Basically impossible, they’re inherently biased by the data they learn from but once they’re done training there you can’t tinker with the system itself, only try to impose external constraints that have unlimited potential to be breached