No, it's impossible. AI-based algorithms are fairly bad at recognizing copyrighted works. It's hard to strike a balance between what is an original composition and what's a remix/cover/copy:

- if you make it too strict you'll just find bit-by-bit copies: reencoding or slightly altering the data would pass screening easily

- if you make it too lax, you will find that many "original compositions" are in fact very similar to or inspired by other works (oh really that bird singing i just recorded is copyrighted?!)

Beyond that, when there's an actual copy involved, an algorithm is fairly bad at determining whether this copy is legal:

- if it looks very similar to another user's video which claims to have copyright on the content, however what proves the copyright is owned by that other uploader and not by a 3rd party? how can you make sure the license was not copyleft or public domain in the first place?

- how do you know whether the content is used in a context of lawful criticism? parody?

- how do you know whether the usage is tied to an education or accessibility exception to copyright?

- how do you know if the usage of the copyrighted content is legal in the uploader's country? in the copyright owner's country? in the host's country? how do you decide which is more important?

If censorship is bad, automated censorship is even worse. There's no way to make it right.

If you can build it and prove this to be true, you'll likely become a multimillionaire when you get acquired.