Implementing Causal Perception: Competing SCMs and Situated Fairness
Researchers provided the first implementation of the causal perception framework by Álvarez and Ruggieri (2025), operationalizing structural and parametrical causal perception. They designed algorithms for computing interventional and counterfactual distributions and proposed distance measures to quantify disagreement. Using the German Credit dataset, they demonstrated that causal perception affects accuracy and fairness in multi-expert decision settings, and that the perception verdict is sensitive to the choice of distance metric and threshold.
This work implements the theoretical causal perception framework, where agents with competing Structural Causal Models (SCMs) infer different probability distributions under the same interventions. The implementation covers both structural disagreement (different causal graphs) and parametrical disagreement (same graph, different weights). Algorithms compute interventional and counterfactual distributions, and distance measures quantify the disagreement. Experiments on the German Credit dataset show that causal perception influences accuracy and fairness assessments, and that the perception verdict depends on the chosen distance metric and threshold. The study highlights how causal perception changes fairness evaluations and threshold-based decisions, revealing that bias proves sensitive to these factors.
The implementation operationalizes causal perception by computing interventional and counterfactual distributions from competing SCMs and quantifying disagreement via distance measures. The sensitivity of the perception verdict to the choice of distance metric and threshold indicates that practical deployment requires careful metric selection and calibration. The use of the German Credit dataset demonstrates applicability to real-world fairness-sensitive domains.
In multi-expert AI systems, such as those used in credit scoring or hiring, causal perception can lead to divergent fairness assessments and decision outcomes. Organizations deploying such systems must account for model disagreement and establish robust fairness evaluation protocols that consider multiple causal perspectives.
This research provides a foundation for building more transparent and fair multi-expert AI systems. By quantifying causal disagreement, businesses can better audit and align AI decisions with ethical standards, potentially reducing legal and reputational risks in regulated industries.
Future work may explore automated selection of distance metrics and thresholds, extension to dynamic causal models, and integration with explainability tools. The framework could be applied to other fairness-critical domains like healthcare and criminal justice. Observing adoption in regulatory guidelines or fairness toolkits would be a next signal.