Start with three different questions
Discussion of AI and jobs becomes confusing when three measures are treated as interchangeable. Exposure asks whether a technology could perform some tasks in an occupation. Observed effects ask whether employment, hours, wages, hiring, or work quality have actually changed. Forecasts estimate what might happen under assumptions about capability, adoption, demand, regulation, and the broader economy. A job can be highly exposed without disappearing: spreadsheets transformed accounting, for example, while increasing demand for some accounting work.
The strongest global exposure estimate available in 2025 came from the International Labour Organization and Poland’s NASK institute. Their task-level index classified about 25% of global employment as having some generative-AI exposure, but only 3.3% in the highest-exposure category. The authors concluded that transformation was more likely than full replacement because most occupations combine automatable and non-automatable tasks. Exposure was higher in high-income economies and for women, largely because clerical employment is distributed unevenly. Those are modeled estimates, not a count of jobs already lost (ILO–NASK, May 2025).
What has actually been observed
As of September 2026, evidence of economy-wide mass unemployment caused by generative AI is not established. Adoption is substantial but uneven, and official employment statistics do not identify a clean “AI-caused” category. Companies may invoke AI while also cutting jobs because of interest rates, over-hiring, outsourcing, or ordinary restructuring. Announced layoffs are therefore weaker causal evidence than a study comparing similar workers or firms with different AI exposure.
The more credible near-term pattern is reallocation within jobs and hiring pipelines. Field experiments have found productivity gains in bounded tasks such as customer support and writing, often with larger gains for less-experienced workers. That does not imply the same effect across an occupation: faster output can reduce headcount, increase demand, raise quality, or simply increase the volume of work. Stanford’s 2026 AI Index reports that labor-market effects are uneven and concentrated in some exposed occupations and younger-worker hiring pipelines, but this evidence is still developing rather than proof that an entire generation has been displaced (Stanford HAI, 2026).
Clerical tasks remain among the most exposed, while exposure has expanded into digitized professional work such as programming, analysis, and finance. Physical presence does not make a job permanently safe; it merely limits what today’s text-and-screen systems can do without robotics — a gap this manual’s state of humanoid robotics page surveys directly. Licensed judgment and interpersonal trust can also slow substitution, although institutions may still automate preparation, triage, documentation, and monitoring around those roles.
Correcting the headline forecast
The World Economic Forum’s Future of Jobs Report 2025 estimated 170 million jobs created and 92 million displaced by 2030, for net growth of 78 million. Those numbers cover five macrotrends together: technological change, geoeconomic fragmentation, the green transition, demographic change, and economic uncertainty. They are not an estimate of jobs created or destroyed by AI alone. Within the report’s employer expectations, AI and information-processing technology were associated with 11 million jobs created and 9 million displaced. The methodology combines a survey of more than 1,000 large employers with ILO employment data, so it represents employer expectations projected globally, not a measured outcome (WEF, January 2025). WEF is an employer-led organization, a perspective readers should keep in mind when interpreting its survey.
Consultancy estimates often measure “hours technically automatable.” They can be useful for scenario planning but are not forecasts of unemployment. Deployment depends on reliability, workflow redesign, legal responsibility, integration costs, labor bargaining, customer preference, and whether automation makes a service cheap enough to expand demand. A model that can perform 60% of listed tasks in a demonstration has not necessarily automated 60% of a worker’s paid role.
What an AGI transition could change
Everything above concerns current generative AI. AGI labor effects are projections, because no agreed test has established that AGI exists. A system able to learn new domains reliably, plan over long horizons, operate tools, and improve workflows could expose a much wider range of cognitive work. If robotics also became cheap and dependable, physical work could follow. Conversely, regulation, ownership structures, human preference, and complementary demand could preserve or create work even with much stronger systems.
The key uncertainty is not just capability but the distribution of gains. Rapid productivity growth can coexist with falling labor income if capital owners capture most returns. It can also finance higher wages, shorter workweeks, public services, or social dividends if institutions distribute the gains. “Will there be jobs?” is therefore less complete than asking who owns the systems, who can bargain, who bears transition costs, and whether households retain income and agency.
Unequal exposure and second-order effects
Transitions are rarely uniform. Administrative centers may be affected before regions dominated by construction or care work. Entry-level tasks may shrink even when senior roles persist, weakening the route by which novices become experts. Workers with disabilities may gain valuable assistance while also encountering automated screening and surveillance. Lower-income countries may gain access to expertise, yet lose outsourced service work on which local economies depend.
Measurement also misses job quality. AI can remove drudgery, but it can intensify work through monitoring, algorithmic targets, and expectations of constant output. A stable employment count can conceal wage compression, fewer hours, degraded autonomy, or a shift from staff employment to precarious contracting. Conversely, it can conceal new accessibility and entrepreneurship benefits. A responsible dashboard should track hiring, wages, hours, vacancies, task content, worker control, and distribution—not layoffs alone.
A practical household reading
Do not make an irreversible career decision from a single forecast. Treat exposure as a prompt for due diligence:
- List the tasks for which you are paid, including relationship, accountability, and physical-context work that job titles omit.
- Test current tools on real tasks and record their reliability, supervision burden, and failure costs.
- Watch changes in your employer’s hiring, procurement, and workflow—not just executive statements.
- Build portable evidence of judgment, domain knowledge, client trust, and AI-assisted productivity.
- Reduce household concentration in one employer or narrow task bundle where feasible.
The honest conclusion in September 2026 is neither “nothing is happening” nor “most jobs are about to vanish.” Current evidence shows substantial task exposure, uneven adoption, productivity gains in bounded settings, and early disruption in some hiring channels. Large net-employment claims remain forecasts. AGI-scale displacement is a plausible scenario whose magnitude depends as much on deployment and political economy as on model capability.