AI Hasn't Created a “Useless Class”—But AI-Exposed Hiring Has Slowed

AI and robots have not produced a broad “useless class.” The stronger conclusion from evidence available in 2026 is narrower: aggregate unemployment has not risen systematically among highly AI-exposed workers, but young people are encountering early hiring pressure in occupations where AI can perform a substantial share of the work.

The important change is that this debate no longer rests entirely on forecasts. Researchers can now compare actual AI use with employment and payroll data. Those measurements remain preliminary and cannot isolate AI from every economic influence, but they reveal where the transition is appearing first: entry routes, routine digital tasks and jobs designed around producing standardized outputs.

The first signal is weaker hiring, not mass unemployment

A March 2026 analysis built an “observed exposure” measure from theoretical model capabilities, professional Claude usage and whether tasks were automated or completed with human participation. Anthropic’s labor-market research found no systematic increase in unemployment among the most exposed US workers since late 2022, although it found suggestive evidence of slower hiring for people aged 22 to 25.

In that analysis, the job-finding rate for young workers entering highly exposed occupations fell by an estimated 14% relative to its 2022 level. The result was only barely statistically significant, and the researchers identify several possible explanations: affected workers may have stayed in existing jobs, moved into less-exposed work, returned to education or been measured imperfectly in survey data. That is a warning signal, not proof of an economy-wide employment collapse.

A separate payroll-based study found a sharper pattern after controlling for changes within firms. In a February 2026 follow-up, Stanford Digital Economy Lab researchers said the relative employment decline for highly exposed workers aged 22 to 25 became notable in 2024 and reached about 16% by October 2025. They also stressed that the analysis cannot establish AI as the sole cause and that earlier declines were probably influenced by other factors.

Together, these findings describe a distributional problem. A technology can leave total unemployment broadly stable while making it harder for a particular generation to enter occupations such as programming, customer service or clerical work. Experienced employees may retain the organizational knowledge needed to supervise AI, handle exceptions and take responsibility, while employers reduce the junior tasks through which that knowledge was traditionally acquired.

Job forecasts do not predict one automated future

Large employment projections are often quoted as if every job gained or lost were caused by AI. That is not what the underlying estimates say. The World Economic Forum’s 2025 jobs outlook projected 170 million roles created and 92 million displaced by 2030—a net increase of 78 million—but attributed the churn to a combination of technology, demographic change, economic uncertainty, geoeconomic fragmentation and the green transition.

The report’s technology-specific estimates are more revealing. Surveyed employers expected AI and information-processing advances to create 11 million jobs and displace 9 million, while robotics and autonomous systems produced a projected net decline of 5 million. These remain expectations from employers, combined with labor data, rather than observed outcomes. The covered workforce of roughly 1.18 billion people is also a subset of global employment, so the headline totals should not be treated as a complete census of the future.

The likely unit of change is therefore the task, not the occupation. A role can survive while research, drafting, data entry, image production or customer triage becomes faster and requires fewer paid hours. Other parts of the same role—judgment, accountability, negotiation, physical execution and relationship management—may become more valuable because they remain bottlenecks.

Why creators face an especially uneven transition

Creative work exposes the weakness of the “jobs versus no jobs” framing. A designer, writer or video producer may remain employed while losing particular commissions to automated production. Another creator may use the same systems to offer more formats, test more concepts or serve clients who previously could not afford professional work.

The immediate risk is not necessarily the disappearance of the creator. It is the compression of rates for work that clients perceive as reproducible, combined with higher expectations for speed and volume. When a tool lowers the cost of generating a first draft, the market may pay less for the draft while still paying for a coherent campaign, defensible factual choices, access to an audience or responsibility for the final result.

That distinction also changes career entry. Junior creators have historically learned through research, rough cuts, simple layouts, transcription and initial copy. If those assignments are delegated to software, organizations can save time now while weakening the pipeline that produces future editors, creative directors and senior specialists. The missing apprenticeship function is a more concrete concern than the idea that an entire category of people will become permanently irrelevant.

Exposure is not the same as replacement

Three questions help separate genuine displacement risk from a dramatic capability demonstration:

• Does AI perform one task, most tasks or the complete chain of work required for the role?
• Can the output be deployed without human verification, legal responsibility or access to information outside the model?
• Does lower production cost reduce labor demand, or does it expand the amount and variety of work buyers commission?

A system may produce competent text or images while remaining unable to gather proprietary context, negotiate a brief, secure rights, evaluate reputational risk or accept liability. Conversely, a narrow back-office process can be highly vulnerable when its inputs are standardized, its success criteria are measurable and exceptions are rare. Occupational labels conceal these differences.

Real-world adoption matters too. Technical feasibility does not guarantee use: regulation, integration costs, customer preferences and verification requirements can slow deployment. But the opposite error is also dangerous. Stable headline unemployment can hide fewer vacancies, reduced freelance demand or workers moving into lower-paid occupations.

The central conflict is over access to productive work

The “useless class” label treats economic value as a permanent personal characteristic. The evidence supports a different interpretation: usefulness is assigned by institutions that decide how jobs are divided, who owns productive systems and how gains are distributed. People do not lose the capacity to contribute simply because a company stops purchasing a particular task from them.

For workers and independent creators, the practical response is to build around complements that remain scarce: domain knowledge, trusted relationships, editorial judgment, accountability and the ability to combine multiple tools into a reliable outcome. That does not guarantee protection, especially where automation removes entry-level opportunities, but it is more grounded than competing with software solely on the speed of producing a standardized asset.

For employers, preserving structured junior work is not charity. If every beginner task is automated without an alternative training path, firms may eventually face a shortage of experienced people capable of supervising systems and resolving unfamiliar cases. Apprenticeships can be redesigned around reviewing AI output, documenting failures, interviewing customers and handling exceptions rather than eliminated with the routine work.

The plausible near-term future is therefore neither universal redundancy nor frictionless abundance. It is a labor market in which overall employment can grow while particular occupations, age groups and income streams deteriorate. The decisive questions are who still gets a first opportunity, who captures the productivity gain and whether institutions create routes from displaced tasks into more valuable work.

Also read:

• Eric Adams Has Been Indicted, But His Crappy Subway Robot Will Be "Redeployed"
• AI Has Become Fintech Infrastructure—But Full Autonomy Remains Rare
• Introducing GPT-5.5: A New Class of Intelligence for Real Work and Powering Agents