You literally don't understand what you're talking about. It's not a database, no copies are being made. The weights of digital synapses are adjusted in a neural network, the same as happens in our brains when we learn something. For all intents and purposes it is indeed learning from an example, not making a copy. There's no "auto-complete" involved, anymore than there is when a human creates something using something else as inspiration.
No copies are being made, no money is being made by something being sold. You can download the models and use them for free. You are only paying for access to the hardware to host the neural network, not the model itself. Again, nothing has been stolen, and no copies have been made. It's learning from data, same as people do from open source projects.
If there was a license that said you are not allowed to learn from reading the code, then you might have a point, but there are none of those that exist to my knowledge. If you weren't meant to learn from them then they wouldn't be publicly available.