Two competing definitions of AI fairness turn out not to guarantee each other when images are involved.
Researchers built new datasets on top of existing image fairness benchmarks to test both counterfactual fairness — whether a model's output changes when a sensitive attribute like secondary sex characteristics is edited — and group fairness, which measures outcome disparities across demographic groups. Prior work on tabular data had suggested the two properties travel together. In image classification, they do not. The study finds the culprit is a latent variable that correlates with the sensitive attribute without being caused by it: secondary sex characteristics are strongly correlated with hair length, so a model can appear individually fair while still keying off a proxy. The team also proposes a method called Counterfactual Knowledge Distillation to reduce reliance on these correlated stand-ins.
The gap matters because most fairness audits pick one definition and call it done. If image models that pass counterfactual checks can still fail group-parity tests — and vice versa — then deploying either standard alone gives a false sense of coverage, particularly in high-stakes domains like hiring tools or facial-recognition systems.
This is a pre-print and has not yet been peer reviewed, so the specific benchmark results should be treated as preliminary — but the conceptual argument, that correlated proxies can launder apparent individual fairness into group disparity, is straightforward enough to take seriously now.