The Economist has delivered an unusually harsh assessment of Daron Acemoglu, questioning whether the enormous influence of the Nobel laureate’s ideas is fully justified. Its criticism extends to artificial intelligence, where Acemoglu has become one of the most prominent advocates of directing the technology toward workers rather than simply using it to replace them.

Criticism of Acemoglu’s forecasts is legitimate. But the debate risks becoming fixated on the wrong question.

The most important issue is not whether Acemoglu has precisely predicted how much productivity AI will generate over the next decade. It is whether artificial intelligence can continue advancing faster than humanity’s ability to understand, govern and distribute the power it creates without producing serious social consequences.

That question becomes more important, not less, as AI grows more capable.

Acemoglu may underestimate AI without being wrong about its direction

In The Simple Macroeconomics of AI, published in Economic Policy, Acemoglu estimated that current AI advances would increase U.S. total factor productivity by no more than about 0.66% over ten years. Once harder-to-learn tasks enter the calculation, his estimate falls below 0.53%.

Those numbers may eventually prove too conservative.

AI capabilities have advanced rapidly since Acemoglu first developed those estimates. Stanford’s 2026 AI Index reports major gains in coding, reasoning and computer-use agents. On the OSWorld benchmark, for example, agent success on real computer tasks jumped from around 12% to roughly 66%.

But suppose Acemoglu did underestimate the pace of technological progress. What exactly would that prove?

It would challenge his productivity forecast. It would not invalidate his broader argument about how societies choose to use technology.

In fact, the faster AI develops, the more urgent that argument becomes.

AI is not simply another productivity tool. It increasingly enters areas that societies once regarded as distinctly human: writing, analysis, programming, research, diagnosis, education, judgment and decision-making. Its development therefore affects not only how much economies produce, but also who holds economic power, which skills retain value and how humans acquire knowledge.

A technology with those characteristics cannot be evaluated through productivity statistics alone.

Someone is already deciding the direction of AI

The Economist describes the idea of “pro-worker AI” — technology designed to complement human labor rather than simply replace it — as attractive but rather obvious. It then raises a reasonable question: who exactly should decide what kind of AI complements humans and benefits society?

But that question has another side.

If society does not consciously influence the direction of AI, it does not follow that nobody will.

Someone will still decide.

Companies decide which systems receive investment. Developers decide which capabilities receive priority. Infrastructure owners determine access to scarce computing resources. Employers decide when AI should augment a worker and when automation offers a cheaper alternative.

Markets are extraordinarily effective at directing resources toward profitable opportunities. But profitability and long-term social welfare are not always identical.

A company can rationally seek lower labor costs, greater automation and higher market share. A society may simultaneously value employment pathways, education, human expertise, social mobility and a broad distribution of economic power.

There is no contradiction here. The two simply operate under different incentives.

That distinction matters because AI development is already highly concentrated.

Stanford’s 2026 AI Index reports that industry produced more than 90% of notable frontier models in 2025. The same report says a single company, TSMC, fabricates almost every leading AI chip, highlighting how much of the physical foundation of advanced AI depends on one semiconductor foundry.

The U.S. Federal Trade Commission has identified another layer of concentration. Its examination of the Microsoft–OpenAI, Amazon–Anthropic and Google–Anthropic partnerships found that technical and contractual relationships can create switching costs for AI developers and can give cloud partners access to sensitive technical and business information unavailable to others.

None of this proves that the companies involved intend to harm society.

That is not the argument.

The concern is structural: a technology capable of reshaping work, knowledge and economic power is increasingly being developed within a relatively small network of companies whose primary incentives are necessarily commercial.

Leaving AI entirely to those incentives is itself a decision about the future of AI.

Human beings should develop with the technology

Acemoglu’s strongest argument may therefore be less about limiting AI than about changing what we expect AI to accomplish.

In February, Acemoglu, David Autor and Simon Johnson defined “pro-worker AI” as technology that expands human capabilities and makes expertise more valuable. Their framework distinguishes systems that merely automate existing tasks from systems that allow people to perform new tasks, develop expertise and exercise better judgment. They also argue that market incentives can lead companies to underinvest in these more worker-enhancing applications.

The distinction is fundamental.

AI can replace part of a nurse’s work, or it can give that nurse diagnostic capabilities that previously required a specialist.

It can remove entry-level programming tasks, or enable junior developers to solve problems that once required years of experience.

It can automate parts of teaching, or give teachers better tools to understand individual students.

All these scenarios involve more capable AI. But they do not produce the same society.

In one model, technology develops while human capability gradually becomes less economically valuable. In the other, technological progress raises what humans themselves can accomplish.

The second should be the more ambitious definition of progress.

Recent labor-market evidence gives us a reason to take the distinction seriously. Researchers at Stanford’s Digital Economy Lab, using payroll records covering millions of U.S. workers, found no evidence of widespread economy-wide AI job displacement. Yet employment among workers aged 22–25 in highly AI-exposed occupations stood about 19% below where it would have been had it followed the employment trend of similarly aged workers in less-exposed occupations. The researchers also found that the divergence was concentrated in jobs where AI tends to automate tasks rather than complement workers.

This does not prove that AI caused the entire gap, and the researchers do not claim that it does.

But the pattern matters because entry-level jobs perform another function beyond producing output: they allow people to acquire experience.

If AI removes the bottom rungs of professional ladders, societies must ask how tomorrow’s experts will acquire the knowledge that today’s experts developed through those jobs.

The deeper risk may be cognitive

Acemoglu’s more recent research goes beyond employment.

In AI, Human Cognition and Knowledge Collapse, written with Dingwen Kong and Asuman Ozdaglar, the researchers construct a theoretical model in which agentic AI improves immediate decision-making but can simultaneously reduce the incentive for humans to learn. Human learning, in their framework, creates both private knowledge and small contributions to society’s broader stock of knowledge.

Under certain assumptions, highly accurate AI recommendations can therefore produce a paradox: individuals receive better answers while society gradually weakens its own capacity to generate knowledge.

The researchers call the extreme theoretical outcome a “knowledge-collapse” steady state.

This is a model, not a prediction that such a collapse will happen. But the question it raises is profound.

Getting the correct answer and developing the ability to discover the correct answer are not the same thing.

If students stop learning how to construct arguments because AI can construct them instantly, if junior professionals lose opportunities to build expertise because machines perform their early-career tasks, or if decision-makers increasingly delegate judgment to systems they cannot independently evaluate, society can become more technologically capable while becoming more intellectually dependent.

That is not necessarily progress.

AI should not advance in isolation from humanity

This is where the AI debate needs a broader definition of development.

Humanity’s technological capability can advance extremely quickly. Our institutions, educational systems, ethical standards and mechanisms for distributing economic power evolve much more slowly.

The widening gap between those two processes may become one of the defining problems of the AI era.

Stanford’s 2026 AI Index already warns that responsible-AI practices are not keeping pace with capability improvements. Documented AI incidents rose from 233 in 2024 to 362 in the latest reporting year, while safety and responsibility reporting among frontier developers remains inconsistent.

The answer is not to freeze artificial intelligence.

AI could accelerate scientific discovery, improve medicine, expand access to expertise and increase human productivity on a scale that is difficult to predict today. Rejecting those possibilities would make little sense.

But technological capability alone cannot define progress.

AI development should advance together with human development — with education, institutional capacity, democratic oversight, economic inclusion and the ability of individuals to think and act independently.

The goal should not be to force humans to adapt endlessly to whatever form of AI produces the highest return on capital. It should be to build AI that increases what humans can understand, create and accomplish.

Acemoglu may prove too pessimistic about AI’s productivity numbers. The Economist may ultimately win that particular argument.

But that would leave the larger question untouched.

If increasingly powerful AI systems remain concentrated within a small number of companies and are developed primarily according to the incentives of capital, who ensures that technological progress also becomes human progress?

That question cannot be dismissed by asking who has the authority to steer AI. Choosing not to steer it simply transfers that authority to those who already control the capital, infrastructure and models.

The greatest challenge of the AI era may therefore not be whether machines become more intelligent than humans.

It may be whether our technological power grows faster than our collective wisdom to use it.