OpenAI Maps 148 Million US Jobs Into Four AI Transition Paths

OpenAI's new AI Jobs Transition Framework finds that roughly 18 percent of U.S. jobs face relatively high automation risk, while nearly half — 46 percent — show less immediate change from AI. Applied across 921 occupations covering approximately 148 million U.S. jobs, the framework splits the remainder into two groups: 24 percent of jobs likely to reorganize around AI, and 12 percent that could actually grow as the technology lowers costs and expands demand.

The numbers arrive at a moment when policymakers, employers, and researchers are arguing over whether AI will displace workers wholesale or quietly rewrite their job descriptions. OpenAI's answer is that the question "which jobs can AI perform?" is the wrong starting point.

"A technology may be capable of completing many of an occupation's tasks without eliminating the occupation itself," the paper's authors write. Humans may remain essential because customers want human interaction, institutions require human accountability, or the work demands physical presence and judgment. Lower costs may also increase demand enough that employment expands even as AI makes workers more productive.

Three questions instead of one

Instead of treating technical exposure as a forecast of displacement, the framework asks three questions of each occupation:

• Can AI perform a meaningful share of the occupation's tasks?
• Is a person still central to delivering, supervising, or taking responsibility for the work?
• If AI lowers the cost of the service, will demand grow enough to absorb the productivity gains?

The categories that result are not predictions that a particular percentage of jobs will disappear, the authors stress. "They are a map of where different kinds of change may emerge first."

That distinction matters for policy. A framework that predicts displacement invites one set of interventions — retraining, safety nets, transition assistance. A framework that predicts reorganization, demand growth, or stability invites very different ones, from updated professional standards to adoption incentives.

Four different paths for work

Jobs at higher automation risk. The roughly 18 percent bucket includes data-entry clerks, telemarketers, proofreaders, and some bookkeeping and administrative roles. Much of this work involves processing, checking, or communicating standardized information — tasks AI can increasingly perform without a person directly delivering the final service. Demand for these activities may not grow enough to offset the productivity gains.

These jobs deserve close monitoring, the authors argue, particularly where employment is geographically concentrated or workers have limited pathways into adjacent roles. But even here, technical capability does not translate automatically or immediately into displacement. Employers still need to redesign workflows, integrate AI with existing systems, manage errors and regulatory risks, and determine whether automation is actually cheaper than continuing to employ people.

Jobs that will reorganize. A second and larger group — 24 percent of jobs — is likely to reorganize rather than disappear. Lawyers, accountants, software developers, financial analysts, and teachers are plausible examples. AI can already draft contracts, summarize evidence, prepare financial analysis, write code, and create lesson materials. But clients and institutions still need people to exercise judgment, take responsibility for decisions, understand unusual cases, and build relationships with colleagues, students, or customers.

AI may allow each lawyer to review more documents, each developer to produce more code, or each teacher to prepare more individualized materials. That could reduce the labor required for some activities, but the occupations remain recognizably human. The central issue becomes how the jobs are redesigned: which tasks are delegated to AI, which remain with workers, and whether entry-level roles and career pathways continue to provide opportunities to learn.

Jobs that grow with AI. Twelve percent of jobs sit in fields where AI may lower costs enough to create more demand: tutors, mental-health counselors, personal financial advisers, and some health-care professionals. These services are currently expensive or difficult to access for many people. In these fields, greater productivity does not necessarily mean fewer workers.

More affordable tutoring could lead more families to purchase it. Lower-cost financial advice could make personalized guidance available to households that cannot currently afford an adviser. AI-supported clinicians might provide more follow-up and ongoing care. If demand expands enough, total employment could rise even as each worker becomes more productive.

Jobs with less immediate change. The largest group — 46 percent of jobs — includes many occupations whose central tasks remain physical and must be performed in a particular place: electricians, plumbers, roofers, construction laborers, and many food-service workers. AI may still help with scheduling, estimates, training, inventory, or customer communication, but it cannot currently install wiring, repair a pipe, replace a roof, or prepare and serve a meal.

These occupations will not be untouched by AI. Their administrative and managerial tasks may change, and advances in robotics could eventually broaden automation. But for now, their core work has relatively low exposure to language-based AI, making large near-term changes less likely.

Early evidence cuts both ways

The framework's most interesting claim is empirical: AI's early effects are likely to appear through changing tasks, workflows, and skill requirements before they appear as the wholesale disappearance of occupations.

The usage data supports the first half of that claim. Actual AI usage is already higher in occupations identified as facing greater automation risk. ChatGPT use is roughly three times as prevalent in the most at-risk occupations as it is across the workforce more broadly.

The unemployment data, however, does not line up neatly with technical exposure. Since the first quarter of 2024, unemployment has risen more in some less-exposed occupations than in the occupations the framework classifies as facing the greatest AI risk.

The authors are careful about what this does and does not prove. "This does not prove that AI is having no labor-market effect," they note. Occupational unemployment reflects many influences, and some changes may appear first in hiring, entry-level opportunities, wages, or the composition of work rather than in layoffs.

That asymmetry — heavy usage in at-risk jobs, no clear displacement signal yet — is exactly why the authors argue against reading technical exposure as destiny. The transmission belt from capability to job loss runs through employer decisions, integration costs, regulation, and demand, and each link can slow or redirect the effect.

A better map for policy

The paper's policy argument follows directly from its taxonomy. AI does not determine one inevitable future of work. Technical capability matters, but so do institutions, consumer demand, business decisions, regulation, and the enduring value of human participation. Different occupations, the framework suggests, require different interventions:

• Jobs facing high automation risk may call for early-warning systems, targeted adjustment assistance, and locally designed transition plans.
• Jobs likely to reorganize may require updated training, professional standards, staffing rules, and clearer expectations about human oversight.
• Jobs with the potential to grow may benefit from policies that encourage adoption, expand access, and prepare more workers to enter the field.

Across all categories, the authors identify a common gap: policymakers need better real-time information. Traditional labor statistics are valuable, but they often move too slowly to capture changes in tasks, employer expectations, and technology use. Combining employment data with measures of AI capability and actual adoption can provide a more timely picture of where pressure is accumulating.

The stakes extend beyond the United States. If the framework's core insight holds — that exposure, human centrality, and demand elasticity together determine outcomes — then the same 921-occupation mapping exercise could be run against any labor market, and the policy prescriptions would shift accordingly. The framework's real contribution is not the 18 percent figure. It is the demonstration that a single number cannot describe how 148 million jobs will meet a general-purpose technology.