AI Has Arrived. The Productivity Miracle Has Not.

Artificial intelligence has attracted billions of dollars in investment and predictions of a productivity miracle, yet its measurable effect on the wider economy remains modest. Nobel laureate economist Daron Acemoglu argues that the central question is no longer whether AI can perform impressive tasks, but whether it can be deployed reliably, widely and in ways that make workers more productive rather than simply making them expendable.

The distinction cuts through much of the language surrounding the technology. A model that drafts an email in seconds is useful. An agent that can navigate software, write code and complete a sequence of instructions is more impressive. Neither, by itself, proves that national output will surge, that companies will earn vast profits or that human labour is about to disappear.

Acemoglu’s estimate remains a stubborn counterweight to the prevailing enthusiasm. In his 2024 paper, The Simple Macroeconomics of AI, he calculated that artificial intelligence might generate total-factor productivity gains of roughly 0.55 percent over a decade, with favorable assumptions producing no more than about 0.66 percent. The associated increase in GDP could amount to roughly 1 to 1.5 per cent over ten years.

Those figures are economically meaningful. They are also far removed from the claims that AI could deliver annual growth of 10 percent or bring an imminent technological singularity. The gap between what the machines can demonstrate and what businesses, workers and institutions can absorb is where the real story lies.

Daron Acemoglu’s predictions

The first mistake is confusing capability with productivity

The modern AI debate often begins with a demonstration. A chatbot produces fluent prose, a coding assistant completes a function, or an image generator creates a convincing scene from a short prompt. The demonstration establishes that a system can do something. It does not establish that the task can be trusted, integrated into a workplace, performed at scale, or converted into higher national income.

Acemoglu’s argument starts with that less glamorous distinction. Productivity is not a score awarded to a machine for looking intelligent. It is an economic result. It depends on whether workers can produce more valuable goods and services with the same resources, whether new products emerge, whether organizations change in useful ways, and whether the gains survive the costs of supervision, error correction, and implementation.

In an interview about the economic impact of AI, Acemoglu separated two kinds of productivity improvement. One can hold output constant while reducing employment. The remaining workers appear more productive because fewer people are producing the same amount. The other kind makes workers more capable, allowing them to create better services, new products and new forms of organization. The second is the productivity miracle that technology companies imply. The first is closer to routine automation.

That difference matters to anyone outside a boardroom. A company that replaces a team of customer-service workers with a cheaper chatbot may lower its wage bill. It has not necessarily increased the value of what it produces. Customers may receive slower or less accurate answers, complex cases may be pushed back to a smaller human staff, and the apparent saving may be offset by complaints, refunds and reputational damage.

The same problem appears in professional work. An AI system can produce a plausible legal summary or financial report. If a qualified employee must check every line, investigate every unsupported claim and reconstruct the reasoning behind the answer, the software has shifted labour rather than removed it. It may still be useful. But the economic calculation is not the one advertised in a product launch.

A Nobel economist’s numbers puncture the growth fantasy

The most provocative part of Acemoglu’s analysis is not that AI has no value. It is that the value is likely to arrive at a pace that resembles a significant technological improvement rather than an economic rupture.

His task-based approach asks what AI can do to particular activities inside occupations, then considers how many of those activities can be automated or improved, how much time they represent and how quickly firms can adopt the technology. This is a more demanding exercise than counting the number of jobs exposed to AI. Exposure is not the same as replacement, and theoretical capacity is not the same as actual use.

On favorable assumptions, his estimates put the increase in total-factor productivity over ten years at no more than about 0.66 percent. More cautious estimates, which account for the difficulty of automating tasks that depend on context and judgment, come in below 0.53 percent. The figure of roughly 0.55 percent is modest by the standards of the rhetoric surrounding AI, though it is not trivial in macroeconomic terms.

In the interview, Acemoglu described the result as “definitely worthwhile productivity gains” but “nothing like what industry insiders or some techno-optimists were predicting”. That is a useful formulation because it refuses two easy errors at once. It does not deny the technology’s benefits. It denies that those benefits should be inflated into a forecast of runaway growth.

The comparison with earlier technologies is revealing. Personal computers changed offices, factories and professional practice. The internet transformed communication, retail, media and the organisation of information. Their effects were not always immediate, but they produced new businesses and new forms of activity that became hard to imagine living without. AI may eventually do the same. The evidence so far does not justify assuming that it has already done so.

A one or one-and-a-half per cent increase in GDP over a decade can improve living standards if the gains are broad and sustained. It can also disappear into existing trends, be captured by a narrow group of firms or fail to translate into better wages. The number is therefore a warning against both complacency and hysteria: the technology may matter greatly while the first economic returns remain surprisingly small.

The missing ingredient is adoption, not another product launch

AI companies have moved at extraordinary speed. Employers and institutions have not. That mismatch is one reason the technology looks revolutionary in demonstrations but incremental in economic statistics.

Most organizations are not blank sheets of paper waiting for an intelligent system to be installed. They are collections of legacy software, informal procedures, legal obligations, security rules, departmental rivalries and workers who have learned how to compensate for the weaknesses of existing systems. Introducing AI means deciding who is accountable when it fails, what information it may access, how its output is checked and whether the organisation is prepared to redesign the job around it.

Those decisions take time. A firm may begin with a pilot program, discover that its data is incomplete, find that employees do not trust the results and then spend months repairing the process. The model may improve during that period, but the business is still trying to work out what it can safely delegate. The technology can advance faster than the institution adopting it.

Estimates discussed in the interview put generative-AI use at about 40 percent among Americans aged 18 to 64, while only about 9 percent of employed workers use it daily. Even those figures describe access and use, not the share of work that has been redesigned around the technology.

The distinction between occasional assistance and operational dependence is crucial. A worker who asks a chatbot to suggest ideas is not working in an AI-enabled occupation in the same sense as a firm whose production, hiring or customer service depends on automated systems. One is an individual tool. The other is an organisational architecture.

The arrival of agentic AI could narrow that gap. Agents are designed to carry out multiple steps rather than provide a single answer. They may operate software, retrieve information, write code and return a completed result. Acemoglu said the development of agentic systems had moved faster than he expected, particularly after a period in which progress appeared limited. But he also described technological progress as sporadic and episodic, with bursts followed by stagnation rather than a smooth, predictable acceleration.

That is a more plausible account of technological change than the straight upward line found in promotional material. A burst can create the impression that the future has arrived. The harder test comes when thousands of ordinary organisations must make the burst useful.

“So-so” automation can make firms leaner without making economies richer

The most politically important risk is not that AI will instantly replace every worker. It is that businesses will use it to remove labor from existing processes without creating enough new tasks, products or opportunities to compensate.

Acemoglu has used the phrase “so-so automation” to describe technologies that are good enough to replace some workers but not good enough to produce a major improvement in output. Such systems can be attractive to managers because the financial calculation is immediate. Reducing headcount is visible on a quarterly budget. Building a new product, training a workforce and redesigning an organization requires patience, capital and a willingness to accept failure.

A self-service system that forces customers to solve more problems themselves may reduce staffing costs while leaving the underlying service unchanged. Automated document review may allow a firm to employ fewer junior professionals, even if it does not improve the quality of legal or financial decisions. A scheduling system may eliminate administrative roles without creating a new service that customers value. Each decision can make sense for the individual company. Taken together, they can produce a labour market in which output rises slowly while bargaining power falls faster.

This is why the employment question cannot be answered by asking only whether AI is productive. The relevant question is productive for whom, and in what way? If the gains come mainly from reducing labour’s share of income, shareholders and senior executives may benefit while workers face lower wages, fewer entry-level routes and greater insecurity. A technology can raise measured productivity and still produce a poor social bargain.

The danger is especially acute in knowledge work because many of the targeted tasks have served as training grounds. Junior lawyers learn by reading documents. Young analysts learn by assembling reports. New programmers learn by writing and debugging routine code. If those activities are automated before workers have acquired deeper expertise, firms may save money in the short term while weakening the pipeline of experienced professionals.

This concern has a historical parallel. During periods of rapid industrial change, technological gains have not always translated into immediate improvements in living standards. Economists have used the term “Engels pause” to describe a period in which output and capital accumulation rise while workers see limited gains. The comparison does not prove that AI will repeat the pattern. It shows why the distribution of gains cannot be treated as an afterthought.

The systems still lack the judgment that workplaces require

The strongest argument against near-term claims of artificial general intelligence is not that current models are unintelligent. It is that they are unreliable in the precise situations where responsibility matters most.

Large language models are skilled at producing answers that resemble the patterns in their training data. They can summarise, translate, classify, draft and recombine information with astonishing fluency. But fluency can conceal weakness. A model may state an incorrect fact with confidence, overlook a change in circumstances or fail to understand what a user is actually trying to accomplish.

Acemoglu’s criticism is direct: current systems lack judgment, social interaction skills, context and physical capability. These are not minor defects that can be ignored in high-stakes work. They are part of the work. A finance system must understand a client’s risk tolerance and recognize when market conditions have changed. A medical system must weigh incomplete evidence and communicate uncertainty. A legal system must distinguish an unusual fact from a routine one and understand the consequences of a mistaken interpretation.

The finance example illustrates the difference between theoretical capability and responsible deployment. A model may appear able to perform a large share of the tasks in a financial job under controlled conditions. That does not mean an institution should allow it to make decisions without oversight. If humans must clean up a stream of errors, the system has not automated 75 per cent of the work. It has created a new supervision problem.

Reliability also has an economic price. Every automated decision requires monitoring, audit trails, staff training and a process for appeal. The more consequential the decision, the more expensive those safeguards become. This does not make AI useless. It makes the business case more specific than the claim that intelligence is simply being sold as software.

An effective system would need to understand context as conditions change, explain its reasoning in a way different users can assess and recognise when it does not know enough. It would need to act within physical and institutional constraints, not merely produce a plausible sequence of words. Until then, AI is best understood as an assistant whose work must be directed and checked, not a general replacement for expertise.

Coding shows what useful augmentation looks like—and what it does not

Coding is often offered as the leading example of an occupation transformed by generative AI. It is also one of the clearest examples of why augmentation should not be confused with substitution.

Coding assistants can complete routine functions, explain unfamiliar code, generate tests and help developers navigate documentation. For an experienced programmer, these capabilities can remove tedious work and speed up a project. The gains may be substantial for particular people and tasks. In the interview, the host described his own productivity as having multiplied several times, and Acemoglu acknowledged that some companies and individuals were achieving remarkable results.

His qualification was the important part: those cases remain exceptions rather than the experience of the entire economy. A tool can be transformative for a small firm with a technically skilled user while producing little change in national productivity. The distribution of expertise, the quality of the underlying data and the ability to integrate the tool determine the result.

Nor does faster code automatically mean better software. Generated code can introduce security vulnerabilities, obscure design choices or solve the wrong problem with impressive efficiency. The developer’s job may shift from typing instructions to reviewing, testing and deciding. That is a meaningful change, but it leaves the need for human judgement intact.

The broader lesson is that AI tends to automate fragments of an occupation before it automates the occupation itself. A profession is a bundle of tasks held together by standards, relationships and responsibility. Removing one task may allow a worker to spend more time on another. It may also leave the entire process unchanged if the bottleneck lies elsewhere.

That is the optimistic route Acemoglu wants to see: AI used to give workers better tools, expand what they can do and create new occupations, products and services. It is different from the managerial route in which the technology is purchased mainly to reduce the number of people performing an unchanged service. The first approach seeks new value. The second seeks a cheaper version of the old model.

Healthcare, education and manufacturing will not absorb AI in the same way

The phrase “AI adoption” hides major differences between sectors. A recommendation system in retail, an image-analysis tool in a hospital and an automated tutor in a classroom do not present the same risks or require the same evidence.

Healthcare illustrates the difficulty. AI can assist with medical images, help identify patterns and support drug discovery. Those are real achievements. But clinical use depends on whether the system works across different populations, whether doctors can understand its limits and what happens when the recommendation is wrong. A tool that performs well in a laboratory or a carefully selected trial may not produce the same result in a busy hospital with incomplete records and unusual patients.

Research discussed in the interview from China suggests that students who use AI may perform better on exercises but worse on examinations and show weaker general knowledge. It also describes a split between students who use AI to support work they have attempted themselves and the larger group that asks the system to do the work for them.

That distinction is not a technical footnote. Learning requires effort before the answer arrives. If students delegate the struggle, they may submit better-looking work while acquiring less knowledge. A technology that helps a capable learner can harm a child who has not yet learned how to judge its output. Parents, teachers and communities therefore become part of the control system.

Manufacturing may offer a more constructive path. Europe and China have made stronger efforts than the United States to explore AI’s integration into production, according to Acemoglu. Factories already contain structured processes, measurable outputs and opportunities to combine machine capability with human skill. That does not eliminate disruption, but it may make pro-worker deployment easier to test than systems that make opaque decisions about people.

The policy implication is straightforward. AI should not be regulated as though every use were identical. Healthcare, education, finance, manufacturing and public administration need different standards because the cost of error, the role of human judgement and the potential benefits differ. A single promise about “AI” is no substitute for evidence in the workplace where it is being used.

The investment boom may be building an industry before it has a business model

The financial story has acquired a momentum of its own. Semiconductor companies, cloud providers, model developers and infrastructure firms have been valued as though explosive demand and enormous profits are already assured. But the value of a technology and the value of every company built around it are not the same thing.

The dot-com comparison is useful only if it is handled carefully. The internet was revolutionary even though many internet companies failed. The collapse of Pets.com did not disprove online commerce; it showed that a business could invest too much, move too quickly and run out of money before customers were ready. The cables and other infrastructure built during the boom later supported a much larger digital economy.

Acemoglu sees both similarities and differences in the present AI cycle. One difference is that the business model for generative AI is less settled. Companies are spending vast sums on computing, research and data centres, while many services are being offered at prices that do not yet reflect their full cost. The source interview describes a system in which customers may receive far more computational value than their subscriptions pay for.

That can be a rational strategy while firms are building a market. It is not proof of durable profitability. If the price of intelligence falls towards a commodity level, model providers may struggle to recover the cost of the infrastructure they have built. The technology could succeed while some of its most heavily funded companies fail.

The scale of the investment intensifies the risk. Acemoglu questioned how firms making enormous commitments could generate the revenue required to justify them. The answer cannot be that every company will sell the same underlying service to one another at ever higher valuations. Businesses must create products that customers value enough to pay for after accounting for errors, oversight and the cost of changing their operations.

Investors who confuse the importance of AI with the inevitability of exceptional returns are making a familiar mistake. A transformative technology can produce poor investments when capital arrives too early, valuations outrun revenue and companies are forced to monetise before the market is ready. The eventual winners may be real. That does not mean every current winner is permanent.

What happens to work depends on who controls the technology

The argument over AI policy is often framed as a choice between speed and restraint. That is too simple. The more consequential choice is whether AI will be developed as a centralising automation system or as a decentralised set of tools that increases workers’ capacity.

Acemoglu’s preferred model is explicit. He wants AI to be pro-worker: technology that helps people become more productive, preserves useful employment and creates new tasks, goods and services. The alternative is a system in which information and decision-making are concentrated in a small number of firms while automation is used to sideline human beings.

That concentration is not an abstract concern. A handful of companies control the models, computing infrastructure and data pipelines on which much of the new industry depends. Their decisions shape what systems can do, what information they are trained on, how they are priced and which risks are treated as acceptable. Acemoglu warned that leaving a transformative technology in the hands of five or ten people could not be optimal.

The political problem is made harder by the speed of development. Industrial societies had generations to argue over the social consequences of earlier technological shifts. AI companies can change the capabilities available to millions of people within months. Legislatures, schools and workplaces are expected to form rules while the underlying systems keep changing.

That is no excuse for doing nothing. It is an argument for targeted policy. Governments should examine whether firms are using AI to create new capacity or merely to cut labour, require accountability in high-stakes decisions and support training that builds judgement rather than dependence. Workers should have a voice in how systems are introduced, because they are often the first to discover where the demonstrations fail.

Acemoglu has also rejected the idea that the central problem is simply a lack of income support. Universal basic income might be preferable to no safety net, he said, but it would not create jobs or restore people’s ability to participate in social and public life. The question is not only how to compensate people after work disappears. It is how to ensure that technology does not remove their agency in the first place.

The policy debate is running ahead of the evidence

Several popular responses to AI fail because they begin with an assumption that has not been established: that automation is spreading too quickly and that mass unemployment is imminent.

Consider the proposal to reduce working hours while preserving pay. Acemoglu dismissed the idea that firms would voluntarily make labour more expensive without changing their incentives. If a company already has a reason to automate at the current cost of employment, raising that cost could strengthen rather than weaken the pressure to replace workers. A rule imposed without a plan for productivity, demand and enforcement could produce the opposite of its intended effect.

An AI tax presents a similar question. Taxing automated systems might slow adoption, which could be useful if firms are replacing workers before society can adjust. But Acemoglu’s assessment is that the immediate problem may be that models are being developed too quickly while adoption in the wider economy remains too slow. A tax on workplace use would not necessarily address a race among companies to build ever larger systems.

The first task, therefore, is diagnosis. Policymakers need to know whether the problem is excessive deployment, excessive concentration, inadequate worker power, weak safety standards or an investment bubble. The answer may differ by industry. A blanket response to a heterogeneous technology is likely to be blunt and easy to evade.

The same discipline should apply to claims about mass unemployment. Employment effects have so far been small in most occupations, while coding appears to have seen stronger change. That could shift with agentic systems, particularly if they make AI easier for small and medium-sized businesses to use. But a possibility is not a forecast, and a forecast is not a fact.

The responsible position is neither to slow every experiment nor to accelerate every application. It is to establish where AI improves a service, where it creates unacceptable risk and where it merely transfers cost from an employer to a worker or customer. Regulation should follow those distinctions instead of following the slogans of either the industry or its opponents.

The personal rule is simple: use AI, but do not surrender judgement

Individuals cannot settle the investment cycle or rewrite the rules of the technology industry. They can decide whether AI becomes a substitute for thinking in their own lives.

Acemoglu’s advice is practical. People need to become informed and involved because democratic oversight depends on citizens who understand what is being introduced in their workplaces, schools and communities. That does not require everyone to become a software engineer. It requires people to ask what a system is doing, who benefits, who is accountable when it fails and what human capacity may be lost through overuse.

For workers, the sensible response is selective experimentation. Learn the tools relevant to the job. Test them on low-risk tasks. Check the output against primary information. Keep responsibility for decisions that affect customers, patients, colleagues or the public. A worker who treats AI as a fast but fallible assistant will often gain more than one who assumes the machine has already replaced the need to understand the work.

The stakes are higher for children. A student who uses AI after attempting an assignment may gain a tutor, a critic or an explanation. A student who asks it to complete the assignment may gain a polished answer while losing the practice that makes later learning possible. The technology does not decide which role it will play. Families, schools and social norms do.

This is also where the rhetoric of artificial general intelligence becomes politically consequential. If AI is presented as the natural successor to human effort, people may begin to treat their own judgement as an obstacle rather than an asset. If it is treated as a tool whose purpose must be chosen, the public has a better chance of directing its use.

The economic evidence does not support panic, but it does not justify complacency either. AI can produce real gains for individuals, teams and specific industries. It can also reduce labour’s bargaining power, reward concentration and encourage institutions to replace thought with plausible output. The difference will not be decided by the models alone.

It will be decided by the choices made around them: whether firms redesign work to help people or simply remove them, whether governments regulate actual risks rather than fashionable predictions, whether investors distinguish technology from valuation, and whether citizens insist that a more capable machine must serve a more capable society.