Our data wasn't actually "user data" in the sense Google usually deals with. It wasn't data collected incidentally after click-through consent from billions of random people on the internet as they use their computers in daily life. It was ~100 people, many of them employees, who voluntarily participated in a one hour in-person data collection session after signing ink-on-paper consent forms, who received monetary compensation for the use of their data, and the data was used solely for training and evaluating models and not cross-linked with any other data for any other purpose. But Google's privacy bureaucracy wanted to apply the same processes and standards as user data collected continuously from billions of internet users.
But of course ultimately it didn't matter. The privacy bureaucracy issues were not at all related to the division-wide strategy pivot that killed our team (and many others). It just made my life very frustrating for the year or so before that happened. And I understand why the bureaucracy exists. In today's climate the PR risk to Google from a hit piece headline like "Google scans your eyeballs and we have the leaked data" is much higher than the probable benefit from a small team's engineering work. So they err on the side of slowing things way down. But that doesn't make it any less frustrating for that small team. And it makes me quite pessimistic about the future development of new technology at Google. I expect that their continuing failure to deploy AI anywhere near as good as GPT-4 can be attributed to similar locally rational risk-averse bureaucracy...