Zusammenfassung
Je mehr KI-gestützte Entscheidungen das Leben von Menschen betreffen, desto wichtiger ist die Fairness solcher Entscheidungen. Dieser Beitrag gibt eine Einführung in die Forschung zu Fairness in KI-Systemen, erklärt die wesentlichen Fairness-Definitionen und Strategien zur Erreichung von Fairness anhand konkreter Beispiele und ordnet die Fairness-Forschung in den europäischen Kontext ein. Weder in der europäischen Gesetzgebung noch in der KI-Forschung kommt es dabei zu einem Konsens, wie Fairness zu definieren und zu erreichen ist. Stattdessen muss für jedes System eine differenzierte und kontextabhängige Betrachtung von möglichen unfairen Ergebnissen und deren Konsequenzen erfolgen. Dieser Beitrag kann dabei unterstützen und richtet sich an ein interdisziplinäres Publikum, indem auf mathematische Formulierungen verzichtet wird und stattdessen Visualisierungen und Beispiele genutzt werden.
Introduction: Dismantling the Myth of Algorithmic Objectivity
As AI systems increasingly take over life-altering decisions—from hiring processes to criminal sentencing—the "myth of algorithmic objectivity" is being dismantled. While these systems appear mathematically objective, they often automate and amplify historical discrimination. The core challenge lies in the fact that "fairness" is not a single, universal metric. This paper provides a framework for navigating these trade-offs, bridging technical implementation with the socio-legal requirements of the upcoming European AI Act.
Defining Fairness: A Multi-Faceted Framework
Fairness research is generally categorized into several distinct pillars, each addressing different aspects of equity:
Group Fairness
Group fairness focuses on comparing outcomes across protected groups (e.g., gender, ethnicity). To analyze this, we can visualize a classification task where individuals are categorized by their actual class and their membership in a protected group.
Key metrics within this pillar include:
- Demographic Parity: Ensuring that the probability of a positive outcome is equal across all groups, regardless of historical base rates.
- Equalized Odds: Balancing error rates (false positives and false negatives) so that no group is disproportionately misclassified.
Individual, Causal, and Dynamic Fairness
Beyond group metrics, Individual Fairness ensures that similar individuals are treated similarly using mathematical distance measures. Causal Fairness utilizes causal graphs to determine if a sensitive attribute has an unjustified path of influence on a decision. Finally, Dynamic Fairness examines long-term feedback loops, such as how a biased hiring tool might discourage certain demographics from applying in the future.
Mitigation and Auditing Strategies
Bias can be mitigated at three distinct stages of the machine learning pipeline:
- Pre-processing: Modifying training data to remove correlations between sensitive attributes and labels.
- In-processing: Integrating fairness constraints directly into the model's loss function during training.
- Post-processing: Adjusting decision thresholds of a pre-trained model to balance outcomes.
Developers can utilize audit tools like Aequitas and AI Fairness 360 to test models for bias without requiring deep access to the underlying code.
Key Insights: The Incompatibility Theorem
A critical realization in this field is the Incompatibility Theorem: it is mathematically impossible to satisfy all fairness definitions simultaneously if groups have different "base rates" (historical success rates). Choosing a fairness metric is therefore a moral and political choice, often reflecting one of two worldviews:
- "We Are All Equal" (WAE): Assumes observed differences are the result of structural discrimination and should be ignored.
- "What You See Is What You Get" (WYSIWYG): Assumes data reflects a reality the AI should accurately predict.
The Regulatory Landscape and Future Outlook
The European AI Act classifies systems in employment, education, and law enforcement as "high-risk," legally requiring rigorous bias testing. However, technical solutions alone are insufficient. Future progress requires interdisciplinary cooperation between engineers, social scientists, and affected communities to translate high-level legal principles into concrete engineering requirements. As we move forward, the focus must shift toward Dynamic Fairness to understand how AI decisions reshape society over decades.
Summary updated