Brookings: AI's productivity boom borrows against future experts

A Brookings essay by Niam Yaraghi, a nonresident senior fellow in the institution's governance studies program, argues the AI productivity story most executives are telling themselves has a hole in it. The gains are real, but they are being paid for out of an expertise inheritance that took decades to build and that current deployment patterns are not replenishing.

The evidence he assembles is worth taking seriously because it points in the same direction from different angles. Research by Brynjolfsson, Li, and Raymond on 5,000 customer service agents found roughly 34% productivity gains for novices and minimal effects for experienced workers, which fits the intuition that AI compresses the easy part of the learning curve. A separate study of BCG consultants found a 19 percentage point performance drop when the same tool was used on unfamiliar tasks that sat outside its training data, what the researchers called the jagged technological frontier. AI is a lever where you already have judgment and a trap where you do not.

The uncomfortable part is what happens to the pipeline that produces that judgment. Yaraghi cites work by Brynjolfsson, Chandar, and Chen finding a roughly 16% relative employment decline for workers aged 22-25 in AI-exposed occupations after 2023, while senior employment held stable. A separate analysis by Hosseini and Lichtinger, covering 65 million workers across 280,000 firms, found AI-adopting companies cut junior hiring while maintaining senior staff. That is an individually rational hiring pattern producing a coordination failure at the market level, since firms cannot later buy back seniority they never trained.

He also flags a diversity effect on the output itself. A Science Advances study by Doshi and Hauser reportedly found writers using generative AI produced individually more creative stories but with collective diversity falling by roughly 10%, which he links to Kuhn's distinction between normal science and paradigm shifts.

The honest caveat is that this is an argument built from a handful of early studies, most of them post-2023, and the essay does not tell you which specific industries will crack first or whether the junior hiring drop reverses as tools mature. What the reporting does not give you is a working example of an AI deployed as a genuine pedagogical tutor rather than an answer engine. The opening for whoever builds that is the interesting part.