It might work, as indicated by the results in your link ("Human raters are only slightly better than random chance at distinguishing short clips of the game from clips of the simulation."), but the result would be a horrific dystopian nightmare, so why would we do this to ourselves?
Anyway, there is one aspect where the STEPS work is similar to this idea, in that it tries to build a more concise model of the system. But it does this using domain-specific languages rather than lossy blobs of model weights, so the result is (ideally) the complete opposite of what you proposed: A less blobby, more transparent and more comprehensible expression of what the system does.