“Researchers then used human annotators and a GPT-4 model to rate whether access to a ChatGPT-powered system would cut down the time required for a human to perform a specific task “by at least 50%.””
OpenAI studied how much of the US workforce its own product exposes and found 19 percent of jobs with half their tasks in reach. GPT-4 helped grade the tasks. The study ignores that the model makes things up and needs a human checking its work. A company asking its product how important its product is will always get a big number.