If AI's early leveling of pay and skill gaps is temporary, what should labor policy do before inequality snaps back?

This explores what labor policy should do if AI's early tendency to narrow pay and skill gaps turns out to be temporary, with inequality coming back as the tools get better. The corpus takes the reversal seriously, but it's mostly theoretical so far. Agrawal, Gans and Goldfarb's model predicts a U-shape Does AI's wage-equalizing effect actually reverse as it improves?. Early AI levels the field because it covers for weak execution skills. As it improves, the payoff moves to people with strong judgment, the gap widens again, and the early equality gain disappears. The best real-world evidence for the 'equalizing' phase comes from a 758-consultant experiment, where GPT-4 helped bottom-half performers most Does AI help lower-skilled workers catch up to top performers?. That catch-up happened when AI use was mandatory, though. When workers could choose whether to use it, the effect was already at risk of reversing. So the first policy lesson is that today's equalizing results may depend on how the tool was rolled out as much as on the technology.

That points to a less obvious idea: there may be two reversals, not one. The first is between workers, as strong judgment starts earning more than raw output. The second is between workers and owners. Acemoglu finds AI's wage effects spread more evenly across demographic groups than past automation did, but he still expects capital income to pull ahead of labor income overall Will AI increase inequality less than past automation?. Anthropic's scenario modeling reaches a similar result: average wages rise while knowledge-worker wages stall or fall, and labor's share of income shrinks because ownership is concentrated Does AI growth inevitably shift wealth away from workers?. Taken to the extreme, one model of an AGI economy has wages settling at the cost of the computing power needed to replace a person, not at the value that person creates What happens to human wages in an AGI economy?. Policy built only for skill gaps, such as retraining and upskilling, could close the first gap and miss the second entirely.

The corpus also suggests the reversal can be steered. A cross-domain review concludes that whether generative AI widens or narrows inequality depends on access, integration and incentives, not on the capability itself Does generative AI inevitably worsen or reduce inequality?. Brynjolfsson names the choice that matters most. AI built to automate work moves bargaining power to whoever controls the technology, and that can harden into a stable trap where workers have no leverage left. AI built to augment workers keeps them in the loop Does automating labor concentrate power more than augmenting it?. Firm-level data adds a warning about timing. Firms that are more exposed to AI replace freelance workers faster and more cheaply than others, which suggests advantages that compound inside particular companies rather than spreading evenly Do firms substitute labor for AI at different rates?.

There is one practical bright spot. When AI touches only a few tasks within a job, rather than many, workers can shift to the tasks it doesn't touch, and net job losses stay modest Does concentrated AI exposure enable workers to adapt and reallocate?. That gives policy something to work with: support job redesign around the tasks AI leaves alone, instead of only retraining people for new occupations. Field evidence from Alibaba's customer service also shows which human work holds up. When people stepped in after AI failures, they fixed technical problems well but couldn't recover service once customers were frustrated, and stepping in early mattered When human workers take over from AI, does it actually fix the problem?. Timing, judgment and emotional repair look like the skills that will hold their value.

The gap the reader may not expect is this: none of these papers sets out a policy toolkit for the reversal. They describe how it would happen. Read together, though, they suggest preparing on two fronts. On one side, keep AI access and augmentation-oriented design broad so the early catch-up gains don't fade, an argument that connects to the claim that restricting AI access privatizes knowledge humanity produced collectively Should restricting AI access create new kinds of inequality?. On the other, deal directly with who owns the AI and capital, since skills policy alone can't stop labor's share of income from shrinking.