If they are training on inbound e-mails from senders without Google accounts residing in or present at the time of sending in two-party consent states, they are likely in violation of the telephone call recording laws of those states.
In fact I doubt there's fully relevant case law, as I think the case would be that the trained model is the recording device, and it could be demonstrated that verbatim strings from presumed confidential communications are regurgitated by the model when appropriately prompted.
