I think the cause might be that the output of the AR is not consistent. Not that each cycle of the loop wouldn’t sound the same to a human, but it’s not a sample accurate reproduction, like a sampled loop is.