besides that, he's also engaging in the logical fallacy of using the human as a more trustworthy replacement of the AI, when the verification of human understanding has the same fundamental problems you would have when verifying AI output and understanding.
i.e. there need to be some meta processes that don't rely on human understanding at all.
in medicine the whole pipeline is optimized exactly for this for short and medium term effects. for long term effects, there honestly isn't a pipeline, and the general mechanical understanding of treatments potential long term effects is surprisingly poor / uncertain when you look into it. and in this case, AI would almost certainly just improve the process.
i.e. there need to be some meta processes that don't rely on human understanding at all.
in medicine the whole pipeline is optimized exactly for this for short and medium term effects. for long term effects, there honestly isn't a pipeline, and the general mechanical understanding of treatments potential long term effects is surprisingly poor / uncertain when you look into it. and in this case, AI would almost certainly just improve the process.