🤖 machine learning

Automatic Causal Fairness Analysis with LLM-Generated Reporting

This paper introduces \textsc{FairMind}, an automated tool that leverages causal fairness assumptions and large language models to perform sound, counterfactual-based fairness analysis on datasets and generate accurate zero-shot reports.

Original authors: Alessia Berarducci, Eric Rossetto, Alessandro Antonucci, Marco Zaffalon

Published 2026-05-01

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Alessia Berarducci, Eric Rossetto, Alessandro Antonucci, Marco Zaffalon

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you have a machine that makes decisions, like a loan officer or a hiring manager. You want to know if this machine is fair, or if it's secretly biased against certain groups of people. Usually, checking for this bias is like trying to solve a complex puzzle while wearing a blindfold; you need a human expert to look at the data, guess the hidden connections, and explain what's going on.

This paper introduces FairMind, a new tool that acts like an "automated fairness detective" that doesn't need a blindfold. It combines hard math with a smart AI writer to automatically find, measure, and explain bias in data.

Here is how it works, broken down into simple concepts:

1. The "Standard Fairness Model" (The Blueprint)

Before FairMind can work, it needs a map of how the world works. The authors use a specific blueprint called the Standard Fairness Model.

The Analogy: Imagine a river flowing from a mountain (the Protected Feature, like gender or race) to a lake (the Target, like getting a job or a loan).
The Problem: Sometimes, the river gets polluted by a hidden factory upstream (a Confounder, like education level or zip code) that dumps trash into the water, making it look like the mountain is dirty when it's actually the factory.
The Solution: FairMind uses a mathematical "filter" to separate the water coming from the mountain from the pollution coming from the factory. It calculates exactly how much of the outcome is caused by the protected feature itself versus how much is just a side effect of other factors.

2. The Math Part (The Calculator)

FairMind doesn't just guess; it uses strict formulas to calculate three main things:

Total Variation: The raw difference you see in the data (e.g., "Men get loans 20% more often than women").
Total Effect: The real causal difference if you could magically change someone's gender without changing anything else about them.
Spurious Effect: The "fake" difference caused by those hidden factories (confounders).

The tool also breaks these down further. It can tell you if the bias happens directly (the machine hates women) or indirectly (the machine hates women because it thinks women work fewer hours, which is a middleman variable).

3. The "LLM" Part (The Translator)

This is where the paper gets creative. Usually, these math results come out as a confusing spreadsheet of numbers that only a statistician can read.

The Problem: If you just ask a smart AI (a Large Language Model or LLM) to look at the raw data and tell you if it's fair, it often makes mistakes. It might do the math wrong, hallucinate facts, or get confused by large datasets. It's like asking a brilliant writer to do your taxes without a calculator; they might write a beautiful essay, but the numbers will be wrong.
The Fix: FairMind does the math first. It crunches the numbers and creates a clean, structured report. Then, it hands this report to the LLM.
The Result: The LLM acts as a translator. It takes the dry, accurate numbers and writes a clear, human-friendly story. It says things like, "The data shows a 5% gap, but 90% of that gap is actually due to education levels, not gender."

4. Why This is Better Than Just Using AI

The paper tested two approaches:

Pure AI: Asking the AI to read the data and write a report.
FairMind + AI: Using the tool to calculate, then the AI to write.

The Verdict: The "Pure AI" approach was unreliable. It got the numbers wrong, missed important details, and struggled with even small datasets. The FairMind approach was accurate, efficient, and produced reports that were easy for non-experts to understand. It's like the difference between asking a poet to build a bridge (Pure AI) versus asking an engineer to build the bridge and a poet to describe it (FairMind).

5. Handling Complex Scenarios

The tool isn't just for simple "Yes/No" or "Male/Female" situations. The paper shows it can handle:

Multiple Steps: If there are many middlemen (like hours worked, job type, and location), it can untangle which one is causing the most trouble.
Ordered Categories: If the protected feature is something like "Education Level" (High School, Bachelor's, Master's), it can check if the bias gets worse as you go up the ladder.
Continuous Numbers: If the outcome is a number (like salary amount), it can analyze the average difference rather than just counting "yes/no" outcomes.

Summary

FairMind is a software prototype that automates the detection of unfair bias in data. It uses rigorous math to ensure the numbers are right, and then uses a smart AI to translate those numbers into plain English. This allows regular people, not just data scientists, to understand why a machine learning model might be unfair and exactly where the bias is coming from.

1. Problem Statement

While Automated Machine Learning (AutoML) aims to streamline the application of ML to real-world problems, existing frameworks largely ignore fairness issues inherent in training data and predictions. Current approaches to causal fairness (e.g., counterfactual fairness) typically require:

Manual modeling decisions by domain experts.
Complex interpretation of causal graphs and identification formulae.
Significant human intervention, which contradicts the "zero-shot" or minimal-human-intervention philosophy of AutoML.

There is a lack of automated tools that can not only compute causal fairness metrics but also interpret and report these findings in a way accessible to non-experts. Furthermore, relying solely on Large Language Models (LLMs) to analyze raw data for fairness is inefficient, prone to hallucination, and raises privacy concerns.

2. Methodology: The FairMind Framework

The authors introduce FairMind, a software prototype designed to automate fairness analysis at the dataset level. The methodology integrates three core components:

A. Theoretical Foundation: Standard Fairness Model (SFM)

The framework is built upon the Standard Fairness Model proposed by Plečko and Bareinboim. It assumes a Structural Causal Model (SCM) where:

$X$ : A protected feature (e.g., gender, race).
$Y$ : The target variable.
$W$ : Mediators (intermediate factors).
$Z$ : Confounders (shared causes of $X$ and $Y$ ).
The model assumes no hidden confounding between specific variable groups ( $X, Z$ and $W, Y$ ).

B. Automated Causal Computation

FairMind automates the calculation of causal effects using identification formulae derived from observational data ( $P(V)$ ). It computes:

Total Variation (TV): Observational difference $P(y|x_1) - P(y|x_0)$ .
Total Effect (TE): Causal difference via intervention $P(Y_{x_1}) - P(Y_{x_0})$ .
Spurious Effect (SE): The difference between TV and TE (confounding bias).
Direct Effect (DE): Effect of $X$ on $Y$ holding mediators constant at their natural state under $x_0$ .
Indirect Effect (IE): Effect of $X$ on $Y$ mediated through $W$ .

The tool uses Bayesian Networks (inferred via the pgmpy library with Laplace smoothing) to estimate these probabilities. It supports closed-form computation of these effects, avoiding the need for complex structural equation modeling by the user.

C. Novel Decompositions and Extensions

The paper extends the standard SFM to handle complex real-world scenarios:

Multiple Mediators/Confounders: New decomposition formulae (Propositions 3 & 4) allow attributing specific portions of the Indirect Effect and Spurious Effect to individual mediators and confounders.
Non-Binary Targets: For continuous or ordinal targets, the framework uses expectations (or utility functions) rather than simple probability differences to preserve signed measures of fairness.
Non-Binary Protected Features: It handles ordinal and multi-class protected variables by averaging pairwise effects or using stepwise decomposition to analyze disparities across ordered levels (e.g., education levels).

D. LLM-Generated Reporting

To bridge the gap between quantitative results and human understanding, FairMind employs a two-stage pipeline:

Deterministic Computation: FairMind calculates all metrics and outputs a structured JSON object.
LLM Interpretation: A Large Language Model (specifically chatgpt-05-mini) acts as a reporting agent. It receives only the structured numerical results (not raw data) and a Chain-of-Thought (CoT) prompt.
- The prompt enforces a strict analytical structure, defining the role as a "causal fairness expert."
- This ensures the report is semantically anchored to the computed data, minimizing hallucinations and ensuring logical coherence.

3. Key Contributions

FairMind Tool: The first AutoML-integrated tool that automates the selection, estimation, and explanation of causal fairness effects with minimal user input.
Theoretical Extensions: Novel decomposition results for multiple mediators and confounders, and extensions to ordinal protected variables and continuous targets.
Hybrid Neuro-Symbolic Approach: A robust architecture combining symbolic causal inference (for accurate computation) and neural LLMs (for natural language reporting).
Empirical Validation: Demonstrations on the UCI Adult Census and Student-Mat datasets showing that the tool can identify whether disparities are driven by direct discrimination, mediation, or confounding.

4. Results and Findings

Accuracy vs. Pure LLM: Experiments comparing FairMind+LLM against a "Pure LLM" approach (where the LLM analyzes raw data directly) show that the hybrid approach is superior.
- Accuracy: Pure LLMs produced incorrect calculations (e.g., wrong Total Effect values) and misinterpreted mediator contributions.
- Efficiency: The hybrid approach required significantly fewer tokens (5,673 vs. 15,150) and could handle larger datasets, whereas the Pure LLM approach became infeasible for larger data.
- Privacy: The hybrid approach protects privacy by processing raw data in a secure local environment and sending only numerical summaries to the LLM.
Case Study Insights:
- In the Adult Census dataset, the observed income disparity between genders was found to be largely spurious (driven by the confounder "relationship status") rather than a direct causal effect.
- In the Student-Mat dataset, the framework successfully quantified how gender affects exam pass rates, distinguishing between direct effects and indirect effects mediated by study time.

5. Significance

This work represents a significant step toward democratizing causal fairness analysis. By automating the complex mathematical requirements of counterfactual fairness and translating them into accessible natural language reports, FairMind enables:

Non-experts (data scientists, policymakers) to conduct rigorous fairness audits without deep causal inference expertise.
AutoML pipelines to include fairness as a standard, automated evaluation metric.
Scalable and Private fairness auditing, avoiding the risks of sending sensitive raw data to external LLMs.

The paper concludes that this neuro-symbolic integration offers a scalable path forward for responsible AI, suggesting future work could involve automated structural learning to infer causal graphs and automated mitigation strategies.