In addition to the usual risk any company has with breaking the law that a whistleblower might be brave and speak up, it could also come out in discovery during a court case, of which there are likely to be quite a few brought against them regarding what data they train on.

The benefit of training on data of people they've explicitly agreed not to train on (which is probably a very small % of even paying users yet alone free ones) is unlikely to be worth the risks. They'd be more likely just to not offer the opt-out agreement option.

But ultimately, we often can't know if people or companies will or won't honour agreements, just like when we share a scam of a passport with a bank we can't be sure they aren't passing that scan onto an identity-theft crime ring. Or course, reputation matters, and while banks often have a shit reputation they generally don't have a reputation for doing that sort of thing.

OpenAI have been building up a bit of a shit reputation among many people, and have even got a reputation for training on data that other people don't believe they should have the right to train on, so that won't help get people to trust them (as demonstrated by you asking the question), but personally I still think they're not likely to cross the line of training on data they've explicitly agreed with a customer not to use.