Well, there's the Einstein test: can AI figure out general relativity if it's trained only on knowledge up to 1915 or so, before it was discovered. Similarly, you could do a Monet test: train AI on everything before Monet and try to get it to create paintings similar to Monet.

Original is something that is out of the data distribution. AI can't do anything original, because it's job is to imitate the data distribution.

Originality in itself is not hard, because pure noise is original. It should be original and beautiful.