Daron Acemoglu estimates that AI will generate only a 0.66% productivity gain globally over ten years, roughly 0.05% per year—far from projections by major investment banks. Acemoglu anticipates a possible widening of the gap between capital and labor income, without demonstrating a general mechanism grounded in initial resources. The political question is who pockets the micro-level gains and to what extent the choices around AI deployment remain open.

The Essentials

  • Only 5% of tasks are profitably automatable in the short term, according to Acemoglu (Economic Policy, 2025).
  • Expected global productivity gain is around 0.05%/year over ten years, far from tech optimists’ forecasts.
  • Microeconomic gains exist but go primarily to companies and workers already well-equipped.
  • The direction of deployment—automation versus augmentation of human capacity—is an institutional choice, not a technical constraint.
  • The fork in the road is happening now: regulatory framework, public R&D, and data rights are the available levers before 2030.

The Number That Bothers Both Sides

In June 2026, Acemoglu repeated it in an interview with Fortune: current AI models are impressive, but their ability to profitably replace entire tasks remains narrow. Only 5% of tasks currently entrusted to humans meet the criterion of immediate automation profitability. The rest requires either a quality of model we don’t yet have or integration costs that erase the benefit.

This figure irritates both sides of the debate. Technopessimists announcing the end of salaried work must revise their timeline. Tech optimists promising productivity gains comparable to electrification or the internet must explain why Fortune cites roughly 1% to 1.5% GDP growth over ten years—a comparison with steam’s contribution to historical growth remaining indicative for lack of directly comparable measures.

Acemoglu and Restrepo had previously described innovation as oriented toward automation rather than task creation. Some major technology companies have favored models based on automation and labor cost reduction rather than widely distributed productivity gains. Result: gains exist, but AI could widen the gap between capital and labor income, with a distribution of gains that remains largely concentrated. On this mechanism, the data align with the analysis of wage bifurcation already documented here.

Automation or Augmentation: A False Technical Inevitability

The key word in Acemoglu’s thesis is “direction.” AI can deploy in two radically different ways. It can replace human tasks—the automation model—or it can make workers more capable at complex tasks—the augmentation model. These two trajectories coexist technically from today. What distinguishes them is a combination of tax incentives, market structures, and design choices.

Current American tax policy leans the wrong way. An employer who replaces a salaried worker with software doesn’t pay social contributions on that software. They amortize the investment faster. They reduce the fixed costs of human resource management. Economic incentives point toward automation even when augmentation would be more productive over the long term.

The bias toward automation stems from the effective tax structure favoring capital over labor.

Proponents of market optimism, Tyler Cowen notably, argue that the history of technology has proven pessimists wrong at each wave. Automating repetitive tasks has always freed human capacity for new activities, often better paid. Electronic cash registers multiplied the number of cashiers in the 1980s by making banks more profitable and thus more numerous. AI could follow the same trajectory.

The argument is solid, but Acemoglu brings a precise limit to it. Previous technological waves created new tasks alongside destroyed ones. Generative AI, conversely, can directly compete with non-routine cognitive tasks—writing, analysis, coding, diagnosis—which until now constituted the refuge of workers displaced by industrial automation. If AI erases both repetitive and cognitive tasks, there is no obvious third refuge. The historical comparison holds up to a point, then it snags.

Who Wins, Who Loses, and Why It’s Being Decided Now

American data from 2025-2026 draw a geography of winners. Tech sector companies, strategy consulting firms, large law firms: all these actors signal real productivity gains among their most qualified teams. Goldman Sachs estimates that its analysts process twice as much data per hour since integrating AI tools into their workflows. These gains are real. But they accumulate where teams already had the skills to exploit the tool.

A Midwest manufacturing SME with no data science department, no continuing training budget, no access to the most performant proprietary models: this SME doesn’t capture the same gains. It may even face increased competitive pressure from large companies that did capture those gains. The productivity gap between large and small companies, already documented before AI, risks widening. On actual adoption in smaller structures, French data on SMEs confirm this gap between declaration and actual use.

The November 2025 ICRIER event included a panel on AI and public policy, but the official source reports no quantified conclusion on comparative gains for data holders. Data is AI’s productive asset. Those who own it—platforms, large companies, states with digitized administrations—are also those who orient model development. Those who don’t own it undergo models designed elsewhere, for other purposes.

This mechanism is both unequal and correctable. Data ownership, workers’ rights over data they generate while working, pooled access to models in public sectors: all of this involves policy choices already under discussion in several legislatures. The European Union opened this project with its data regulation. The United States debates it less structurally but genuinely.

Decisions Still Available to Institutions

AI’s regulatory framework is being built now, according to Acemoglu, within a limited time window. This early-lock-in mechanism is well documented in the literature on digital platforms. This window is what Acemoglu emphasizes in his recent public interventions.

Three levers are available and actively discussed. The first is fiscal: taxing automation and augmentation differently. An employer investing in AI to train their teams, increase their productivity, and maintain employment deserves different tax treatment from one automating to eliminate positions. This distinction is difficult to implement in practice; the boundaries are blurry, but not impossible. France experimented with tax credits conditional on job preservation.

The United States has precedent in the research and development tax credit.

The second lever is public R&D. Current foundation models are developed by a handful of private laboratories with commercial objectives. A public investment in pro-work-oriented models, designed to augment human capacity rather than replace it, would change the technological base available. The United States funded the foundations of GPS, the internet, and modern biotechnology before the private sector took over: such an approach has already produced concrete results.

The third lever is data rights. If workers and communities had explicit rights over the data they generate, the power structure in AI development would shift. American labor unions in film and journalism negotiated clauses on using their productions in model training. It’s primitive, but it’s a start.

The 2030 Fork: Two Trajectories for the Same Tools

Projecting ten years remains uncertain, and Acemoglu is the first to say so. But identifying the conditions for each trajectory is possible right now, from already-visible signals.

In the first scenario, AI deployment remains driven by cost-reduction logic. Productivity gains would concentrate on capital holders and highly skilled workers. Wage inequality could increase. Low-wage service jobs persist because they cost less than a robot, but they’re squeezed from above and below: by machines above and wage deflation beneath. The democratic legitimacy of technology policies weakens.

The question of what institutions can still decide becomes pressing at the same rate inequalities accumulate.

In the second scenario, public institutions reorient deployment toward human complementarity. Productivity gains would distribute more broadly, while continuing education would receive funding commensurate with the stakes.

The Fortune range of roughly 1% to 1.5% GDP over ten years approaches its ceiling rather than its floor.

What separates these two trajectories is not technology, but the full set of policies, institutions, and deployment choices.

The Signal to Watch: Where AI R&D Goes by 2027

An indicator will allow us to read which trajectory is prevailing before macro data becomes available. Watch where AI R&D money goes, not just its volume. If the share of investments in augmentation applications—medical tools for field practitioners, educational assistants for teachers, translation interfaces for multilingual workers—increases relative to pure automation investments, that’s a bifurcation signal. If it stagnates or declines, the first scenario’s trajectory confirms.

Current signals are mixed. Microsoft and Google invest massively in professional augmentation Copilots. Simultaneously, announcements of AI-linked job cuts in call centers, content moderation, and administrative services accumulate. Both dynamics coexist. Regulation, tax policy, competition, and R&D choices can influence AI’s trajectory.

Acemoglu emphasizes the importance of the current period for policy choices. As market concentration advances, the capacity to reorient deployment from outside diminishes. The European Union was right to legislate on AI before that concentration was complete. The United States will have to decide: let it happen and hope the market corrects, or intervene while intervention still has leverage.

These institutional choices, jointly with technology, will influence the distribution of benefits from productivity gains.

Sources

  1. Daron Acemoglu, “The Simple Macroeconomics of AI,” Economic Policy, vol. 40, no. 121, January 2025, https://academic.oup.com/economicpolicy/article-abstract/40/121/13/7728473
  2. Fortune, interview with Daron Acemoglu, June 21, 2026 (available on fortune.com)
  3. MIT Technology Review, analysis of Acemoglu’s work on AI and productivity, March 2025 (available on technologyreview.com)
  4. ICRIER, panel “AI & Policy Choices,” November 4, 2025 (available on icrier.org)