1. There exist laws to prevent discrimination against people based on protected attributes 2. ML models make predictions based on attributes without interpretability (it's not possible to prove that protected attributes are not factoring into model predictions) 3. Empirical observation that a model proxies a protected property exposes corporation to liability for regulatory non compliance 4. Therefore any study that could expose bias of a model used in production is to be road blocked or prevented ...
To combat flows like above -- seems like regulators are going to need to update rules with third party audits and an incentive structure that encourages self-regulation and derisks self-detection and self-reporting of non-intentional violations... ideally google should not be put into a position where it is incentivized to police its own ai ethics research to ensure that such research doesn't expose their own illegal/non-compliant activity ...
In this case, there were recent changes to the statute of limitations for CA laws that extended it from a year to 3 years, which could be why this whole process seems weird.
https://www.ebglaw.com/news/ab-9-extends-employees-statute-o....