As for your point 2), we already have humans in the loop of other humans and this commonly leads to the "why do companies have so many people in them" trope we see posted on HN commonly. I think in that sense you may be underestimating how incredibly stupid people can be too. This said, AI doesn't remove all humans from the loop, but may significantly reduce how many are in said loop.
As someone who has worked at companies buying the databases that come from manually reading papers, the quality of them is hugely variable due to this. Lots of manual errors, missing context, etc.
Plenty of drug activity databases don’t distinguish between assays in solution, cell-based, and in vivo. When these produce hugely different values for inhibition due to metabolism and transport in biological environments.
GPT4 opens the possibilities for the true experts to focus on prompt engineering to extract these details, developing mechanisms to test performance, think deeply on how the data is stored and organized, etc.