Implementing Causal Perception: Competing SCMs and... | AI Research

This paper provides the first practical implementation of the causal perception framework originally proposed by Álvarez and Ruggieri (2025). It addresses the problem of how different agents—such as humans or machine learning models—can interpret the same system differently because they rely on competing Structural Causal Models (SCMs). By operationalizing these models, the research shows that fairness and decision-making are "situated," meaning they depend on the specific worldview or causal assumptions held by the agent.

Operationalizing Causal Perception

The framework models how agents disagree on the causal structure of a system. The implementation distinguishes between two types of disagreement: structural perception, where agents disagree on the causal graph (the cause-effect pairs), and parametrical perception, where agents agree on the graph but disagree on the weights of those causal relationships.
To make this framework executable, the author developed algorithms for both single-intervention and multi-intervention settings. These algorithms allow for calculations at the second rung (interventional) and third rung (counterfactual) of Pearl’s causal ladder. For counterfactual reasoning, the process includes an abduction step to infer latent variables from observed data before simulating hypothetical scenarios.

Measuring Disagreement

Because the framework does not mandate a specific way to measure the difference between two agents' worldviews, the implementation utilizes three standard distance measures:

• Wasserstein-2 distance: Measures the "cost" of moving probability mass between distributions, capturing differences in both location and shape.

• Kullback-Leibler divergence: An information-theoretic measure that quantifies distributional surprise.

• Total Variation distance: A bounded measure that identifies the maximum probability difference between two events, providing a clear scale for interpreting thresholds.
  The research uses these metrics to quantify the distance between the probability distributions implied by competing SCMs. When this distance exceeds a chosen threshold, the agents are classified as having different causal perceptions.

Impact on Fairness and Decisions

Using the German Credit dataset, the author illustrates how these competing worldviews affect real-world outcomes. In a multi-expert decision setting, the study found that even minor disagreements in a single causal edge led to significant differences in fairness metrics and individual loan decisions. For example, two receivers with slightly different SCMs produced demographic parity gaps of -0.594 and -0.170, respectively, and disagreed on 21.7% of loan approvals.

Considerations for Implementation

The research notes that the "perception verdict"—whether two agents are considered to be in disagreement—is sensitive to the choice of distance metric and the threshold used. Because the number of possible interventions can grow exponentially with the number of variables, the current implementation relies on a subset of relevant interventions based on domain knowledge. The author suggests that future work should focus on developing more efficient algorithms to handle larger sets of variables. The findings emphasize that because bias is situated relative to an agent's SCM, it is not possible to treat fairness as a universal, objective fact when competing worldviews exist.