Depends on your definition of "create", but Beatrix also creates :) It's not just about link sharing - sign up and choose (for example) the Technology category and you'll see original pieces of content as well as link sharing in your content feed.

Given that rewriting a text using AI is not that hard, I wonder. At what point does it stop being copying and start being creating ?

I sort of wonder about stuff like that. If you look at what we know about the algorithm that "is" the human mind, it's not actually capable of creating. So nothing is created. It's copies of other things, usually mixed together.

Most commercial AI products these days avoid patent issues by using AI. Every AI program that isolates and OCR's text uses dozens of patent pending techniques. It's just that this is not visible in the actual code. The reason that it's not in there is that that code encodes something akin to a VM, and the real program is the training results. It is generally very hard to determine what exactly the program does, but for trivial patents that (e.g. energy-lines to separate individual letters, or tracing likely pen movements and recognizing the derivative) is something these algorithms can be shown to do. Yet as far as I know, no-one's been successfully sued.

The algorithms themselves have the advantage that most are quite old, and have obvious roots in the 60s and 70s. So they are not patented in the US (effectively). Is this the perfect way around software patents ?

I wonder. Copyright-wise. If I use an algorithm analyse all text from an author and then have it produce "his next book", would that be legal ?