OpenAI absolutely does not have licenses for 99%+ of the content used to build their models. They're following the standard tech company model of "negotiate forgiveness rather than ask for permission".
The legal status of copyright in this context is unclear, because the laws were written without imitation machines in mind. Presumably they blew right on by "is this right?" and went for "can we get away with this and pay less in fines than we'd make in profits?" As long as there's a 51% chance of that being the case, they saw it as worth a go.