It is hard not to see “fragile white people” as a bias. Look at these comments around you. The more typical HN lens of trying to understand the technical causes is overcome by cultural posturing and positioning. If I had to guess, either the training set was incorrectly tagged like with a simpler model creating mislabeled meta data, or a deliberate test was forked to production. Sometimes you run tests with extreme variables to validate XYZ and then learnings are used without sending to prod. But what do I know as a PM in big tech who works on public facing products where no one ever has DEI concerns. No DEI concerns because not everything is a culture war like the media or internet folks will have you believe. Edit: not at Google