most current models cannot satisfy this causal requirement.
Furthermore, there is an inherent friction between data protection law (which encourages data minimization) and non-discrimination law (which requires access to sensitive attributes to detect bias). Resolving this requires an integrated techno-legal framework where transparency and accountability serve as primary tools to overcome the limitations of quantitative metric optimization.
Key Insights: The "Ground-Truth" Myth
A provocative insight from the project is the deconstruction of the "ground-truth." In structurally unequal societies, the data collected as "truth" is often a reflection of historical injustice. The authors suggest that AI data curation should adopt "source criticism" and archival practices from the humanities to better understand the context and limitations of training data. Bias is not merely a data problem; it permeates the entire pipeline, including data annotation processes and the potential exploitation of data laborers.
Conclusion and Future Directions
The NoBIAS project provides a structured taxonomy for AI fairness, offering a curated path through the state-of-the-art while linking technical metrics back to their philosophical and economic origins.
However, limitations remain. The policy advice is heavily centered on the European context, and the field is evolving so rapidly that no list of best practices can be exhaustive. Future work must focus on the "implementation gap"—developing practical methods for deploying these interdisciplinary frameworks within fast-moving, real-world production environments.
Summary updated