The productivity dashboard cannot see the cost MIT measured

The distinction the MIT finding rests on is not that people hand work to tools. Calculators and search engines already did that ^[4]. It is that a generative model returns a conclusion rather than raw material for someone to work through ^[13], so the step being skipped is the one the tool was nominally supporting: reasoning, memory, imagination and judgment ^[14].

That difference is what makes the throughput figures awkward rather than wrong. A 25% cut in email time and 26% more completed tasks ^[12] are both counted while the assistance is switched on. The MIT deficit surfaces only after it is switched off ^[1]. An organisation running the usual dashboard therefore collects the first number continuously and the second one never, which is not a measurement problem so much as a sampling one: nobody schedules the unaided test.

Rationing adoption by task type ^[5] is a blunter instrument than it sounds, and that is the argument for it. It does not require detecting cognitive debt in a named individual, which Kapoor concedes remains hard to establish conclusively ^[7]. It requires a prior decision about which tasks exist to produce output and which exist to build the person doing them. Education is the clean case in Kapoor's telling, on the grounds that its purpose is developing the brain in the first place ^[6], with original research alongside it ^[5].

The workplace version is messier. Kapoor's claim is that AI amplifies weak judgment as efficiently as strong judgment ^[15], and that a workforce can stay highly productive while losing the expertise to challenge an output, spot a flawed assumption or decide under unfamiliar conditions ^[16]. If that is right, the exposure sits with the firms that rolled out tools fastest to the juniors who had the least independent practice to lose.

Worth naming what is carrying all this. One EEG study of essay writers ^[1], and Kapoor's own observation of a decline of roughly 40 to 50 points on a 1,000-point brain training scale after using AI for reasoning ^[3], which is four to five percent ^[11], self-measured, one subject. Kapoor is a researcher and strategist at major multinational banks rather than the study's author ^[9]. His personal discipline, think before prompting, and do not outsource the first question or the final decision ^[8], costs nothing and survives thin evidence intact. A policy that restricts tool access by task category costs something, and that is the version that will need more than one paper before anyone signs it.