> the algorithm only knows how to output the things that it has already outputted

That's a very old problem that people building recommendation systems solved 10 years ago.