> Once adopted this will lead to an increase in randomly locked out accounts due to model false positives

I would assume the model was also trained on negatives that look like CSAM but aren’t.