Implementing Causal Perception: Competing SCMs and Situated Fairness

Imagine you are trying to solve a mystery, like figuring out why a plant in your room is wilting. You might look at the sunlight, the water, or the soil. In the world of Artificial Intelligence (AI), scientists use special maps called "Structural Causal Models" to do the same thing. These maps are like flowcharts that show how different things cause other things to happen. For example, a map might show that "low income" causes "poor credit score." But here is the twist: two different people (or AI programs) might look at the exact same plant and draw two completely different maps. One might think the soil is the problem, while the other thinks it's the sunlight. This isn't just a mistake; it's a difference in how they see the world. This idea is crucial because when we ask AI to make fair decisions—like who gets a loan or a job—those decisions depend entirely on which map the AI is using. If the maps disagree, the idea of "fairness" can change depending on who is looking at it.

This paper, written by José M. Álvarez, takes a big step from theory to reality by building the first computer program that can actually measure these "clashing maps." The author calls this phenomenon "causal perception." Think of it like two friends watching a movie. Even if they see the same scenes, they might interpret the characters' motivations differently based on their own life experiences. In the AI world, this means two agents (or "receivers") might look at the same group of loan applicants and see two different stories about why some people get rejected. The paper tests this using a real dataset of 1,000 loan applicants from Germany. The researchers created two digital "credit officers" who agreed on most things but disagreed on one specific rule: whether a person's gender directly affects their credit risk.

The results were eye-opening. Even though the two officers only disagreed on a single connection in their mental maps, they ended up seeing the world very differently. When the researchers measured how far apart their views were, they found that the two officers disagreed on about 21.7% of individual loan decisions. The paper also highlights that such disagreements can lead to vastly different fairness assessments; for instance, in a hypothetical scenario described in the introduction, two receivers differing on a single causal edge could reach demographic parity gaps of -0.594 versus -0.170. The paper also discovered that the answer to "are they seeing things differently?" depends entirely on which measuring tool you use. If you use one type of ruler (a distance metric called Wasserstein-2), they might look almost the same. But if you use a different ruler (like Kullback-Leibler divergence or Total Variation), the gap looks massive.

Ultimately, the paper shows that bias isn't just a fixed fact floating in the air; it is "situated," meaning it depends on the specific worldview of the person or AI doing the judging. The author argues that we can't just ignore these competing worldviews or force everyone to agree on a single "correct" map. Instead, we need to acknowledge that in a world with multiple experts, different people will genuinely perceive fairness differently, and our tools need to be ready to handle that disagreement rather than pretending it doesn't exist.