AI and Work: What It Can Do, What It Cannot, and Who Decides

This week I'm taking a slightly different turn. There's almost no techie stuff in this one, but I think it's a cool topic anyway. Have a read for yourself.

In “Will AI replace workers? Not if we build it right,” Nobel laureate and MIT economist Daron Acemoglu argues that the central question is not whether AI can imitate people. It is whether we choose to build systems that replace workers or help them do better work.

My reading is less dramatic than the usual job-loss headline. Acemoglu does not promise that AI will leave every job untouched; he argues that full job automation will arrive slowly because most occupations combine technical tasks with judgment, social interaction, exceptions, and responsibility.

The fear is real: the essay cites Pew’s finding that 52% of Americans worry about AI’s impact on their jobs. Acemoglu’s point is that this anxiety should not decide the technology’s direction by itself. Extreme optimism and extreme fear can both push organizations toward the wrong target, automating a job instead of improving the work.

That is a useful frame, but it needs a practical reading. AI already automates some tasks and improves performance on others. Those effects vary by task, worker, workflow, and organization. A tool can reduce minutes on one step while adding review, correction, and coordination elsewhere. A capability demo alone cannot tell us whether a whole job will disappear, whether output will improve, or who will receive the gain.

The distinction to keep in view is task exposure is not job replacement. Jobs are bundles of tasks, responsibilities, relationships, and judgment. AI may take over part of a bundle, change the order of work, or make a person more capable. The result depends on what employers do with the remaining work and whether they create new tasks where people can contribute.

The practical boundary: AI can produce an answer. A person or institution still has to decide whether the answer is good enough to use.

What Acemoglu is arguing

Acemoglu’s essay separates the capability of a model from the economic outcome of deploying it. A model may perform a task in a controlled test, yet a company still has to connect it to data, software, decisions, users, and accountability. That integration is slow. An error that looks small in a demo can be costly when it changes a medical recommendation, a customer record, or a production system.

He also challenges the assumption that technology naturally creates broad prosperity. In his task-based research with Pascual Restrepo, automation can displace workers from tasks they previously performed. New tasks can reinstate labor demand, but that outcome does not happen automatically or at the same pace. Their framework calls these the displacement and reinstatement effects. The balance between them shapes jobs, wages, and labor’s share of income. Their task framework is summarized in this NBER paper.

The essay’s proposed direction is pro-worker AI: systems designed to complement human expertise, provide information workers could not previously access, and support new work rather than merely reducing headcount. His teaching example is specific: use student work to identify where learners are struggling, then give a teacher better information for adapting instruction. The AI helps surface patterns; the teacher decides what to do with the class. I agree with that direction: the useful question is whether the tool gives a worker better information and new options, not whether it can imitate the worker.

Complementarity is a design choice: the model can surface a pattern, but the worker keeps the context, judgment, and authority to decide what happens next.

That is an argument about design and incentives, not a claim that AI is harmless. A company can use the same model to assist a worker, monitor them, deskill their role, or remove the role. The technology sets some possibilities; management choices, labor institutions, competition, and public policy influence which possibilities get funded and deployed.

What AI can do well today

Modern generative AI is especially useful when a task starts with a large amount of digital information and ends with something a person can inspect. It can draft, classify, summarize, search, translate, extract, and generate plausible alternatives. It can also act as a conversational interface to tools and data, provided the system gives it the right permissions and checks what it does.

There is evidence of real gains in some settings. In a field study of 5,179 customer-support agents, an AI assistant increased issues resolved per hour by nearly 14% on average, with larger benefits for novice and less-skilled agents and little or no benefit for the most experienced agents. The authors suggest that the tool helped spread practices used by stronger workers. The NBER study is a concrete example of augmentation, but the result belongs to that company, tool, and task.

In a randomized experiment involving professional writing tasks, participants with ChatGPT completed work faster and received higher-quality evaluations on average. The benefits were larger for people who initially performed less well, narrowing the gap in output quality. It was a controlled experiment with bounded writing assignments, not a measure of how an organization performs after months of using AI. Noy and Zhang’s experiment makes the potential clear while also showing why we need task-specific evidence.

What AI cannot be trusted to do on its own

AI systems can produce fluent answers without having verified them. Their output can be stale, incomplete, internally inconsistent, or confidently wrong. They may miss a rare exception because it is rare, not because it is unimportant. Giving an agent access to a tool adds another failure mode: it can take an action based on a mistaken interpretation, so permission design and auditability matter as much as answer quality.

• It cannot guarantee truth. A plausible explanation is not evidence. High-impact answers need independent checks against authoritative sources or observed facts.
• It cannot infer your actual goal from vague instructions every time. It can guess at missing context, but the guess may quietly change the work.
• It cannot accept accountability. A model does not carry professional duties, legal responsibility, or the consequences of a decision. A named person or institution still has to own them.
• It cannot reliably handle every edge case or physical situation. Software agents can only act through their sensors, interfaces, and tools; a language model alone does not inspect a room, feel a machine failing, or understand an unrecorded local constraint.
• It cannot make a workflow good just by speeding up one step. Review, integration, data cleanup, and exception handling can consume the saved time, or even take longer.

These are practical boundaries, not proof that AI will never improve at reasoning, planning, or physical tasks. Capabilities change. The safe claim is narrower: today’s systems should be evaluated in the environment and workflow where people will rely on them, with a way to detect and recover from errors.

The evidence is mixed because the work is different

The positive writing and customer-support results do not mean every employee will become faster. In a randomized trial on mature open-source codebases, experienced developers took 19% longer on average when allowed to use the early-2025 AI tools in the study. The sample was small, with 16 developers and 246 tasks, and the work involved repositories the participants already knew well. It should not be generalized to every developer or newer tools. It does show that review and interaction costs can outweigh generation speed in some contexts. METR’s study and limitations are worth reading alongside claims about coding speedups.

At the economy-wide level, Acemoglu’s 2024 paper estimated that AI would increase total factor productivity by no more than 0.66% over ten years, based on task-level estimates then available. That was a model-based estimate with explicit assumptions, not a measurement of what happened or a ceiling on what future systems can do. It helps explain why a breakthrough model does not immediately translate into a large GDP effect: adoption, process redesign, new products, and worker transitions take time. The paper makes those assumptions visible.

Occupational exposure studies also measure potential task overlap, not layoffs. The ILO’s 2025 index estimates that one in four workers globally are in occupations with some GenAI exposure, while 3.3% of global employment is in its highest exposure category. The ILO says job transformation is more likely than full automation for most occupations. That is a useful map for asking where to investigate; it is not a forecast that one in four jobs will disappear. See the ILO paper and methodology.

The studies answer different questions: Can people finish a defined task faster? Does a deployed tool change team output? Which job tasks overlap with model capabilities? Does a firm reorganize work or reduce hiring? What happens to wages and new job creation? A headline result from one layer does not answer the others.

Evidence rule: a productivity result belongs to a task, a tool, and a setting before it belongs to “AI” in general.

Where AI in work is likely to go

Near-term change is likely to show up as task redesign before job disappearance. Some routine digital tasks will be automated. In other roles, AI will make preparation faster, create more drafts or analyses, and shift time toward decision-making, customer interaction, or quality control. Some jobs will shrink; others will acquire new responsibilities. The transition will be uneven across industries and workers.

Over time, the direction depends on three forces:

1. What firms choose to build: systems that reduce staffing are easy to explain in a budget; systems that improve judgment, training, or service quality may require more complex deployment and measurement.
2. Whether new tasks appear: productivity gains can lower costs and expand demand, but businesses and public institutions still have to create work where people’s skills matter.
3. Who has a voice in deployment: workers often know where the real bottlenecks and exceptions are. Involving them in design can produce better tools and fairer transitions; excluding them can make an apparently efficient system fail in use.

Acemoglu also points to incentives outside the model. In the United States, the essay cites an effective marginal tax rate of more than 25% on labor income and close to zero on capital income. That difference can make replacing or avoiding workers look cheaper than hiring and training them. He also argues that concentration matters: the seven largest technology companies account for about 60% of the NASDAQ, making it harder for new companies with different, more worker-friendly ideas to compete.

Acemoglu’s electricity comparison is useful here. New York and London had central generating stations by 1881, yet only about half of factories and homes used electricity by the 1920s. An enabling technology does not change productivity simply because it exists. Organizations need complementary investments, redesigned processes, skills, and applications. AI may spread faster in some areas, but the same adoption problem remains: firms have to reorganize work around it.

How to tell whether AI is helping workers

Measure more than adoption, prompts, or time saved on an isolated step. Before a pilot, write down the outcome that matters and measure the whole task from start to accepted result. Depending on the work, that might include:

• time to complete and time spent reviewing or correcting;
• error rates, rework, customer outcomes, and consistency across experience levels;
• whether workers take on higher-value work, learn new skills, or lose discretion;
• the number and type of tasks removed, changed, or newly created;
• who receives the productivity gain: workers, customers, owners, or some combination.

Keep a human decision-maker for consequential actions. Give that person enough information and authority to disagree with the system. Use restricted permissions for agents, logs for actions, and a fallback process when the model is uncertain or unavailable. Even a claimed 99% accuracy can be inadequate when the remaining 1% changes the meaning of a message or creates expensive review work. These controls do not make AI infallible; they make its errors easier to catch and contain.

The role of AI should be chosen, not assumed

AI can search and transform information at a scale people cannot match, generate useful first passes, and make some expertise easier to access. It can remove repetitive work and help people perform tasks they could not previously do alone. It also makes mistakes, struggles with unfamiliar context, and can shift power and income away from workers when firms use it mainly to substitute for labor.

Acemoglu’s strongest point is that the labor-market outcome is not written into the model. The same technology can displace or assist. A sensible test is simple: does the deployment improve the work people can do, and can we show that with evidence from the actual workflow? If the only success measure is fewer people doing the same work, the organization has chosen an automation strategy. If it measures quality, learning, service, and new human capabilities as well, it has at least made room for a different result.

Sources and further reading

• Daron Acemoglu, “Will AI replace workers? Not if we build it right,” The Humanist Review of AI
• Acemoglu and Restrepo, “Automation and New Tasks: How Technology Displaces and Reinstates Labor”
• Acemoglu, “The Simple Macroeconomics of AI”
• Brynjolfsson, Li, and Raymond, “Generative AI at Work”
• Noy and Zhang, “Experimental Evidence on the Productivity Effects of Generative Artificial Intelligence”
• METR, “Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity”
• ILO, “Generative AI and Jobs: A Refined Global Index of Occupational Exposure”