Who captures the productivity gains from AI?

A technology can raise output and lower wages at the same time. That is not a paradox or a pessimist's talking point; it falls straight out of the standard framework. Acemoglu and Restrepo show that automation shifts the task content of production against labour, and therefore always reduces labour's share of value added, and may reduce labour demand even as it raises productivity.

Which means "AI makes us more productive" and "AI makes us better off" are two claims, not one. Almost every argument in this field answers the first and quietly presents it as the second.

Two effects, pulling opposite ways

The framework is simple enough to hold in your head. Production is a set of tasks allocated between capital and labour.

Automation lets capital take over tasks labour was doing. That is the displacement effect, and it always moves the share of value added away from labour. Against it runs the reinstatement effect: new tasks get created where labour has a comparative advantage, and that always moves the share back.

Their decomposition of US industry data attributes three decades of slower employment growth to an accelerating displacement effect, a weaker reinstatement effect, and slower productivity growth than in earlier periods. Note what that means. The problem was not that automation happened. It was that displacement outran the creation of new work, and the productivity gains that were supposed to justify it came in smaller than expected.

The estimates disagree by an order of magnitude

Acemoglu's macroeconomic estimate puts total factor productivity gains from AI at under 0.53 per cent over ten years. That is a long way below the forecasts that circulate in company announcements, and it comes from working the task-level effects through to the aggregate rather than extrapolating from demonstrations.

Task-level studies find much bigger numbers. Brynjolfsson, Li and Raymond measured roughly 14 per cent higher productivity among customer support agents. Dell'Acqua and colleagues found large effects inside the frontier of what the tool does well, and negative ones outside it. The METR developer study found experienced open-source developers were slower with AI assistance while believing they had been faster.

These can all be true together. Large gains on particular tasks do not aggregate cleanly into economy-wide productivity, because most work is not the task that was measured, and because the gains have to survive contact with everything else an organisation does. The gap between the task studies and the macro estimate is the most interesting number in this area. Almost nobody discusses it.

Who the measured gains actually reached

Where studies do find gains, a consistent pattern shows up: they concentrate among the least experienced workers. Brynjolfsson, Li and Raymond found exactly this, and identified the mechanism: the system disseminated what the best performers already knew to everyone else.

That sounds unambiguously good, and half of it is. The other half is that compressing the distance between a novice and an expert raises average output while reducing what expertise is worth. If a year-one employee now performs like a year-five employee, the market rate for four years of experience has changed. That is a distributional shift inside the workforce, running alongside the shift between labour and capital, and it points at the same problem as the missing rungs.

It also raises a question the productivity numbers cannot answer: if the expertise is in the tool, who was building the next generation of experts? See capability debt.

Why this is a choice and not a forecast

Nothing in the framework makes any of this inevitable. That is the part worth insisting on. Displacement reduces the labour share. New tasks raise it. Which dominates depends on what firms decide to build and what the surrounding rules reward.

Acemoglu's own conclusion is that better wage and inequality outcomes depend on creating new tasks for middle and low-pay workers specifically, rather than on automation getting cheaper. Autor's argument for rebuilding middle-class work runs on the same logic from the other direction: the useful question is whether the technology extends what a person can do or removes the need for them to do it. That gets decided repeatedly, in ordinary meetings, by people who mostly do not think of themselves as deciding it.

Which is why "the technology will decide" is the least accurate thing said about this subject. See design versus drift.

What to watch instead of the productivity number

• Ask where the saved time went. If a tool saves an hour and the hour is absorbed into more output at the same headcount, the gain has been captured before it reaches anyone doing the work.
• Watch the labour share, not the output figure. Output can rise while labour's portion of it falls. Reporting only the first is how the two questions get merged.
• Count new tasks. The reinstatement effect is the only thing that reliably pushes the share back. If nothing new is being created for people to do, the direction is already decided.
• Distinguish gains that reach novices from gains that reach the firm. They are frequently the same measurement described two ways.
• Treat aggregate forecasts with the caution their spread deserves. When credible estimates differ by an order of magnitude, confident planning against either end is not evidence-led.

What this page does not claim

It does not claim AI will reduce wages. The empirical work in the task-based tradition concerns industrial automation and robotics, and predates generative AI; the framework is well established, the application to this technology is not yet settled by data.

It does not claim the productivity gains are illusory. Several are well measured. The argument is narrower and, I think, harder to dismiss: measuring a gain tells you nothing about who ends up holding it, and the second question has an answer that the first cannot supply.

Related SuperSkills research

On the employment side of the same question, will AI replace my job and will AI replace entry-level jobs. On what the gains do to career structure, the missing rungs and capability debt. On the gap between adoption and benefit, usage theatre. On what individuals can control, staying valuable in the age of AI. On the organisational choice, design versus drift.

Key research and primary sources

• Acemoglu, D. and Restrepo, P. (2019). Automation and New Tasks: How Technology Displaces and Reinstates Labor. Journal of Economic Perspectives, 33(2).
• Acemoglu, D. (2024). The Simple Macroeconomics of AI.
• Autor, D. (2024). Applying AI to Rebuild Middle Class Jobs.
• Brynjolfsson, E., Li, D. and Raymond, L. (2023). Generative AI at Work.
• Dell'Acqua, F. et al. (2023). Navigating the Jagged Technological Frontier.
• METR (2025). Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity.

About this research

Rahim Hirji is the author of SuperSkills (Kogan Page, 2026) and founder of The SuperSkills Intelligence Company. The task-based framework is applied here to a technology its empirical work predates, and that limit is stated rather than glossed. Estimates that differ by an order of magnitude are reported as differing rather than averaged. Reviewed quarterly.