I think if Google trained current models on private data, confidential info would leak constantly, it would be an absolute trainwreck. If Gemini leaked your Gmail and Chrome activity, Google would get sued and regulated into oblivion.

But Google needs to leave this option open in the future, in case they have to go all-in on an arms-race against China, if Chinese AI starts becoming an actual threat somehow. And it's easy to predict the USA gov would prioritize that race over privacy concerns.