Algorithmic Lending and the New Face of Discrimination Law

A credit-scoring model that never directly uses caste, religion, or gender as an input can still produce outcomes that track those categories closely, through variables that correlate with them without naming them — pin code, employment category, even device type. This is not a hypothetical concern; it is the standard failure mode of algorithmic underwriting, and regulators reviewing digital lending are increasingly unwilling to accept "we don't use protected characteristics" as a complete answer.

The legal exposure here sits at the intersection of RBI's Digital Lending Guidelines, which require documented, explainable credit decisions, and the broader constitutional and statutory anti-discrimination framework that doesn't disappear simply because the discriminating logic runs through a model rather than a human underwriter's stated bias.

A model that never asks someone's caste can still discriminate on caste, if the variables it does use are close enough proxies. Intent was never the legal test for discrimination, and it isn't the test here either.

A defensible position requires more than the absence of protected variables in the model's feature set. It requires an active bias audit examining disparate outcomes across those categories even when they were never inputs — testing the model's actual behavior, not just its stated architecture — and a documented remediation process for any disparity that audit surfaces.

This is precisely the kind of documentation that satisfies a regulator under real time pressure: not a claim that the model is fair, but evidence that its actual outcomes were tested against the categories that matter, with a record of what was found and what was done about it.