AI Is Going the Wrong Way. A Nobel Economist Explains Why.

Daron Acemoglu, the 2024 Nobel laureate in Economic Sciences, together with David Autor and Simon Johnson, published an NBER working paper in February 2026 titled Building Pro-Worker Artificial Intelligence. The paper opens with a definition: pro-worker technology expands workers’ capabilities and makes their skills and expertise more valuable. The opposite is technology that improves efficiency by making machines run faster or cheaper.

That definition draws the first line in the current AI debate.

Who decides what AI is for?

The AI industry has never publicly declared that “doing the same work with fewer people” is its primary goal. But that goal is hidden in budget approval forms. When a company applies to purchase an AI system, the easiest line item to approve is “reduce labor costs.” If the application says “improve employees’ ability to handle complex tasks,” the finance department will ask: How do you quantify the return? How soon will it materialize? Who is responsible?

Acemoglu and Restrepo’s research shows that the effective tax rate on labor in the United States is between 25.5% and 33.5%, while the effective tax rate on capital is only 5% to 10%. When a company decides whether to hire an employee or buy an automation system, the tax system pushes it toward the latter. Institutional subsidies, not technological inevitability, set the direction of substitution.

Education shows the split. AI can take over lectures, grading, and Q&A, allowing schools to complete existing teaching with fewer teachers. AI can also help teachers identify each student’s knowledge gaps, making personalized teaching that was previously difficult to achieve possible. Both applications share the same underlying model, but which one an organization chooses depends on where the saved resources go. If the saved teaching positions are used to cut budgets, the first application will spread. If they are used to let teachers do more tutoring, the second application has room to grow.

Productivity data has not caught up with model capabilities

Acemoglu’s estimates show that even assuming AI can affect 20% of tasks and bring about a 25% cost reduction, the cumulative increase in total factor productivity over the next ten years will be about 0.66%, equivalent to less than 0.1 percentage point of annual growth. This figure is far below Goldman Sachs’s forecast of 1.5% and McKinsey’s range of 0.5%–3.4%.

Real-world data supports his conservatism. OpenAI’s own 2025 enterprise AI report shows that 75% of users believe AI helps them work faster or produce higher-quality work, but the actual time saved is only 40–60 minutes per day. At the enterprise level, 88% of organizations use AI, but only 28% place employees in positions where they can achieve business impact that changes metrics; among integrated AI pilots, about 5% produce substantial P&L impact.

The gap between lab and enterprise performance comes from task structure. A radiologic technologist needs to handle more than 30 different tasks at work, from recording medical history to organizing mammography files. Human workers can naturally switch among different formats, databases, and work styles. AI needs many separate tools and protocols to complete the same work. Engineers spend a lot of time debugging AI-generated code. Doctors struggle to combine their expertise with AI tools. These phenomena point to the same problem: the coordination costs between tasks are underestimated.

Self-checkout machines are another case. They reduce the need for cashiers but do not move retail productivity in a measurable way. Acemoglu calls this type of technology “so-so automation.” It replaces workers but does not create enough new value to compensate those displaced.

The asymmetry of displacement and reinstatement

Acemoglu and Restrepo proposed the “displacement-reinstatement” framework in 2019. Automation produces a displacement effect: capital replaces tasks performed by labor, and labor’s share of value added falls. The creation of new tasks produces a reinstatement effect: labor gains new areas of comparative advantage, and labor’s share and demand recover.

The two effects are asymmetric. Displacement is visible and executable; it turns into cost savings quickly. Reinstatement requires organizational adjustment, skills training, and process redesign. It is slow to materialize and hard to quantify. Automation can be completed in a few months, while new tasks often take years from emergence to large-scale employment.

The ATM case illustrates reinstatement. After ATMs became widespread, bank tellers were not eliminated. Because the cost of opening branches fell, banks hired more tellers and shifted them to customer service. But the conditions for this case to hold are: banks have an incentive to open more branches, tellers can be retrained, and new tasks are created fast enough. If these conditions are missing, reinstatement will not happen automatically.

The distribution problem: who gets the gains

If AI continues to develop in the direction of substitution, workers’ importance in production will continue to decline. Their claim on income no longer follows automatically from production.

One path is to use taxes and redistribution to spread the dividends of technology to the majority. This depends on the reliability of redistributive institutions and political will. Another path is to let technology enhance more people’s productive capacity, so that they have a basis for sharing returns and bargaining power in the process of value creation. Acemoglu advocates the latter.

The NBER paper proposes nine policy directions: equalize the tax burden on labor and on software and hardware capital at the margin; make public investments in health care and education; use antitrust enforcement to prevent tech giants from killing pro-worker business models through defensive acquisitions; have the government act as a first purchaser to embed demonstrations of pro-worker AI applications; and support worker representation in AI policymaking.

These policies face criticism. Economist Joshua Gans points out that the definition of pro-worker technology contains a contradiction: part of the reason a technology is valuable is that not everyone possesses the relevant skills. If a new technology spreads skills more widely, it may help more workers while reducing the wage premium of those who once had scarce skills. Gans uses education to illustrate this tension: universal education expanded literacy and raised skills, but it also eroded the wage premium once enjoyed by a minority. By the paper’s logic, even education would produce mixed effects.

Critics also point out that trying to design technology around specific employment goals may backfire. ATMs did not eliminate tellers, spreadsheets did not eliminate accountants, and GPS gave rise to ride-hailing and modern logistics systems. Government planners lack the expertise and real-time market information to pick winning technologies and understand their long-term effects.

The debate should focus on outcomes, not substitution versus augmentation

Acemoglu’s argument has boundaries. Augmenting AI does not guarantee more employment: if a tool doubles an employee’s efficiency, and demand does not grow, the company may still reduce staffing. Reducing the scarcity of certain professional skills may also expand access to services: if AI allows more people to enter occupations once protected by high barriers, consumers get cheaper services. This itself may increase social welfare, although the incomes of existing practitioners may be hurt.

The focus should be on outcomes: better services, lower costs, broader capability, and mitigated transition losses.

Acemoglu’s main point is that measuring progress only by headcount reduction misses a second standard: what work becomes possible that was not possible before.

That standard gets tested on budget sheets, tax schedules, and pilot reports.