A worker watches software produce in seconds a first draft that once occupied part of an afternoon.

It is not perfect.

Somebody still has to check it. Somebody still has to understand what the client wanted, whether the underlying information is correct and what happens if the answer is wrong.

But the amount of time has changed.

Once time changes, the organisation of work can change with it.

That is where the argument about artificial intelligence and employment becomes more difficult than asking whether a machine can "do a job."

Jobs are not single actions. They are bundles of tasks, responsibilities, relationships and decisions.

AI can automate one part of a job while making the person doing the rest of it more productive. It can make an employee more valuable. It can also make another vacancy unnecessary.

All three things can happen inside the same company.

The labour market has not collapsed

It is worth beginning with what the evidence does not show.

By the middle of 2026, there was still no persuasive evidence of a broad, economy-wide employment collapse caused by generative AI.

A major Stanford Digital Economy Lab analysis published in 2025 and revised in August 2026 examined high-frequency payroll data covering millions of US workers through June.

Its first finding was straightforward: the researchers found no evidence of widespread economy-wide job displacement associated with AI.

That conclusion matters because public discussion has often leapt from an impressive demonstration to the assumption that entire professions must therefore already be disappearing.

They are not.

But the same research contains a much more uncomfortable second finding.

No

Evidence of widespread economy-wide AI job displacement in Stanford's payroll analysis through June 2026.

19%

Estimated employment shortfall for workers aged 22 to 25 in highly AI-exposed occupations relative to less-exposed peers.

Hiring

The young-worker divergence appeared mainly through fewer hires rather than a surge in separations.

The first warning may be at the bottom of the ladder

Stanford's researchers found that workers aged 22 to 25 in highly AI-exposed occupations were increasingly falling behind similarly aged workers in less-exposed occupations.

By June 2026, employment in those highly exposed groups stood about 19 per cent below where it would have been if it had kept pace with employment among less-exposed peers.

The researchers found no comparable gap among experienced workers.

That distinction is more important than another dramatic headline about robots taking jobs.

It suggests that disruption can appear first in the opportunities that never open.

A company does not have to fire a junior analyst if it decides not to hire the junior analyst in the first place.

There is no redundancy announcement.

The missing worker never appears on the payroll.

The first AI job loss may not look like a layoff. It may look like a vacancy that was never created.

Important limitation

Nineteen per cent does not mean AI destroyed nineteen per cent of those jobs.

Stanford's authors explicitly describe their results as descriptive indicators rather than causal estimates.

The gap becomes smaller when education is taken into account, some differences existed before generative AI became widespread, and results from the ADP payroll sample are stronger than some national survey benchmarks.

The evidence is concerning. It is not permission to pretend that every missing junior job has one proven cause.

Hiring can change before unemployment does

This is one reason unemployment statistics may be a poor early-warning system for technological change.

Imagine a team of ten people that previously expected to become a team of twelve.

New software allows the existing staff to handle the additional workload.

The company remains at ten.

Nobody lost a job.

Two jobs nevertheless disappeared relative to the company's previous plan.

Multiply that decision across thousands of organisations and entry into a profession can become harder without producing a sudden wave of redundancies.

The Stanford data point in precisely this direction. The deterioration among younger workers appeared primarily through reduced hiring rather than increased separations.

Why younger workers may be unusually exposed

The mechanism is not mysterious.

Junior employees often begin professional careers by doing work that is relatively structured and codified.

They summarise documents.

They prepare first drafts.

They clean datasets.

They conduct routine research.

They write straightforward code.

They answer recurring customer questions.

They turn one format of information into another.

These are precisely the kinds of tasks on which language models have improved rapidly.

Stanford's 2026 revision also found a distinction between codified and tacit knowledge.

Employment weakness among young workers was more visible in occupations relying heavily on formalised, documented knowledge. More experienced workers appeared stronger in roles drawing heavily on tacit knowledge acquired through practice, mentorship and repeated exposure to real situations.

That is an intuitively important distinction.

AI has consumed enormous quantities of material that human beings have already written down.

Experience is harder to upload.

Entry-level work has a second purpose: creating senior workers

A junior task can look inefficient if examined only as today's unit of production.

It may also be training.

A senior analyst does not become senior without analysing ordinary cases.

An experienced journalist does not learn how to recognise a suspicious document without reading many unsurprising ones.

A lawyer develops judgment partly by drafting, researching and observing decisions repeatedly.

A software engineer learns architecture after writing, breaking and repairing less glamorous code.

This is where organisations face a problem that will not appear neatly in next quarter's productivity figures.

If AI removes too much of the work through which beginners develop expertise, companies may eventually discover that they have made junior employees less necessary while making experienced employees harder to replace.

Global exposure is broad, but exposure is not unemployment

The International Labour Organization's 2025 global assessment reached a similarly cautious conclusion from a different direction.

It estimated that roughly one in four workers worldwide is employed in an occupation with some degree of exposure to generative AI.

Only 3.3 per cent of global employment fell into its highest exposure category.

Exposure also rises substantially in richer economies, where office, professional and highly digitised occupations represent a larger share of employment.

Clerical work remains among the most exposed.

But the ILO's central conclusion was not that one quarter of humanity is about to lose its job.

It argued that transformation is more likely than complete replacement for most occupations because jobs still contain tasks requiring human input.

ILO global assessment

One worker can be highly exposed without being easily replaceable.

Occupational exposure measures whether AI can perform significant tasks within a job. It does not demonstrate that an employer will automate the entire occupation, that deployment is economically worthwhile or that human responsibility can be removed.

That distinction is essential when reading percentages about "jobs affected by AI."

Jobs are bundles of tasks

This may be the single most useful way to understand what is happening.

A job title hides enormous variation.

An accountant may inspect documents, speak to a client, apply tax rules, enter figures, prepare a narrative explanation and take responsibility for a filing.

A doctor may review records, speak with a patient, interpret symptoms, communicate risk, make a diagnosis and accept responsibility for a decision.

A programmer may write repetitive code, understand an undocumented system, negotiate requirements, investigate a production failure and decide whether a technically elegant solution is worth deploying.

AI may be excellent at one part and unreliable at another.

That changes the value of the bundle.

Automation and augmentation can happen at the same time

Companies frequently describe AI as augmenting employees.

Critics describe the same technology as automation.

Both can be correct.

Suppose software allows a worker to produce twice as many first drafts.

That employee has been augmented.

If the employer consequently decides that four workers can handle output that previously required six, part of the process has also substituted for labour.

If cheaper production doubles customer demand and the company hires eight workers instead, productivity has expanded employment.

The technology alone does not tell us which outcome occurs.

Demand, competition, prices, regulation and management decisions all sit between technical capability and employment.

One real experiment shows why the productivity story is complicated

A widely cited study by Erik Brynjolfsson, Danielle Li and Lindsey Raymond examined the rollout of a generative-AI assistant to 5,179 customer-support agents.

Workers with access to the system increased the number of issues they resolved per hour by about 14 per cent on average.

The improvement was much larger for novice and lower-skilled workers, around 34 per cent, while highly experienced workers gained much less.

The system appeared partly to spread the practices of stronger workers to less experienced colleagues.

5,179

Customer-support agents included in the productivity study.

+14%

Average improvement in issues resolved per hour after access to the AI assistant.

+34%

Improvement reported for novice and lower-skilled workers.

That sounds like an optimistic result for workers.

It may be.

But a productivity study does not tell us how a company responds once the productivity gain becomes normal.

Management can improve service.

It can reduce prices.

It can increase output.

It can employ fewer people.

Or it can do some combination of all four.

The important question is not only what AI can do. It is what organisations decide to do after it can.

Customer service offers a preview of task compression

Customer service has been targeted by automation for decades.

The difference with modern generative systems is that they are less dependent on rigid menus and pre-written scripts.

They can classify unusual wording, summarise previous conversations and draft responses to situations that have not been typed exactly the same way before.

The economic attraction is obvious.

If software resolves a larger share of routine questions, fewer human agents may be required for the same volume.

But the work that remains may be more difficult.

Humans increasingly receive the angry customer, the unusual billing dispute, the ambiguous policy case and the situation where a machine has already failed.

Automation can therefore increase the skill required of the remaining worker while reducing the number of routine cases that once made the job easier.

Software development presents a similar paradox

Coding systems can now generate boilerplate, suggest fixes, explain unfamiliar functions and produce working first drafts.

They also produce incorrect assumptions, insecure code and solutions that appear plausible to someone who does not know enough to recognise the problem.

This creates a peculiar situation.

AI can make a beginner far more productive while simultaneously making genuine expertise more valuable.

If producing ordinary code becomes cheaper, employers may want fewer people whose only advantage is producing ordinary code.

Architecture, security, debugging, domain knowledge and responsibility can become a larger share of the profession.

The unresolved question is how new workers obtain those skills if fewer employers are willing to pay them to perform the simpler work first.

Creative work can survive while becoming harder to sell

Employment statistics can also miss another form of disruption.

A profession does not have to disappear for its economics to deteriorate.

Writers, illustrators, translators, photographers, designers and other creative workers increasingly compete with systems that can produce large quantities of acceptable material at extremely low marginal cost.

Premium human work may remain valuable where clients care about authorship, originality, accountability, relationships or a recognisable personal style.

Commodity work can become much harder to price.

Somebody can remain technically employed while earning less per project and competing for fewer projects.

A simple count of employed people would not capture the whole change.

Employers themselves still do not agree on the outcome

Corporate surveys reveal how unsettled the response remains.

Microsoft's 2025 Work Trend Index found that 33 per cent of surveyed leaders were considering reducing headcount as AI changed workflows.

In the same survey, 78 per cent were considering hiring for new AI-related roles.

Those findings are not contradictory.

A company can remove one category of work and create another.

It can reduce a team of routine operators while hiring AI engineers, security specialists, workflow designers or people capable of supervising automated systems.

The net outcome depends on the numbers on either side.

More importantly, employer expectations are evidence of what managers intend to do, not proof of what the entire economy will eventually do.

The 2026 workplace is already becoming more supervisory

Microsoft's 2026 Work Trend Index surveyed 20,000 workers using AI across ten countries and analysed anonymised workplace usage.

Two-thirds of the AI users surveyed said the technology allowed them to spend more time on higher-value work.

Fifty-eight per cent said they were producing work they could not have produced a year earlier.

The most revealing result may be about responsibility rather than speed.

Eighty-six per cent said they treated AI output as a starting point rather than a final answer and remained responsible for the thinking.

The data come from Microsoft's own ecosystem and survey population, so they should not be mistaken for a neutral measurement of the entire global labour market.

They do illustrate how the worker's role can move from producing every component manually towards directing, evaluating and correcting machine output.

Judgment becomes valuable when production becomes cheap

This is a recurring feature of technological change.

When the cost of producing something falls, value can migrate elsewhere.

If drafting text is cheap, knowing what should be written becomes more important.

If generating code is cheap, knowing whether it should be deployed becomes more important.

If producing an image is cheap, taste and originality may become more important.

If summarising information is cheap, finding the right information and knowing whether it is trustworthy becomes more important.

This does not prove that judgment is immune to automation.

It explains why technical capability at one task does not immediately remove the surrounding occupation.

Experience may become more valuable before it becomes less valuable

The Stanford findings on experienced workers are particularly interesting.

So far, the pronounced employment weakness visible among young workers in highly exposed occupations has not appeared in the same way among older, experienced workers.

That could be temporary.

It could also reflect the fact that experienced workers hold knowledge that is not contained neatly in written procedures.

They know which client is likely to change their mind.

They recognise when an apparently ordinary case is actually unusual.

They know which shortcut causes problems six months later.

They have lived through enough failures to distrust an answer for reasons they may struggle to formalise.

That is tacit knowledge.

A machine trained predominantly on recorded information can be enormously capable while still being weaker at knowledge that organisations never wrote down.

But experience creates a pipeline problem

This brings us back to younger workers.

Today's twenty-three-year-old is tomorrow's experienced worker only if someone employs them long enough to become one.

An economy cannot preserve a permanent class of senior professionals while deleting the process by which senior professionals are created.

Employers may eventually have to redesign entry-level jobs rather than eliminate them.

A junior employee may spend less time producing routine material and more time checking AI output, interacting with clients, handling edge cases and learning the context that software lacks.

That would be a transformed apprenticeship rather than the end of one.

Whether companies actually make that investment is another question.

The danger of allowing AI to replace practice

Workers themselves face a related temptation.

AI can help a beginner produce output that looks dramatically more advanced than their underlying skill.

That is useful.

It can also conceal the fact that the person has not learned how the work is done.

Someone who asks a model to write every first draft may complete more assignments while developing less intuition about writing.

A programmer can generate working code without understanding why it works.

A student can submit a plausible explanation without learning the subject well enough to recognise a plausible error.

Productivity today and capability five years from now are not always the same objective.

For workers

The practical response is neither panic nor pretending nothing is happening.

  • Learn where AI is genuinely useful in your occupation rather than only where it looks impressive.
  • Understand the underlying work well enough to recognise when automated output is wrong.
  • Develop domain knowledge that is difficult to obtain from generic information alone.
  • Build relationships and trust where the job depends on more than producing information.
  • Practise judgment, verification and responsibility rather than outsourcing all of them to a tool.
  • Watch actual hiring behaviour in your field, not only predictions about what software may eventually do.

Productivity does not automatically mean fewer jobs

There is another reason to resist simple arithmetic.

Suppose AI cuts the cost of producing a service in half.

If demand remains unchanged, fewer workers may be required.

But lower prices can create new demand.

Smaller companies may suddenly afford analysis that previously belonged only to large corporations.

More businesses may commission software because development has become cheaper.

Individuals may consume services that were previously too expensive to provide personally.

New demand can absorb productivity gains.

Sometimes it can create more employment than existed before.

That is why one hour saved by software should not automatically be counted as one hour removed from the labour market.

Distribution matters as much as total productivity

Even an optimistic productivity outcome does not guarantee an optimistic outcome for every worker.

An economy can become richer while a particular occupation becomes poorer.

A company can become more productive while bargaining power moves away from employees.

Consumers can benefit from lower prices while the workers who previously provided the service lose income.

Owners of scarce technology can capture a large part of the gain.

These are distributional questions, not arguments against productivity itself.

Technology can create value while politics, institutions and bargaining determine who receives it.

AI exposure is unusually concentrated in educated work

Earlier waves of automation often centred public anxiety on factories and routine manual work.

Generative AI feels different because it can operate directly on language, software, images and other forms of knowledge work.

IMF research has estimated that approximately 60 per cent of employment in advanced economies is exposed to AI in some form.

Exposure does not mean replacement.

In many occupations AI may complement workers rather than substitute for them.

But it does explain why professional workers who once viewed automation as somebody else's problem are now watching it happen inside software they already use.

The same technology can widen and narrow inequality

The customer-support experiment offers one optimistic mechanism.

Less experienced workers benefited much more than highly experienced workers.

If AI gives beginners access to effective practices that once took years to learn, it can narrow productivity gaps.

But if employers respond by hiring fewer beginners, the same technology can narrow the gap between workers who already have jobs while widening the gap between those inside and outside the profession.

This is why "AI helps junior workers" and "AI threatens junior hiring" are not necessarily contradictory findings.

One concerns productivity after a worker is employed.

The other concerns whether the worker is employed at all.

Company announcements should be treated as evidence of intention

Corporate leaders regularly announce that AI will allow smaller teams, greater productivity or entirely new forms of organisation.

Those announcements matter because managers act on beliefs.

They should not be confused with final labour-market evidence.

A company may discover that a planned automation performs worse than expected.

Another may become more efficient and expand.

A third may cut workers for conventional financial reasons while describing the change as an AI transformation.

A fourth may automate successfully and create a genuine long-term reduction in labour demand.

The headline "CEO says AI will replace jobs" tells us what the CEO believes.

It does not settle what happens next.

Organisations are slower than demonstrations

Software demonstrations create a distorted sense of speed.

A model can generate a convincing result in thirty seconds.

A large organisation may need years to rebuild a workflow around it.

Companies have old databases, security rules, regulators, contracts, unions, procurement departments and customers who expect existing systems to continue working.

A tool can be technically capable of a task long before an organisation trusts it enough to remove the person doing that task.

This friction slows change.

It does not guarantee permanent protection.

Human responsibility can slow automation too

Some work is not difficult because producing an answer is difficult.

It is difficult because somebody must be responsible for the consequence.

A medical system can suggest a diagnosis.

A person may still need to decide whether treatment begins.

Software can draft a legal argument.

A lawyer may still be accountable to the client and court.

AI can produce a financial recommendation.

A regulated institution may still need an identifiable person to approve it.

Accountability is therefore an economic feature of work, not simply an ethical decoration.

History gives perspective, not a guarantee

Every technological revolution attracts historical comparisons.

Computers automated enormous amounts of clerical work.

Industrial machines transformed manufacturing.

The internet destroyed some occupations, weakened some business models and created others that barely existed before.

None of those examples proves that generative AI must have the same employment outcome.

AI interacts directly with cognitive tasks in ways many earlier technologies did not.

The more useful historical lesson is that economies adapt in aggregate while particular workers can experience extremely painful transitions.

"Technology eventually creates new jobs" is little comfort to someone whose occupation disappears twenty years before the new opportunity reaches them.

What we know in September 2026

We know generative AI can already perform economically useful work.

We know controlled studies have measured substantial productivity improvements in some settings.

We know a large share of workers in richer economies occupy jobs exposed to at least some of its capabilities.

We know younger US workers in highly AI-exposed occupations are showing a worrying employment divergence in one major payroll dataset.

We know that divergence appears more strongly in hiring than in firings.

We know employers are actively redesigning workflows around AI.

What we do not know

We do not know whether the young-worker employment gap will continue widening, stabilise or reverse.

We do not know how much of it will ultimately be shown to have been caused directly by AI.

We do not know how quickly organisations will trust agents with entire workflows.

We do not know which new occupations will become large enough to absorb displaced workers.

We do not know whether productivity gains will flow primarily to workers, consumers, shareholders or some combination of all three.

We do not know whether cheaper intelligence will reduce total labour demand or create so much new economic activity that employment expands elsewhere.

The uncertainty is not an excuse to ignore the evidence

There are two easy positions available.

One is to declare that mass unemployment is inevitable because AI can perform impressive tasks.

The other is to point out that employment remains high and conclude that there is nothing to worry about.

The evidence supports neither position.

There is no economy-wide AI jobs collapse.

There are credible signs of pressure in specific parts of the labour market.

That is exactly the kind of period in which good measurement matters most.

By the time disruption becomes obvious in a headline unemployment number, the earlier changes in training, hiring and career progression may already have been happening for years.

Workers are being asked to prepare for a future that employers cannot yet describe

This is perhaps the strangest feature of the current transition.

People are told to "future-proof" themselves while nobody can say with confidence which version of the future will arrive.

Learn AI, but do not become dependent on it.

Develop expertise, but recognise that some traditional routes into expertise may shrink.

Become more productive, while understanding that greater productivity may alter how many people an employer needs.

Build human skills, while machines continue expanding into tasks previously described as uniquely human.

That advice is unsatisfying because the situation itself is unsatisfying.

The job market is changing before the answer exists

Technology does not determine the labour market alone.

Employers decide whether a productivity gain becomes more output or fewer workers.

Customers decide whether lower prices create new demand.

Governments decide how education, tax, competition and social insurance respond.

Workers decide which tools to adopt and which skills to preserve.

Regulators decide where human responsibility must remain.

The capabilities of AI matter enormously.

What human institutions do with those capabilities matters just as much.

The job market is already changing.

The uncomfortable part is that workers have to live through the transition before anybody knows what the final version will look like.

Reporting note

Sources used for this analysis

This article draws on the Stanford Digital Economy Lab's August 2026 revision of its research using ADP payroll data, the International Labour Organization and NASK's global index of generative-AI occupational exposure, published productivity research by Erik Brynjolfsson, Danielle Li and Lindsey Raymond, International Monetary Fund research on AI exposure and occupational mobility, and Microsoft's 2025 and 2026 Work Trend Index research. Forecasts, exposure estimates, employer intentions and observed employment effects are treated separately throughout the article.

If you believe this article contains a factual error, visit our corrections page .

Portrait of Michael Frisch

Michael Frisch

Michael Frisch is the founder of Crazy News and a senior reporter and analyst. His work focuses on institutions, politics, economic change, technology, human rights and the longer historical context behind claims about how society is changing.

Reporter profile