"We don't collect race, so our model can't discriminate by race." It is one of the most common things teams say about their hiring, lending or screening models, and one of the most expensive misunderstandings. A model doesn't need to see race to learn it.
What a proxy is
A proxy is an input that looks neutral but is closely linked to a protected characteristic. Common ones:
- Schools, clubs and past employers, which can stand in for race, sex or age.
- Gaps in employment, which can stand in for caregiving, disability or pregnancy.
If a model is trained on past decisions that were biased, it will use whatever proxies it finds to reproduce them. Deleting the race column changes nothing; the information is still in the other columns. New Jersey's disparate impact rules give this exact scenario as an example: an automated tool trained on a mostly white, male workforce that favors applicants who resemble current employees, with no intent to discriminate.
Why it matters legally
Disparate impact law looks at outcomes, not intent. If a neutral-looking practice disadvantages a protected group, the organization has to show it is necessary for a legitimate purpose, and a plaintiff can still win by pointing to a less discriminatory alternative. Not collecting race is not a defense; it only means you can't see the problem coming.
Federal agencies have pulled back from disparate impact enforcement since 2025, but the theory remains part of Title VII, and states are moving the other way. New Jersey's rules, in effect since December 15, 2025, explicitly name automated decision-making tools. Colorado has also passed legislation on automated decisions in employment, lending, housing, insurance and other areas, currently set to apply from January 1, 2027, though a federal court order and pending rules make its timing worth checking.
You can measure bias you can't see
Not collecting race does not mean you can't check for it. An independent review can show whether your decisions treat some groups worse, and by how much, without asking anyone their race. It works for groups, never for deciding what one person is.
Where to start
List the decisions in your organization that a model or a scoring rule makes or ranks: résumé screening, loan approvals, tenant screening, complaint routing, fraud flags. For each one, ask whether it uses inputs like name, address or school, and whether anyone has ever checked the outcomes by group. If the answer is no, that is where to look first.
Fairspire's proxy bias check shows you whether that gap exists in your own decisions, and how big it is. Personal data never leaves your systems.