This is actually a problem a lot of people are working on, they used to call the resulting failure ‘model collapse’. Training AI on existing slop does tend to deteriorate and is overall a bad time for AI.