11–16 minutes

Whenever a new technology is introduced, particularly one as pervasive as artificial intelligence, the debate about its consequences on work tends to revolve around a familiar question:

Will it create more jobs than it destroys?

It is an understandable question. Jobs are visible, politically relevant and relatively easy to count. They determine income, social status, taxation and access to welfare systems. For many people, having or not having a job remains one of the most consequential distinctions in life.

But is the number of jobs still the right metric for understanding what is happening to work?

I increasingly suspect that it is not.

The problem is that a job is not a unit of work. It is a social and legal arrangement within which a highly variable quantity of work is performed. Two economies may have the same number of employed people while requiring radically different amounts of human effort. One may rely mostly on full-time workers putting in long weeks; the other may distribute shorter working times among a larger proportion of its population. Counting jobs could make the two societies appear similar even though the amount of human life absorbed by economically relevant labor is very different.

The same ambiguity exists within a single economy over time. If productivity increases and the average working week becomes shorter, the number of jobs might remain unchanged while the total volume of human labor decreases. Conversely, employment may grow only because work is being divided into more precarious, fragmented or part-time positions.

In both cases, the employment count may move independently of the phenomenon we are actually trying to observe.

What, then, should we measure?

From jobs to human time

A more meaningful indicator might be the total number of hours worked in a year divided by the entire population, rather than only by the number of workers.

The distinction matters.

Average annual hours per worker tell us something about the experience of people who are already employed. Hours per capita, by contrast, would help us estimate how much economically relevant human labor a society requires from its population as a whole.

In simplified form:

Annual human labor per capita = total annual hours worked / total population

This metric would not answer every question. No single number could. It would not tell us whether working time is distributed fairly, whether the work is meaningful, whether it is adequately paid or whether some groups are carrying a disproportionate burden.

But it would bring us closer to the essence of the phenomenon under discussion: how much human time and effort does society need in order to function and produce its output?

Official statistics already contain many of the components required to observe this dimension. The OECD, for example, maintains separate indicators for employment and hours worked, while its productivity analysis treats hours worked and numbers of employed people as distinct ways of measuring labor input. The conceptual distinction is therefore hardly exotic. What may need reconsideration is which of the available measures should occupy the center of the public debate1.

When discussing AI, we continue to ask whether there will be enough jobs. Perhaps we should first ask how many hours of human labur will actually be necessary.

These are not equivalent questions.

Our extraordinarily “lazy” brain

Reducing the human effort required to obtain a given result is not an accidental by-product of technology. It may be one of the deepest and most persistent forces behind technological development.

In everyday language, we are inclined to describe the desire to avoid effort as laziness. The word carries a moral judgment: effort is virtuous; avoiding it is suspect. Yet from an evolutionary perspective, expending energy unnecessarily would hardly have been a virtue.

For living organisms, energy has never been free. Acquiring it requires time, exposure to danger and further energy expenditure. A system that consistently achieves the same useful result at a lower metabolic cost has an obvious advantage over one that wastes scarce resources.

The human brain itself is an extraordinarily expensive organ. Although it accounts for only about two per cent of body mass, it consumes roughly twenty per cent of resting energy. Research into brain evolution and neuronal organization shows that energy availability imposes important constraints on neural architecture and that nervous systems are optimized through trade-offs between information processing and energy consumption2.

In this sense, the brain’s tendency to search for shortcuts is not merely a character flaw. It is part of the machinery that made human intelligence possible.

We recognize patterns so that we do not have to reconstruct the world from scratch at every moment. We form habits so that frequently repeated actions require less conscious control. We create abstractions so that many individual cases can be handled through a single mental model. We build tools so that physical and cognitive effort can be transferred to objects, machines and, increasingly, software.

We automate because our brains are reluctant to keep solving problems they have already solved.

If we are here discussing artificial intelligence, it is partly thanks to this innate “laziness.”

The wheel, the lever, writing, mechanical engines, industrial machinery, computers and AI can all be seen, from this particular perspective, as stages in the externalization of effort. Each allows a result to be achieved with less direct human energy, or allows the same human energy to produce more.

Civilization might therefore be read as the history of an animal attempting to reduce the amount of itself required to obtain a desired outcome.

Technology does not have to destroy jobs to reduce work

This distinction helps reveal a weakness in the conventional debate about automation.

Suppose a new technology increases productivity but does not reduce the number of jobs. Should we conclude that it has not reduced human labor?

Not necessarily.

It may have shortened the time required to perform each task. It may have made it possible to produce more without a proportional increase in working hours. It may have changed full-time roles into part-time ones or redistributed tasks among a broader population. It may also have reduced work in one part of the economy while generating new activities elsewhere.

Looking only at the number of jobs compresses all these possibilities into a binary classification: employed or unemployed.

But work is not binary. It is a quantity of human time.

A person working twenty hours per week and a person working sixty hours per week each count as one employed person. From the perspective of employment statistics, both represent one job. From the perspective of human labor, however, their contributions differ by a factor of three.

The number of jobs can therefore remain stable, or even increase, while the aggregate amount of labor per capita falls.

This may help explain why the familiar observation that “technology has always created new jobs” does not settle the question. Even if the statement continues to be true, those new jobs might not require enough total human time to compensate for the labor eliminated elsewhere.

The real question is not only whether new forms of work will emerge. They almost certainly will.

The question is whether the additional human labor required by those activities will exceed the labor that automation removes.

I am not convinced that it will.

Artificial intelligence inside a much older trend

Artificial intelligence is often presented as an unprecedented rupture. In some respects, it may be. Its ability to automate or support cognitive activities distinguishes it from technologies mainly designed to replace muscular effort.

Yet, viewed from a wider historical perspective, AI may also represent a continuation of something much older: the progressive reduction of human input required to deliver a given output.

The difference may be one of scope and acceleration.

Previous machines automated individual physical operations or clearly defined administrative procedures. AI-based systems can increasingly participate in activities that once appeared difficult to formalize: writing, programming, classification, analysis, design, planning and interaction through natural language.

This does not mean that all these activities will be fully automated, nor that human contribution will disappear. It means that the ratio between output and human time may change even when people remain involved.

An engineer assisted by several AI agents may produce more in the same working day. A small company may perform activities that once required a much larger organization. A professional may delegate research, drafting, comparison and routine analysis while retaining responsibility for judgment and decisions.

None of this necessarily eliminates a job.

But it may reduce the human labor required per unit of economically valuable output.

My expectation is therefore not simply that AI will destroy occupations. It is that AI will consolidate, and possibly accelerate, the long-term tendency towards requiring less economically relevant human labor per capita.

My earlier essay, Workless Society — Part 2, described this as a continuation of the historical decline in working time accompanied by an extraordinary expansion in the goods and services available to ordinary citizens. It also proposed replacing job counts with “economically relevant human labor,” expressed as hours worked per capita. The spread of generative and agentic AI makes that proposal more urgent because the new technology may change the volume of labor without immediately changing the formal structure of employment.

The demographic paradox

There is another reason why automation may become less of a choice and more of a necessity.

For a very long time, societies could assume that relatively large younger generations would replace older ones. In many countries, that assumption is becoming increasingly difficult to sustain. Ageing populations and changing demographic structures mean that fewer working-age people may be available to support larger elderly populations.

In that context, reducing the human labor required to maintain output is not merely a corporate strategy for cutting costs. It may become a condition for preserving living standards.

Agriculture, manufacturing, logistics, healthcare and social assistance will still need to function. Infrastructure will still require maintenance. Goods will still need to be produced and distributed. If the proportion of people conventionally expected to perform these activities declines, productivity-enhancing technologies will become increasingly important.

This makes the conventional fear of “machines taking jobs” somewhat paradoxical. Some societies may soon need machines precisely because there will not be enough people available, or willing, to perform all the work demanded by the inherited economic structure.

Automation and AI could consequently reduce the amount of labor required for each activity while helping compensate for a shrinking labor supply.

Both things can be true at once.

And this is another reason employment counts alone may mislead us. A high employment rate in an ageing society could coexist with a decline in total hours worked per capita. Alternatively, a smaller working-age population might maintain or increase total output only because each hour of human work has been amplified by machines.

The central transformation would concern the relationship between output, population and human time. The number of jobs would reveal only part of it.

Counting jobs answers a political question

None of this makes employment statistics irrelevant.

The number of jobs remains crucial within societies organized around wages. If access to housing, healthcare, pensions, social recognition and personal autonomy depends on employment, then losing a job is a profound event even if the economy as a whole requires less labor.

But this means that job counts answer primarily a political and institutional question:

How many people have access to the mechanism through which income and social participation are currently distributed?

Hours worked per capita answer a different question:

How much human labor does society actually require?

Confusing these questions makes it difficult to understand either one.

If technology reduces the second quantity while our institutions continue to base income, dignity and citizenship on the first, the resulting problem is not necessarily a shortage of useful output. It is a mismatch between our productive capacity and the way we distribute its benefits.

Trying to preserve jobs merely to preserve access to income may then conceal the real achievement of automation: the ability to produce what society needs with less expenditure of human life.

This does not imply that work has no value beyond production. Work can provide purpose, structure, learning, identity, and social connection. A society requiring less economically necessary labor would still need activities through which people contribute, cooperate and feel useful.

But economically necessary work and meaningful human activity are not synonymous. Treating them as though they were may lead us to create or preserve labor because our institutions cannot imagine distributing income and social recognition in any other way.

A different dashboard for the age of AI

If we genuinely want to understand the effects of technological change, we may need a broader dashboard.

Alongside the employment rate and the number of jobs created or destroyed, we should examine at least:

  • total annual hours worked;
  • annual hours worked per resident;
  • annual hours worked per working-age resident;
  • the distribution of those hours across the population;
  • output produced per hour of human labor;
  • the proportion of involuntary part-time and excessive work;
  • uncompensated care and domestic labor;
  • and the relationship between productivity gains, income and leisure.

Each measure would illuminate a different aspect of the transformation. Together, they would help distinguish between less work, worse employment, unequal work and more efficient production.

Most importantly, they would prevent job creation from being treated automatically as evidence of progress and job reduction as evidence of decline.

A society that produces more, distributes its output fairly and requires fewer hours of human labor has not necessarily failed because it contains fewer conventional jobs. It may have achieved one of civilization’s oldest ambitions.

Conversely, a society that creates millions of poorly paid jobs merely to preserve employment statistics may not have created more prosperity. It may simply have distributed inefficiency.

The success we have forgotten how to recognize

For billions of years, evolution has selected organisms capable of acquiring and using energy effectively. In our species, that pressure contributed to the development of a brain able not only to solve problems, but also to remember solutions, generalize them, teach them and embody them in tools.

We call the resulting tendency laziness when it appears in individuals.

When it operates across generations, we call it progress.

Artificial intelligence may be the latest expression of that ancient drive: obtain the same result with less effort, or a greater result with the same effort. Whether it ultimately reduces human labor remains an open empirical question. New desires, services and industries may expand enough to absorb all the time released by automation, and perhaps even demand more.

But counting jobs will not give us the answer.

If we want to know whether AI and other technologies are increasing or reducing work, we must measure work itself: the portion of human time and energy that society must devote to economically relevant production.

The crucial curve may not show the number of employed people.

It may show the number of hours demanded from each member of society.

And if that curve continues to fall, perhaps we should not immediately ask how to force it upwards again. Perhaps we should ask how to distribute the remaining work, the resulting wealth, and the newly available time.

After all, if the human brain had never tried to avoid unnecessary effort, we would probably not be here to complain about its laziness.


Credits

  1. OECD Employment Database, Hours worked, OECD Compendium of Productivity Indicators 2024 ↩︎
  2. Neuronal energy use and brain evolution, Power saving in the brain ↩︎