The "risks" of Machine learning with large models is you won't know who you are even sampling, which exposes you to unknown risk.
What happens with machine learning models is more similar to someone learning to play music: everyone learning some form of musical expression will by default learn music which has been written previously, and this can't but help to influence them to a degree.
Of course, there are systems to police this too, but high profile cases in recent years have shown how imperfect (or difficult?) this is in reality [0].
And it will be much more fair / accurate than the droves of human artists who intentionally omit attributions.
When a human artist composes a new song, they often have no idea where they got the compound basis for it either.
If anything, human attribution is vastly inferior to ML as-is, nevermind of what's to come.