Causal Dependence Plots

CDPs provide more realistic explanations by accounting for real-world relationships: Unlike traditional explanation methods that artificially freeze other variables in place, Causal Dependence Plots show how changing one input (like income) naturally affects not just the model's output, but also related variables (like employment status), mimicking how these factors actually interact in the real world.

Causal assumptions must be made explicit for trustworthy AI explanations: The paper emphasizes that "causal conclusions require causal assumptions," meaning organizations using CDPs need to clearly document their beliefs about how variables relate to each other—this transparency is essential for responsible deployment of AI systems in high-stakes domains.

CDPs can be combined with fairness and sensitivity analysis tools: The research demonstrates that CDPs work modularly with other methods for detecting bias and testing model robustness, making them practical for organizations conducting algorithmic fairness audits or stress-testing AI systems before deployment.

This method bridges how humans naturally think about causality and how AI systems work: Since people intuitively reason about cause-and-effect relationships, CDPs make model behavior more interpretable to non-technical stakeholders, regulators, and affected communities—improving the ability to identify when AI systems behave unexpectedly or unfairly.