He’s not. He’s saying there’s an implied license to read(and by extension manipulate with a computer in order to view) what’s been published openly. And he’s right. But there’s not an implied license to copy and republish, which is what training models on that data does.