A new paper on arXiv addresses the challenge of making fairness claims about AI systems when the causal relationships behind outcomes are not fully known. The authors propose a principle they call 'causal humility' to guide how such claims should be made and governed.
The paper notes that AI systems are increasingly used in consequential areas like employment, lending, healthcare, and public services, making fairness a key concern for auditing and governance. However, fairness assessment often depends on causal assumptions that may be uncertain, and the paper argues for a more cautious approach.
Causal humility, as the name suggests, would temper the authority of fairness claims when causal evidence is incomplete. This could affect how auditors and regulators interpret and act on fairness metrics.