TLDR: This research introduces a novel framework using formal abductive explanations and background knowledge to diagnose proxy discrimination and unfairness in individual AI decisions. It identifies features acting as unjustified proxies for protected attributes, revealing hidden structural biases. By introducing “aptitude” and “mapping functions,” the framework assesses fairness by ensuring individuals with equivalent aptitudes receive similar treatment across different groups, even suggesting that sometimes bias might be necessary for fairness.
Artificial intelligence systems are increasingly making important decisions in areas like finance, healthcare, and criminal justice. While these systems offer many benefits, they also raise serious concerns about fairness, discrimination, and transparency. One major issue is “proxy discrimination,” where seemingly neutral features in an AI model can indirectly encode sensitive information, leading to unfair outcomes for certain groups.
Current methods for auditing AI fairness often struggle to uncover why unfairness occurs, especially when it’s deeply rooted in structural biases within the data or the system’s design. This new research introduces a groundbreaking framework that uses “formal abductive explanations” to shed light on these hidden forms of discrimination in individual AI decisions.
Understanding the Problem: Proxy Discrimination
Imagine an AI system deciding on credit applications. It might not directly use a protected attribute like gender. However, if certain features, like marital status or credit purpose, are strongly correlated with gender in the training data, the AI could inadvertently use these features as “proxies” for gender, leading to discriminatory decisions. This is proxy discrimination – where neutral inputs act as stand-ins for sensitive attributes, producing biased results.
The challenge is that traditional bias detection methods, which often rely on statistical checks or simply looking for explicit use of protected attributes, can miss these subtle, indirect forms of bias. The goal of this research is to provide a more profound understanding, moving beyond just detecting that bias exists to explaining why it arises.
A Novel Approach: Abductive Explanations and Background Knowledge
The core of this framework lies in “abductive explanations.” Unlike “what-if” scenarios (counterfactual explanations), abductive explanations provide logical proofs for why a specific decision was made. They answer the question: “Why did the system produce this outcome for this individual?” By identifying the minimal set of features sufficient to guarantee a decision, abductive explanations can pinpoint the causes embedded in decision outcomes.
A crucial element introduced by the researchers is “background knowledge.” This refers to a set of real-world constraints or relationships within the data. For example, in a credit dataset, background knowledge might reveal that it’s impossible to find a female applicant whose credit purpose is a car while also being single. This kind of knowledge helps identify when a variable is acting as a proxy. A variable is considered a proxy if, within a certain context defined by background knowledge, knowing its value allows one to infer the value of a protected attribute.
The paper demonstrates that when background knowledge is considered, an AI system that appears unbiased by conventional definitions might still exhibit bias through proxy discrimination. This means that an explanation for a decision might not explicitly mention a protected attribute, but it could apply only to individuals sharing a specific protected feature (e.g., only male applicants).
Ensuring Fairness Through Aptitude Equivalence
To address unfairness, the framework introduces the concepts of “aptitude” and “mapping functions.” Aptitude is defined as a task-relevant property that should be independent of group membership. Fairness, then, requires that individuals with equivalent aptitudes receive similar treatment, regardless of their protected attributes.
The researchers propose “mapping functions” to align individuals of equivalent aptitude across different groups. This allows for the comparison of explanations between subgroups. For instance, an explanation for a male applicant’s credit approval can be mapped to a “counterpart explanation” for a female applicant. If both explanations, representing equivalent aptitudes, lead to the same decision, the decision is considered fair.
Interestingly, the research highlights that in some situations, including a protected attribute or its proxies in an explanation might be necessary to ensure fairness and subgroup equivalence. This nuanced perspective moves beyond simply removing protected attributes to understanding their complex interplay in achieving equitable outcomes.
Also Read:
- Measuring Spatial Fairness in AI: The GeoBS Approach
- Uncovering Hidden Biases in AI: Introducing humancompatible.detect
Looking Ahead
This formal framework, developed by Belona Sonna and Alban Grastien, offers a powerful new way to diagnose subtle, structural discrimination in AI decisions. By leveraging abductive reasoning and domain knowledge, it provides interpretable, case-specific explanations of bias, going beyond aggregate statistical checks. The work, detailed in their paper available at arXiv:2509.25662, also suggests potential for extension to non-binary data and supports intersectional fairness without complex re-engineering. While currently demonstrated with examples from the German Credit dataset, future work will involve empirical studies on larger, real-world datasets and exploring interactive explanation tools to enhance transparency for non-expert users.
