Does AI's wage-equalizing effect actually reverse as it improves?

Agrawal, Gans and Goldfarb's July 2025 NBER working paper (w34034, not yet peer-reviewed) — the same model described in How do cognitive tools reshape the three types of worker skill? — carries a further, more specific prediction beyond explaining today's data: that AI's current inequality-reducing effect is temporary. The conclusion states the model "predicts that this equalising effect may reverse as AI tools become more sophisticated, eventually favouring workers with superior judgment capabilities and creating a U-shaped pattern in wage inequality."

The mechanism follows from the paper's distinction between implementation skill, which cognitive tools substitute for, and judgment, which tools complement. Early, less capable AI primarily replaces weak implementation skill, so workers who previously lacked that skill gain the most — the inequality-reducing effect the paper cites in call centers (Brynjolfsson et al., 2025) and coding (Cui et al., 2025). But the same logic that makes opportunity judgment "always complement cognitive tools" implies that as AI gets better, the returns to judgment — recognizing when and how to act on an AI's output — grow alongside it. Past some point, the authors argue, "the benefit of responsive human judgment rises as AI improves," so the workers who gain most shift from those with weak implementation skill to those with strong judgment, flipping the direction of the inequality effect and producing a U-shape over time rather than a one-directional trend.

This complicates any reading of present-day equalizing findings as a stable description of AI's distributional effect. Can AI narrow the education performance gap? and When does AI actually boost worker productivity? describe snapshots consistent with this model's early phase; under this prediction, those snapshots are one point on a trajectory the authors expect to reverse, not an endpoint. The prediction also qualifies What makes accountable judgment scarce when AI cognition is cheap? by giving a capability-driven reason, not only an institutional one, for why judgment becomes the scarce asset: judgment's complementarity with cognitive tools strengthens as the tools themselves improve.

The U-shape is a theoretical prediction derived from the model's assumptions about the relative variance of judgment and implementation skill, not an observed trend — the paper cites no longitudinal data showing inequality already reversing direction within an AI-adopting workforce, and its working-paper status means the claim has not been peer-reviewed. If the prediction holds, it implies that policies or workplace designs calibrated to today's AI-driven equalization, such as using AI to close skill gaps, may need to anticipate a later shift back toward rewarding judgment, rather than treating current equalization as a stable gain.