Artificial Intelligence and the Changing Nature of Work

The Public Ledger

Technology

Published: July 27, 2026 - 10:33

Last Updated: Never

(Stock Free Images - Data Centre - https://l1nq.com/zuq7kdh)

Artificial intelligence is beginning to change not simply what people produce, but how they produce it. The evidence suggests that the most consequential question may not be whether machines will replace workers, but how much of a worker’s job will be reorganized around machines.

For decades, technological change has been accompanied by a familiar fear: that new machines will make human labour obsolete. The arrival of generative artificial intelligence has revived that concern with unusual intensity. Systems capable of writing, programming, summarizing documents, analyzing information and producing images have moved rapidly from research laboratories into ordinary workplaces.

Yet the emerging evidence does not fit neatly into either of the dominant narratives. Artificial intelligence is neither simply replacing human workers nor universally making them more productive. Instead, it is changing the composition of work itself.

That distinction matters.

A job is rarely a single task. A software developer does not merely write code. A customer-service representative does not merely answer questions. A manager does not merely produce documents. Most occupations consist of dozens or hundreds of individual activities, some of which can be automated, some of which can be accelerated, and some of which remain dependent on human judgement.

The future of work may therefore be determined less by whether AI can perform an occupation than by which parts of that occupation it can perform — and what humans do with the time and capacity that technology creates.

From replacing jobs to replacing tasks

Economists Daron Acemoglu and Pascual Restrepo have proposed a useful framework for understanding technological change: rather than asking whether technology replaces labour altogether, examine which tasks are transferred from workers to machines.¹

Their research distinguishes between two competing effects.

The first is displacement. When a machine becomes capable of performing a task previously carried out by a person, demand for human labour associated with that task can decline.

The second is the creation of new tasks. Technology can create activities in which human workers have a new comparative advantage, potentially restoring labour demand that was lost elsewhere.

This framework helps explain why previous technological revolutions did not simply result in permanent mass unemployment. The mechanization of agriculture, industrial automation and computerization eliminated or reduced demand for many particular forms of labour while simultaneously creating new occupations and new categories of work.

But this historical pattern should not be mistaken for a guarantee.

Acemoglu and Restrepo's research also finds that automation can reduce labour demand and the share of income going to labour, particularly when the creation of new tasks fails to keep pace with displacement.¹ Their work therefore presents a more complicated picture than either technological optimism or technological pessimism: innovation can increase productivity while still leaving some workers worse off.

That distinction is particularly relevant to artificial intelligence.

AI does not have to replace an entire occupation to substantially alter it. If a system can perform half of the routine tasks associated with a job, the remaining human work may become more specialized, more supervisory, or more dependent on interpersonal skills. Alternatively, the employer may decide that fewer workers are required.

The technology itself does not determine which outcome occurs.

The organization does.

What happens when workers receive an AI assistant?

Some of the strongest evidence available so far comes not from predictions about the future but from experiments conducted with workers performing real jobs.

In a widely cited study, economists Erik Brynjolfsson, Danielle Li and Lindsey Raymond examined more than 5,000 customer-support agents after the staggered introduction of an AI assistant.² The system provided suggestions to workers during customer interactions, but the human agent remained responsible for deciding whether to use them.

The result was significant: productivity increased by approximately 14 percent, measured by the number of customer issues resolved per hour.²

The distribution of the gains was particularly interesting.

The least experienced and lowest-performing workers benefited the most, with productivity improvements of roughly 34 percent. Experienced and highly skilled workers saw considerably smaller gains.²

This suggests that AI can function as a form of technological knowledge transfer. Instead of simply performing the worker's job, the system can provide less experienced employees with suggestions resembling the practices of more experienced colleagues.

In other words, AI can sometimes make workers more capable without making the workers themselves unnecessary.

This finding is consistent with a broader pattern emerging in research on generative AI: the greatest benefit may occur when the technology is used as an augmentation tool rather than as a fully autonomous replacement.

A 2023 experiment by Shakked Noy and Whitney Zhang provides another example. Their study randomly assigned 453 college-educated professionals to complete occupation-specific writing tasks either with or without access to ChatGPT.³ Participants using the system completed tasks approximately 40 percent faster, while the quality of their work increased by approximately 18 percent.³

The implications are significant, but they should not be overstated.

A worker completing a writing task 40 percent faster does not necessarily mean that the employer now needs 40 percent fewer workers. The saved time might instead be used for additional work, greater quality control, more research, customer interaction, or tasks that were previously neglected because there was not enough time to perform them.

Productivity is not synonymous with job destruction.

The productivity frontier is uneven

There is another reason to be cautious about assuming that AI can simply be handed a job and expected to perform it.

Research on highly skilled knowledge workers has found that AI's abilities are uneven.

In a field experiment involving 758 consultants, Fabrizio Dell'Acqua and colleagues examined how GPT-4 affected performance on tasks resembling the work performed by management consultants.⁴ The researchers found that AI significantly improved performance on tasks within its capabilities. Participants completed those tasks faster and generally produced higher-quality results.

But the researchers also found that the technology could perform worse than humans on tasks outside its capabilities.⁴

They described this uneven boundary as a “jagged technological frontier.”

That concept may prove more useful than the idea of a simple threshold separating jobs that AI can and cannot perform.

AI can be extraordinarily capable at one task and surprisingly unreliable at another that appears, to a human, to be of similar difficulty. A system might generate a useful first draft in seconds but confidently invent a source. It might identify a pattern in a large dataset but misunderstand why the pattern matters. It might write functional code while introducing a subtle security vulnerability.

The result is a new category of worker: neither someone who works entirely independently nor someone who merely supervises a machine. Instead, the worker becomes responsible for deciding when the machine should be trusted. That is a substantially different skill.

The worker becomes the verifier

This creates an important paradox.

The better AI becomes at producing plausible work, the more difficult it can become to identify when that work is wrong.

A calculator that produces an incorrect answer can often be checked through another calculation. A language model can produce an entire page of convincing prose containing a fabricated citation, an incorrect interpretation, or an invented fact.

This means AI-assisted work still requires domain knowledge.

The person using the system must understand enough about the subject to recognize an implausible answer, identify missing context and determine whether the result is suitable for the intended purpose.

The implication is that AI may reduce the amount of time spent producing certain kinds of information while increasing the importance of evaluating it.

That is not necessarily a reduction in skill. It may be a shift in where skill is applied.

Recent research on doctoral writing provides a useful illustration. A 2025 study of doctoral students found that generative AI was being used for numerous stages of dissertation work, including exploration, confirmation and execution, while students simultaneously expressed uncertainty about the boundary between legitimate assistance and inappropriate dependence.⁵ Another 2025 study examining doctoral theses found evidence that generative AI was altering patterns of authorial stance and potentially weakening explicit expressions of authorial identity.⁶

The lesson extends beyond universities.

When a machine becomes capable of producing a first draft, a summary or an analysis, the human contribution may increasingly lie in deciding what deserves to survive the drafting process.

AI can make workers more equal — and potentially less equal

One of the more promising findings from workplace AI research is that these systems can narrow performance differences between workers.

If an inexperienced customer-service employee can access guidance derived from thousands of successful interactions, the gap between that employee and an experienced colleague may shrink.²

That could make workplaces more productive while also making some forms of expertise easier to acquire. But there is no guarantee that AI will reduce inequality everywhere.

Research by Acemoglu and Restrepo has shown that automation and the creation of new tasks can have very different effects on wages and inequality depending on which workers benefit from the newly created activities.⁷

More recent research points in the same direction. An experimental study of UK workers found that generative AI could reduce productivity differences within certain groups while still raising questions about inequalities between different groups of workers.⁸ The distinction is important.

A technology can make two employees more similar in their ability to complete a particular task while simultaneously increasing the value of workers who possess the skills needed to manage, integrate or verify that technology. AI literacy may therefore become another form of workplace capital.

Those who know how to work effectively with AI may gain an advantage over those who do not.

Exposure does not mean replacement

This is where some of the most dramatic claims about AI and employment become difficult to interpret.

An occupation can have a high degree of AI exposure without being likely to disappear.

The International Labour Organization's 2025 update to its global occupational exposure index found that generative AI is more likely, in many occupations, to affect particular tasks than to automate entire jobs. The organization's analysis distinguishes between automation and augmentation, and finds that augmentation is generally the more prevalent possibility.⁹ This is particularly important for office-based and knowledge-intensive work.

An accountant may use AI to classify documents while still making the final judgement. A lawyer may use it to summarize case material while remaining responsible for legal interpretation. A programmer may use an AI coding assistant while deciding what software should actually be built. A journalist may use AI to organize notes while remaining responsible for determining what is true.

In each case, the technology can affect a substantial portion of the workflow without eliminating the occupation.

That does not mean the labour market will remain unchanged.

Research from Menaka Hampole and colleagues finds evidence that occupations with greater AI exposure subsequently experienced reduced labour demand, while also finding that occupations where AI exposure was concentrated in only some tasks could offset some of those losses by allowing workers to redirect their effort toward other activities.¹⁰

This may be the central tension of AI and employment.

The same technology can simultaneously substitute for some labour and complement other labour.

The gains do not automatically belong to workers

Even if AI raises productivity, there is another question that is often overlooked: Who receives the benefit?

Suppose an employee can complete a day's work in six hours instead of eight because AI has removed several hours of routine work.

There are at least three possible outcomes.

• The employee could work fewer hours for the same pay.

• The employee could use the additional time to perform more valuable work.

• Or the employer could simply increase the amount of work expected from the employee.

Technology does not decide between these outcomes. Labour markets, management practices, contracts and bargaining power do.

Recent research by Wei Jiang, Junyoung Park, Rachel Xiao and Shen Zhang offers reason to take the third possibility seriously. Their 2025 study finds that greater occupational AI exposure was associated with longer workdays and reduced leisure time, arguing that productivity improvements can sometimes increase the intensity of work rather than reduce it.¹¹

This challenges an appealing assumption about automation: that if machines make us more productive, humans will naturally work less. History does not provide such a guarantee.

Technological progress can make a worker capable of producing more in an hour. Whether that means the worker gets an extra hour of leisure or is expected to produce even more is ultimately a social and economic decision.

The uncomfortable possibility of substitution

The evidence for augmentation should not be used to dismiss the possibility of displacement. AI systems are improving.

A technology that currently assists a worker with a task may eventually become capable of completing that task independently. If the remaining human contribution becomes sufficiently small, employers may have an economic incentive to reduce staffing.

There is already evidence that labour demand can fall in areas with greater AI exposure.¹⁰ Financial-market research following the release of ChatGPT has also found evidence consistent with investors anticipating labour substitution in firms whose workforces were more exposed to generative AI.¹² And the history of automation provides an important warning.

Acemoglu and Restrepo's research shows that productivity improvements do not inherently guarantee rising labour demand. When machines replace tasks faster than new tasks are created, technological progress can reduce the demand for workers even while increasing output.¹

There is therefore no sound basis for claiming that AI will not replace jobs.

The more defensible conclusion is that job replacement is only one possible outcome — and probably not the most informative unit of analysis.

The better question is what happens to the collection of tasks inside a job.

A future of AI-assisted work

A recent doctoral thesis by Erik Engberg at Örebro University provides another example of why the relationship between AI exposure and employment cannot be reduced to a simple replacement story. His research, based partly on German data, finds that occupational AI exposure was associated with wage gains and an increased focus on knowledge-intensive tasks — a result that contrasts with the effects observed for robotics.¹³

This does not prove that AI will raise wages universally. It is one study, and its findings should be interpreted alongside the broader literature. But it reinforces an important point.

The labour market does not respond to technology in a single direction. Workers adapt. Firms reorganize. New tasks appear. Old tasks disappear. Skills become more or less valuable. Some workers benefit. Others do not.

Artificial intelligence is entering that process at remarkable speed.

The question for workers, employers and policymakers is therefore not simply whether AI should be adopted. It is how it should be integrated into human work.

A worker who uses AI to remove repetitive administrative tasks may become more productive without becoming less valuable. A company that uses AI to eliminate every entry-level task in an occupation may save money in the short term while inadvertently removing the training ground through which future experts would have developed.

The second problem is easy to overlook.

If junior workers rely on AI for every difficult task, they may become more productive today while developing less expertise for tomorrow. The technology can accelerate learning when used as a guide, but it can also make it possible to bypass the learning process entirely. That distinction will become increasingly important as AI becomes more capable.

The changing nature of work

The most reasonable conclusion from the evidence available today is neither that AI will save the labour market nor that it will destroy it.

It is that work is being reorganised around a new technological capability.

Some tasks will disappear. Some will become faster.

Some will become more valuable because AI makes them easier to perform.

Others will become more important precisely because AI cannot reliably perform them.

And some occupations will likely shrink or disappear as the economics of automation change.

The important question is what replaces them.

If AI allows workers to spend less time formatting documents, searching through information or producing routine drafts, the technology could make human work more creative, analytical and interpersonal.

If those same productivity gains are instead used primarily to reduce headcount or increase workloads, the same technology could produce very different social consequences.

That is why the debate over AI and employment should move beyond the question of whether machines can replace people.

They already can replace people for some tasks.

The real issue is what happens next.

The history of technological change suggests that the most consequential effects occur not when a machine performs something humans once did, but when society reorganizes around the fact that the machine can do it.

Artificial intelligence is beginning that reorganization now.

The future of work will be determined not solely by what AI becomes capable of doing, but by what humans decide is worth having humans do.

Bibliography

¹Acemoglu, D. & Restrepo, P. (2019). “Automation and New Tasks: How Technology Displaces and Reinstates Labor.” Journal of Economic Perspectives, 33(2), 3–30. DOI: 10.1257/jep.33.2.3.

² Brynjolfsson, E., Li, D. & Raymond, L. R. (2023). “Generative AI at Work.” NBER Working Paper No. 31161. DOI: 10.3386/w31161.

³ Noy, S. & Zhang, W. (2023). “Experimental Evidence on the Productivity Effects of Generative Artificial Intelligence.” Science, 381(6654), 187–192.

⁴ Dell’Acqua, F., McFowland, E., Mollick, E., Lifshitz-Assaf, H., Kellogg, K. C., Rajendran, S., Krayer, L., Candelon, F. & Lakhani, K. R. (2023/2026). “Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of AI on Knowledge Worker Productivity and Quality.” Organization Science.

⁵ Hoomanfard, M. H. & Shamsi, Y. (2025). “Generative AI in dissertation writing: L2 doctoral students’ self-reported use, AI-giarism, and perceived training needs.” Journal of English for Academic Purposes, 78, Article 101570. DOI: 10.1016/j.jeap.2025.101570.

⁶ Zhao, W. (2025). “Reconstructing stance in EFL doctoral thesis writing through generative artificial intelligence.” Humanities and Social Sciences Communications, 12, Article 1963.

⁷ Acemoglu, D. & Restrepo, P. (2020). “Unpacking Skill Bias: Automation and New Tasks.” AEA Papers and Proceedings, 110, 356–361. DOI: 10.1257/pandp.20201063.

⁸ Haslberger, M., Gingrich, J. & Bhatia, J. (2024/2025). “No Great Equalizer: Experimental Evidence on Productivity Effects of Generative AI Use in the UK Labor Market.” Working paper.

⁹ International Labour Organization (2025). Generative AI and Jobs: A Refined Global Index of Occupational Exposure. ILO Working Paper 140.

¹⁰ Hampole, M., Papanikolaou, D., Schmidt, L. D. W. & Seegmiller, B. (2025). “Artificial Intelligence and the Labor Market.” NBER Working Paper No. 33509. DOI: 10.3386/w33509.

¹¹ Jiang, W., Park, J., Xiao, R. J. & Zhang, S. (2025). “AI and the Extended Workday: Productivity, Contracting Efficiency, and Distribution of Rents.” NBER Working Paper No. 33536. DOI: 10.3386/w33536.

¹² Eisfeldt, A. L., Schubert, G. & Zhang, M. B. (2023/2026). “Generative AI and Firm Values.” NBER Working Paper No. 31222. DOI: 10.3386/w31222.

¹³ Engberg, E. (2026). The Impact of AI on the Labour Market: Essays on Transformative Technology, Occupations, and Firms. Doctoral dissertation, Örebro University School of Business.