Automatic Causal Fairness Analysis with LLM-Generated Reporting
New tool computes counterfactual bias and generates zero-shot explanations using LLMs
FairMind, a new software prototype from researchers at IDSIA (Alessia Berarducci, Eric Rossetto, Alessandro Antonucci, Marco Zaffalon), tackles a critical gap in AutoML: most frameworks ignore dataset-level fairness. The tool implements the standard fairness model by Plečko and Bareinboim, enabling sound causal fairness evaluation through counterfactual queries. It computes direct, indirect, and total effects of a protected attribute on a target variable, properly accounting for confounders and mediators. This approach goes beyond correlation-based fairness metrics by quantifying actual causal harm.
After automatic data preprocessing, FairMind performs closed-form computation of causal effects—including novel decomposition results for ordinal protected variables and continuous targets. The tool then leverages LLMs in a zero-shot setup to generate human-readable fairness reports that explain which groups are disadvantaged and by how much. The authors demonstrate that LLM-generated reports from FairMind's structured causal outputs are more accurate and less hallucinated than asking an LLM to analyze raw data directly. This makes it practical for ML teams to audit training data for bias without deep causal expertise.
- FairMind uses the standard fairness model by Plečko and Bareinboim for causal counterfactual analysis, computing direct, indirect, and total effects of protected attributes.
- The tool handles ordinal protected variables and continuous targets with novel decomposition results, enabling application beyond binary sensitive features.
- LLMs generate zero-shot fairness reports from structured causal outputs, outperforming direct LLM analysis on raw data (less hallucination, more accurate).
Why It Matters
Brings rigorous causal fairness into AutoML pipelines, making bias detection accessible to non-experts via LLM summaries.