Comment by fittingopposite

> While unlikely, we cannot rule out that de-identified data derived from their usage of our products helped improve our models .

Wow. That sounds like an admission of guilt.