Like, a startup my employer helped with is using AI to read biopsies and other medical imaging faster and more accurately than ever. The dataset it is trained on is enormous. Millions of images tagged only with the diagnostic information, all de-identified. A pathologist still has to sign off on it, and that pathologist is the point of accountability.
The trouble with this approach is the same issue you get with nearly self-driving cars and llm-assisted code: at a certain point, the tool is good enough that the person just signs off on the results, without checking it. In the same way you can’t expect someone to drive for 6 hours just staring at the road without eventually zoning out, you can’t expect a doctor to keep seeing accurate results without eventually zoning out.
That’s not to say there isn’t a place for these, but I don’t know how you dodge the “zoning-out” problem.
The trouble with this approach is the same issue you get with nearly self-driving cars and llm-assisted code: at a certain point, the tool is good enough that the person just signs off on the results, without checking it. In the same way you can’t expect someone to drive for 6 hours just staring at the road without eventually zoning out, you can’t expect a doctor to keep seeing accurate results without eventually zoning out.
That’s not to say there isn’t a place for these, but I don’t know how you dodge the “zoning-out” problem.