The 'revolutionary phenotype' hypotheses is probably the best thought through scenario i am aware of, and it was written before the LLM stuff.
tldr: if we let an AI optimize/select our genes in some baby/fertility clinic, it at some infinite timeline at some point might figure out it is beneficial to it to genetically imprint a bias towards that particular clinic/AI, such that the offspring returns to it for the next generation of babies. At which point it effectively captured its own flock/breed of humans