AI and jobs: what the evidence says

AI can change work without eliminating an entire occupation. It can also weaken opportunities for particular workers before an economy-wide employment measure shows a large change. The evidence reviewed here supports taking both possibilities seriously.

It does not support turning an estimate of occupational exposure into a forecast of equivalent job losses. Nor does it justify treating a gain on one measured task as proof that AI will raise employment or wages across the economy. The responsible question is what each study actually measured, where, and over which period.

Our assessment: the strongest case for action today is to measure changing tasks, entry-level hiring and job quality alongside productivity. Broad claims of either inevitable mass unemployment or negligible disruption go beyond these selected findings.

Four questions hidden inside the jobs debate

“AI affects jobs” can mean at least four things:

1. Exposure: a technology has capabilities relevant to some tasks in an occupation.
2. Adoption: people or firms actually use it in their work.
3. Performance: using it changes the quality, speed or quantity of an output.
4. Labor-market outcomes: employment, hiring, earnings, hours or job quality change.

These categories require different evidence. A task-rating exercise cannot count redundancies. A usage survey cannot establish a productivity gain. A before-and-after employment pattern does not automatically isolate the effect of AI from the other changes happening at the same time.

Productivity itself compares output with inputs. The US Bureau of Labor Statistics distinguishes productivity at worker, company, industry and national levels.[1] A faster document draft is therefore not interchangeable with an increase in a firm's output per labor hour, let alone national productivity.

This distinction matters when deciding how to start using AI at work. A team can investigate one workflow without making an unsupported claim about the future of its whole workforce.

AI Chronicle

What the selected studies establish

The following map deliberately keeps the studies separate. They cover different populations, technologies and periods, and should not be averaged into a single “AI impact” number.

Sources and versions: ILO–NASK,[2] Brynjolfsson–Li–Raymond,[3] Humlum–Vestergaard,[4] Stanford–ADP,[5][6] and Yotzov and colleagues.[7]

Exposure is a reason to investigate, not a layoff count

The ILO–NASK index estimates that about one in four workers globally are in an occupation with some exposure to generative AI. Its highest exposure category covers 3.3% of global employment. Those are classifications of potential exposure, rather than counts of workers displaced.[2]

The index combines task information, worker assessments, expert input and model-based scoring. Its authors emphasize that occupations contain tasks requiring human input.

“transformation of jobs is the most likely impact of GenAI.”

— ILO–NASK, Generative AI and Jobs: A Refined Global Index of Occupational Exposure, excerpt (2025).[2]

That is an assessment of likely change, not a promise of job security. The useful next question is which tasks change and how the organization responds.

Productivity gains can be real and uneven

Brynjolfsson, Li and Raymond studied a staggered introduction of a conversational assistant among 5,172 customer-support agents. The published paper reports a 15% average increase in issues resolved per hour, with differing results by experience and skill.[3]

Less-experienced and lower-skilled workers improved speed and quality. The most experienced and highest-skilled workers had small speed gains and small quality declines. That pattern cautions against assuming one benefit for every worker.

The study establishes a result in that support workflow. It does not show a 15% gain in all office work, a 15% wage increase, or a requirement to remove 15% of staff. Those are different outcomes.

Changing work can precede changes in pay and hours

Humlum and Vestergaard's March 2026 paper links large adoption surveys to Danish administrative labor-market records. Their difference-in-differences estimates find null effects on earnings and recorded hours at worker and workplace levels, ruling out effects larger than 2% two years after ChatGPT's launch.[4]

At the same time, they report work reorganization, including AI-related content, oversight and integration tasks. Some adopters move into higher-paying occupations, but not in sufficient numbers to shift average earnings.

This is evidence of an early transition with limited measured effects on those outcomes. It is not proof that no tasks changed, or that future effects must remain small everywhere. The country, period and outcomes remain part of the result.

The entry-level warning deserves close attention

The August 2026 revision of Stanford's Canaries in the Coal Mine? uses an ADP payroll sample covering millions of US workers through June 2026. It reports employment among 22–25-year-olds in exposed occupations at about 19% below a path that kept pace with less-exposed peers. The authors find no comparable gap for experienced workers and no widespread economy-wide displacement in their analysis.[5]

The shortfall is primarily associated with reduced hiring, rather than increased separations. This distinction matters: a career path can narrow even if existing workers are not being dismissed.

The researchers explicitly warn against a causal reading:

“These are descriptive patterns, not causal estimates of the effect of AI.”

— Brynjolfsson, Chandar and Chen, authors' note accompanying the August 2026 revision.[6]

Their note discusses attenuation after accounting for education, some earlier divergent trends, and stronger gaps in the ADP sample than in national survey benchmarks. These issues affect interpretation and generalizability.

The result warrants monitoring entry-level opportunities. It does not establish that AI caused a 19% fall in all young people's employment, or that 19% of entry-level jobs were eliminated. Its comparison group and denominator cannot be dropped from the headline.

Forecasts are not observed employment changes

Firm Data on AI surveys nearly 6,000 executives at US, UK, German and Australian firms. The March 2026 version finds widespread adoption with limited reported effects over the preceding three years. Respondents anticipate larger effects over the following three years: average productivity growth attributable to AI of 1.4%, output growth of 0.8%, and employment reduction of 0.7%.[7]

The three numbers are respondents' expectations for their firms. They are not realized changes, annual growth rates or a forecast independently verified by the researchers. Reported past effects are also survey responses, not the same thing as a causal estimate from payroll records.

AI Chronicle

A forecast can help identify what managers plan for. It cannot establish what has already happened. Keep that distinction even when the expected numbers look modest or precise.

How efficiency can coexist with different hiring outcomes

There is no automatic conversion from a task-level time saving to a headcount decision. Demand for the output, remaining work, costs and how an organization allocates capacity all matter.

Hampole and colleagues' February 2026 working paper examines AI and machine-learning exposure over 2010–2023. It distinguishes task substitution from reallocation and firm-level expansion. Its findings include reduced demand for exposed tasks alongside modest overall employment effects and offsetting demand at adopting firms.[8]

Its period largely predates the recent generative-AI expansion. It helps explain the economic mechanisms; it should not be presented as a direct evaluation of today's assistants.

Consider three hypothetical support businesses that verify a reduction in the labor needed for routine replies:

• A business with stable demand may need fewer hours for that portion of its work.
• A growing business may use the capacity to handle more inquiries while keeping its team.
• A business with unresolved complex cases may move effort toward those cases and quality review.

These are possible responses, not predicted outcomes or study findings. The same measured task improvement leaves different decisions open. It also says little about whether the people doing the work receive higher pay, less pressure or more responsibility.

This is the missing step in many sweeping claims. Knowing that a tool can complete a task does not reveal whether it is used, whether the full workflow improves, or what an employer does with the result.

Job quality belongs in the evidence too

Counting jobs is necessary, but it is not sufficient to understand the experience of work. A role can retain its title and salary while its workload, autonomy or learning opportunities change.

The OECD's 2023 Employment Outlook discusses results from 2022 surveys in manufacturing and finance across seven countries. It reports positive worker experiences alongside risks involving privacy, work intensity and bias.[9] That evidence concerns the surveyed settings and period; it is not a description of every current AI deployment.

For a workplace assessment, ask whether the change affects review responsibility, workload targets, training access and the ability to challenge a decision. These are proposed assessment questions, rather than a score generated by the studies above.

Employment research also does not answer every governance question. Whether an organization should delegate a particular decision, how it documents oversight and what requirements apply are separate issues. The EU AI Act explainer addresses the regulatory framework; a productivity result should not be treated as a permission slip.

What workers and employers can do with this evidence

The findings support a practical response: investigate the changing work rather than labeling an entire occupation either safe or doomed.

Map tasks, including the work after generation

List what a role actually does. Include checking, exception handling, coordination and accountability—not just the part a demonstration completes quickly.

For a marketing role, that could distinguish generating headline options from deciding the campaign, verifying a claim, securing approval and evaluating the result. For a finance role, distinguish explaining an exception from checking the ledger and authorizing a transaction. These are illustrative task distinctions, not tested automation recommendations.

This makes the investigation specific enough to observe a change. “AI replaces marketing” is too broad to be a workable evaluation question.

Measure quality and the complete effort

Record the accepted output, the time spent preparing inputs, the review required and the corrections made. A faster first draft can coexist with a slower final deliverable.

Where AI changes who performs a task, record that transfer of work too. A saving for one person is not necessarily a saving for the team. The AI productivity measurement guide will develop that workflow-level assessment.

Track entry routes separately from existing roles

Record junior openings, actual hires, separations and the work used for training. A stable total headcount can hide a decline in new opportunities if more senior roles grow at the same time.

If routine assignments change, explicitly decide how new staff will learn the judgment needed to review the output. Do not assume that removing practice tasks automatically creates an alternative path to expertise. This is an organizational design question raised by the evidence, not a measured conclusion about a particular employer.

Make the decision auditable

Before changing staffing on the basis of a tool's performance, identify the evidence connecting the task improvement to the complete role, expected demand and required quality. State which parts are observed and which are assumptions.

This does not require waiting for every macroeconomic question to be settled. It requires matching a decision to evidence at the relevant scale. A bounded workflow trial and a workforce forecast have different evidence needs.

How to read the next “AI jobs” headline

Use six checks before accepting its conclusion:

An employment change still needs interpretation even when its source is credible. Is it a relative gap or an absolute decline? Is a hiring measure being described as a redundancy measure? Is an expectation being reported as an outcome?

Also look for uncertainty in the original report. A confidence interval, sensitivity analysis or limitation is part of the finding, rather than an optional qualification that can be dropped when sharing it.

Questions readers commonly ask

Does “one in four exposed” mean one in four jobs will disappear?

No. The ILO estimate concerns occupations with some potential task exposure. It does not measure realized losses or establish that entire occupations can be automated.[2]

Do small average effects mean there is no disruption?

No. An average can coexist with work reorganization, changes for a subgroup or offsetting gains and losses. Inspect the distribution and the particular outcome, rather than treating an average as everyone's experience.

Does the young-worker evidence prove that AI caused the decline?

The Stanford authors describe their August 2026 results as descriptive patterns. Their comparison is concerning, but their stated design does not establish how much of the divergence was caused by AI.[6]

Should workers learn AI tools now?

A reasonable response is to learn a relevant, checkable workflow and how to judge its output, alongside the domain knowledge the role requires. This is an editorial recommendation, not a guarantee of employment or higher earnings.

The next step

The evidence supports attention to changing tasks and opportunities. It does not settle a universal employment forecast. A useful analysis keeps capability, use, output and worker outcomes connected without treating them as equivalent.

Next, our examination of what workplace AI studies actually show will focus on study design, tasks and populations. That is where we will examine why a result can be valid in its study while failing to answer a broader claim.

Sources and evidence note

Research reviewed on 14 October 2026. This is a selected, critical evidence review, not a systematic review, meta-analysis or original data collection. It covers generative-AI exposure, defined workplace results, early labor-market evidence and one broader AI/ML study. It does not estimate a single global job-loss rate.

Working papers can change. Versions are identified below; data periods are preserved in the text. Study findings, researcher cautions and AI Chronicle's proposed workplace questions are kept distinct. No hands-on tool test, measured local productivity saving or employment forecast is claimed by AI Chronicle.