Let's say that Google has developed a deep learning paradigm to identify -- and ban -- low-value users who violate their ToS. The low-value group would probably include all non-paying users, as well as small G Suite accounts. Let's also say that the paradigm was deployed as an automated solution, since it was 99% accurate during testing at picking out ToS violators, though the exact percentage doesn't really matter. I would imagine that the relevant Google execs did a cost-benefit analysis, and figured that the PR hit and revenue loss associated with a liberal and fully automated use of the banhammer against all supposed ToS violators -- including the 1% that were false positives -- was justified by the benefits to Google's bottom line.

Based on this possibility, it seems likely that Google employees would not _know_ why the banhammer fell in any particular instance -- nor would they care to know, since that would involve digging through the data to figure out which events triggered a positive hit in their machine learning implementation. It presumably wouldn't be worth Google's time and effort on behalf of users already classified as "low-value".

This might also explain why, as mentioned in another post, it took a few days for the banhammer to fall on a start-up, after a dev with a "poisoned" Google account joined the team. Presumably it took those few days worth of traffic and usage data for ToS-Hammer(tm) to figure that the dev had spawned new accounts elsewhere.

Again, I'm not an expert, so pardon any flaws in the relevant logic.