The problem with measuring AI productivity is that the people doing the actual job (paralegals, developers, etc.) are doing it for someone else (judges, managers, etc.). More work, or even a speedup does not actually benefit them. So when you give professionals a tool that speeds them up, they increase their slack and/or focus on other, less productive activities rather than work more.

The article captures this too, mentioning a couple of examples of startups where presumably this feedback loop is tighter.