I'd advocate for a hard line on this.
Something that's utter random garbage is actually less harmful thatthan almost correct garbage that needs an expert to work out.
If you're using the output of a MLmachine learning or AI tool WITHOUT verification and/or disclosure, and we're going to have to waste time working this out, the user clearly isn't here with the right intentions. As such I'd be inclined to treat the user as a severely as needed, and I'd be tempted to start with longer suspensionsuspensions.
ItsIt's a good way to prevent folks who are clearly wastewasting people's time from wasting more of it.