The recent post about Candy Japan's attempt to combat credit card fraud gives a good sense of what fighting fraud and abuse is like:
https://news.ycombinator.com/item?id=11431881As an email service provider it's like that and more, since (1) once an abuser uses your service, they've gotten the benefit immediately and keep it even if their account is discovered as fraudulent later, e.g. stolen CC number & chargeback. (2) Abusive users can directly harm good users such as by harming the deliverability of the overall platform. It's not just bad debt, it's bad experience too. (3) Unlike Candy Japan where fraudsters mostly just wanted to check CC numbers and not actually buy product, email abusers really want to send emails (4) It can be hard to tell good and bad senders apart because some companies with an internet presence aren't email savvy and might make mistakes or might get hacked.
Spam filters are always tough because if you give someone transparency into which actions of theirs that you consider abuse, then they will quickly detect and route around your attempt to block them. (See Candy Japan article) It's pretty easy for a human to guess what might be the sign of their fraud and run a few experiments to see what gets flagged e.g. By comparison a machine learning system might be hard to outsmart, but then it's also challenging to explain and troubleshoot false positives. Hence what's effective is often a combination of machine-learned filters and heuristics along with manual overrides by human judgment.
All other things equal, new users are a lot more likely to engage in fraud than existing ones, and so tend to be under more suspicion. Aside from B2B fraud where companies take out lines of credit and then go bankrupt intentionally, it's uncommon for existing established customers to turn fraudulent - they're already vetted. (Consider: who is more likely to be fraudulent. The first time subscriber to Candy Japan, or a subscriber who has been using it for 12 months and is about to buy their 13th month?) It's not a great experience as a new user to be under suspicion, but if it's temporary and easily overridden by a human it can be a decent trade-off - the need to reach out acts a deterrent to spammers but does not deter legitimate users as much (speaking generally).