Proxy Test

The proxy test, which NYDFS Circular Letter No. 7 calls the proxy assessment, is an exercise the insurer runs, not one the Department runs on the insurer. It asks whether an external consumer data source is a proxy for a protected class, even though the protected class itself never enters the model.

There are two moves. The first is measurement: how far does the data source track membership in a protected class where that tracking could produce unfair or unlawful discrimination? NYDFS treats “correlated with” and “proxy for” as the same question, and accepts either data the carrier already has or a reasonable inference drawn with accepted statistical methods. The second move is justification, and only if the first turns something up: is this use required by a legitimate business necessity? The letter does not define legitimate business necessity, and it does not mention less discriminatory alternatives here.

Two limits are easy to miss. The proxy assessment as written covers external consumer data, not artificial intelligence systems. A proprietary model built on traditional data is caught by the letter’s separate three-step assessment instead. And the quantitative side reaches only those protected classes the carrier can identify or reasonably impute from what it has. Going out to gather more data about individuals is explicitly not expected.

A variable can fail the proxy test even if it is facially neutral. NYDFS expects insurers to document how they test for proxy effects and what they do when a proxy turns up. Our guide to NY DFS Circular Letter No. 7 explains both assessments in underwriting practice, and our AI in underwriting guide covers the broader compliance framework.