Can Fairness be Automated? Guidelines and Opportunities for Fairness-aware AutoML

TL;DR

This paper analyzes whether fairness can be fully automated and argues that while fairness cannot be completely automated, fairness-aware AutoML can substantially aid practitioners by integrating best practices, accelerating fairness evaluation, and enabling systematic exploration of fairness–performance trade-offs. It reviews foundational concepts in algorithmic fairness and AutoML, discusses user-centric and system-centric challenges, and highlights risks such as data bias, metric misalignment, automation bias, and benchmarking pitfalls. The authors propose guidelines (e.g., clear assumptions, seamful design, robust evaluation) and outline opportunities where AutoML can improve modeling practices, transparency, and learning process inspection, while calling for contextual benchmarks and interdisciplinary collaboration. The work emphasizes that responsible, context-aware collaboration between humans and AutoML is essential for achieving fair real-world outcomes and proposes a research agenda toward interactive, user-centered fairness tooling.

Abstract

The field of automated machine learning (AutoML) introduces techniques that automate parts of the development of machine learning (ML) systems, accelerating the process and reducing barriers for novices. However, decisions derived from ML models can reproduce, amplify, or even introduce unfairness in our societies, causing harm to (groups of) individuals. In response, researchers have started to propose AutoML systems that jointly optimize fairness and predictive performance to mitigate fairness-related harm. However, fairness is a complex and inherently interdisciplinary subject, and solely posing it as an optimization problem can have adverse side effects. With this work, we aim to raise awareness among developers of AutoML systems about such limitations of fairness-aware AutoML, while also calling attention to the potential of AutoML as a tool for fairness research. We present a comprehensive overview of different ways in which fairness-related harm can arise and the ensuing implications for the design of fairness-aware AutoML. We conclude that while fairness cannot be automated, fairness-aware AutoML can play an important role in the toolbox of ML practitioners. We highlight several open technical challenges for future work in this direction. Additionally, we advocate for the creation of more user-centered assistive systems designed to tackle challenges encountered in fairness work

Figures (2)

• Figure 1: Example of an ML workflow adapted from the CRISP-DM ( ? ) process. Developing ML systems is an iterative process (dotted arrows) that can require frequently revisiting decisions made in previous stages. Most existing fairness-aware AutoML methods and systems address fairness analogous to fairness-aware ML techniques: the problem is formulated as an optimization task under a fairness objective or constraint. However, many important design choices are made outside of the modeling stage which is typically the part of the workflow that is tackled by AutoML systems. Taking fairness into account adds additional considerations (in blue) to every step of the ML workflow.
• Figure 2: Multi-objective perspective (left) and constrained perspective (right) showing exemplary unfairness and error values of found ML models (black dots). We depict optimal solutions returned by the minimization procedure as blue triangles, the approximation of the Pareto-front with a dashed line and the constraint (unfairness $>$ 0.075) as a red area.