TLDR: Joachim Baumann’s dissertation, “On the Societal Impact of Machine Learning,” provides a comprehensive framework to address the ethical challenges of AI. It outlines three core objectives: to measure unfairness using domain-specific criteria (e.g., ‘sufficiency’ in insurance), to decompose complex ML systems to understand the dynamics of various biases and feedback loops, and to develop practical interventions. These interventions include optimal decision-making under fairness constraints, algorithmic collective action for user empowerment, and real-world applications like an equitable rental assistance allocation system. The work emphasizes a proactive, long-term approach to ensure ML systems align with societal values.
Machine learning (ML) systems are increasingly integrated into our daily lives, influencing critical decisions and recommendations across various sectors. From insurance pricing to job applications and social service allocations, these data-driven systems carry a significant risk of perpetuating or even amplifying existing societal biases, leading to discriminatory effects. A recent dissertation by Joachim Baumann delves into this crucial issue, offering a comprehensive framework to understand and address the societal impact of ML.
The thesis, titled “On the Societal Impact of Machine Learning” by Joachim Baumann, explores three interconnected goals to tackle the challenges of fairness in ML systems. These goals are to measure, decompose, and fix the issues that arise from the widespread use of these powerful algorithms.
Measuring Fairness in ML Systems
The first goal focuses on establishing robust frameworks to quantify ML’s impact on society, particularly in terms of unfairness. The research highlights that defining fairness is not a one-size-fits-all problem; it requires domain-specific ethical considerations. For instance, in insurance pricing, the study argues that certain fairness criteria, like ‘independence’ (statistical parity) and ‘separation’ (equalized odds), are not morally appropriate. Instead, ‘sufficiency’ (well-calibration) is proposed as a defensible criterion, allowing insurers to detect systematic biases in premiums across demographic groups. This approach connects technical fairness criteria to philosophical theories of distributive justice, revealing the normative choices embedded in technical implementations.
Decomposing Bias and Feedback Loops
The second goal involves breaking down complex ML systems into their fundamental components to systematically analyze how biases emerge and evolve over time. Traditional fairness assessments often adopt a static view, overlooking the dynamic interplay between ML models and their environment. The research introduces a framework for generating synthetic datasets with controlled forms of bias, enabling the isolation and study of historical, measurement, representation, and omitted variable biases. This decomposition helps in understanding how biases are amplified through feedback loops, where algorithmic decisions influence future data collection and outcomes. For example, predictive policing algorithms can direct increased surveillance to certain neighborhoods, leading to more arrests that further justify heightened policing, creating a self-reinforcing cycle.
Also Read:
- Unlocking AI’s Reasoning: A Causal Path to Interpretable Models
- Navigating Uncertainty: How Partial Calibration Enhances Trustworthy AI Decisions
Developing Practical Interventions
The third goal is dedicated to developing effective interventions that mitigate unfairness and contribute to a positive societal impact of ML. The thesis explores various bias mitigation techniques, including pre-processing (transforming training data), in-processing (modifying learning algorithms), and post-processing (adjusting decision rules after model training). A notable contribution is the derivation of optimal solutions for decision-making under specific fairness constraints, such as positive predictive value (PPV) parity and false omission rate (FOR) parity. Interestingly, some optimal solutions might involve counter-intuitive approaches, like selecting individuals with the smallest utility for one group to achieve overall fairness, highlighting trade-offs between group and individual fairness.
Beyond direct model modifications, the research also proposes ‘algorithmic collective action’ strategies. This innovative approach empowers users to leverage their coordinated behavior to indirectly steer recommendation systems, especially when direct model modification is not possible. An example is promoting emerging artists on music streaming platforms by strategically reordering playlists. Furthermore, the thesis demonstrates how responsibly developed ML systems can benefit society. In collaboration with Allegheny County’s Department of Human Services, a proactive outreach system was built to distribute rental assistance more effectively, helping to prevent homelessness. This system has moved from concept to active deployment, with a randomized controlled trial underway to evaluate its effectiveness.
This comprehensive work provides a foundation for ensuring that the societal impact of machine learning aligns with broader social values, addressing continuous challenges and future research directions as ML systems, including generative AI, become increasingly integrated into society. For more details, you can read the full dissertation available at arXiv:2510.23693.
