You may be able to deduce that percentage from the percentage of humans who make art, and the percentage of art that contains original elements. (Whether it's aesthetically pleasing has no bearing on whether it is art).
All art is derivative to some extent, because all artists have absorbed cultural influences and have seen prior art. But some art contains elements and ideas which are not synthesized from prior art. You can prove this. If art were only synthesized from prior art, then there would never be any addition to its vocabulary. There are conceptual "breakthroughs" which cannot happen just by looking at and iterating upon existing art.
If an AI had been trained only on classical Greek sculpture, it could not invent Cubism or Impressionism or Surrealism. Not just that: It would have no reason to invent these schools of art. The only impetus it might have would be if a human asked it to invent a school of art; and then it could only draw upon its training data.
That's why to call AI output "art" is to fundamentally misunderstand what art is. Art is not the final result or product. An aesthetically pleasing painting is not automatically art, outside the limited commercial sense. Art is the intention of the artist and the unique characteristics of the artist made manifest in the creative process which required discovering something new. The actual output, the thing on the canvas, is just evidence of that process, it is not the art itself.
More often than not, this is also a physical process involving trial and error with real materials in a world that is many orders of magnitude more complicated than what AI currently understands.
An equation on a blackboard is not a mathematical proof, it is the residue of the logic of the proof. In the same way, a painting is the residue of art. A sculpture is a residue of the artistic process by which a person learned to turn a shapeless mass into an imagined 3D object.
This is why AI can only make simulacrum of the final result of art, the same way it can simulate coming up with a proof for an algorithm. But as we frequently see when we ask it to analyze or create an algorithm, it cannot provide a true proof, because it cannot think of failure modes unless we explicitly point them out, nor can it think of concepts that are not in its existing canon of knowledge.
Maybe with AGI this will change. But passing a Turing test and making pictures doesn't mean it can actually create anything like art.