Which Jobs Are Safest and Riskiest From AI? The Most Authoritative Research of 2026

The Research Has Finally Caught Up to the Reality

For years, AI job displacement forecasts were largely theoretical. Researchers modeled what AI could do and mapped those capabilities onto occupational task lists, producing exposure scores that told you which jobs were conceptually vulnerable. What they could not tell you was what was actually happening in the real world, in real workflows, right now.

That gap closed meaningfully in 2026. Three major research outputs published between March and April alone -- Anthropic's Labor Market Impacts study, Goldman Sachs's April displacement analysis, and a large-scale worker survey also from Anthropic -- have created the most data-grounded picture of AI's job impact to date. Combined with projections from the World Economic Forum, the International Monetary Fund, and the U.S. Bureau of Labor Statistics, a coherent and in some ways surprising picture is now visible.

The central finding: AI is not displacing workers in the order most people expected. The jobs most exposed right now are not factory floors or call centers staffed by low-wage workers without degrees. They are well-paid, highly educated, majority-female knowledge roles -- computer programmers, financial analysts, market researchers -- where AI has already taken over a significant share of actual daily tasks. Meanwhile, the so-called 'low-skill' jobs requiring physical presence are, by the data, the most protected.

Observed Exposure: The Metric That Changes the Conversation

The most methodologically significant contribution of 2026 came from Anthropic researchers Maxim Massenkoff and Peter McCrory, whose March 2026 paper introduced a measure called 'observed exposure.' Unlike earlier AI impact studies that scored occupations based on theoretical AI capability, observed exposure tracks which tasks are actively being automated in real professional workflows, using anonymized, privacy-preserving data from Claude usage across thousands of companies and individual users.

The distinction is not trivial. Theoretical exposure tells you what AI could do; observed exposure tells you what employers and workers are actually delegating to AI today. The gap between the two is still large in most sectors, but in a handful of occupations, the observed rate has grown close enough to the theoretical ceiling to represent a structural shift.

Computer programmers sit at the top of the observed exposure ranking at 74.5% -- meaning nearly three-quarters of their core job tasks are already being handled by AI in real workplace settings, not in labs or hypothetical scenarios. Customer service representatives follow at 70.1%, then data entry keyers at 67.1%, medical record specialists at 66.7%, market research analysts at 64.8%, wholesale and manufacturing sales representatives at 62.8%, financial and investment analysts at 57.2%, software QA analysts at 51.9%, information security analysts at 48.6%, and computer user support specialists at 46.8%.

The demographic profile of workers in these exposed roles is just as striking. They earn 47% more on average than the least-exposed group, with a median hourly wage of $32.69 versus $22.23. They are nearly twice as likely to hold a graduate degree (17.4% versus 4.5%), and they are 16 percentage points more likely to be women. The assumption that AI automation primarily threatens economically precarious workers is not supported by this data.

The Safest Jobs Right Now Are Not the Ones People Expected

Anthropic's analysis of theoretical AI coverage -- what AI could potentially do across all tasks in a given field -- is also useful for understanding the floor of risk. Ground maintenance workers have the lowest theoretical coverage at just 3.9%, meaning fewer than 4% of their core job tasks even fall within the range of what an LLM could plausibly handle. Transportation comes in at 12.1%, agriculture at 15.7%, food and serving at 16.9%, construction at 16.9%, personal care at 18.2%, installation and repair at 18.4%, and production at 19%. Every one of these sits below 20% theoretical exposure -- and actual adoption is far lower still.

A separate analysis of 784 occupations from education research group EdSmart reached a consistent conclusion: construction and extraction roles account for 66% of the top 50 safest jobs from AI, followed by building and grounds maintenance at 10% and hands-on healthcare practitioners at 8%. The structural reason is straightforward. These roles require physical dexterity, on-site problem-solving, and real-time adaptation to unpredictable environments. A plumber diagnosing an unusual problem in a century-old house or a lineman working live electrical systems in a storm cannot be replaced by a system that outputs text.

The labor market is reinforcing this. According to data from Metaintro, which tracks active hiring across the US, approximately 500,000 skilled trade positions sat unfilled in 2026, driven by a decades-long collapse in vocational enrollment, a retiring boomer workforce, and accelerating infrastructure demand tied to federal spending and AI data center construction. Electricians now earn a median of $62,350 per year with top earners clearing $106,000, and BLS projects 9% job growth through 2034 -- faster than average. Plumbers, HVAC technicians, and welders sit in a median band of $51,000 to $63,000, with experienced specialists regularly reaching $80,000 to $120,000.

Hands-on healthcare roles -- surgical nurses, physical therapists, home health aides, paramedics -- share comparable structural protection. While AI can assist with diagnosis or imaging analysis, it cannot replicate the physical, emotional, and context-specific demands of direct patient care. Goldman Sachs's framework explicitly classifies physicians and construction managers as high augmentation potential rather than high substitution risk, because human judgment and physical presence remain essential even as AI handles adjacent tasks.

Goldman Sachs Puts a Number on It: 16,000 Jobs Per Month

In April 2026, Goldman Sachs Global Investment Research published the first major Wall Street attempt to isolate AI's actual contribution to US job losses from broader economic trends. Their estimate: AI-driven automation is now responsible for the net displacement of approximately 16,000 US jobs per month, or roughly 192,000 positions annualized.

The methodology involved three layers -- occupational task automation scoring based on real deployment data (not just theoretical capability), quarterly employer surveys tracking whether headcount reductions were directly attributed to AI tool deployment, and a macro model to separate AI-driven displacement from offshoring and cyclical factors. Goldman's updated 2026 figure for the share of US work tasks automatable by AI is 34%, up from 25% in their 2023 analysis.

The breakdown within that 16,000 net figure matters: AI substitution eliminates roughly 25,000 jobs per month. AI augmentation -- creating new roles or expanding demand -- adds back approximately 9,000. The gap is real and so far growing. Office and administrative support has the highest sector-level task automation share at 46%, followed by legal at 44% and architecture and engineering at 37%. In March 2026 alone, the outplacement firm Challenger, Gray and Christmas counted 15,341 layoffs in which companies explicitly cited AI as the reason, representing 25% of all US job cuts that month and the single leading cause.

Goldman was careful to note that the aggregate long-term impact is likely smaller than these near-term figures suggest, partly because their model does not fully account for the hiring surge tied to AI infrastructure buildout in data centers, power systems, and construction. Their longer-term estimate is that AI will ultimately displace roughly 6 to 7% of the US workforce, approximately 11 million workers, while the productivity gains it enables tend to expand markets and generate new categories of employment.

The Entry-Level Collapse and Who Is Taking the Hardest Hit

The most immediately damaging effect of AI on the labor market may not be visible in unemployment rates. Anthropic's March 2026 paper found no systematic increase in joblessness among workers in heavily exposed occupations since late 2022. What it did find -- and what Goldman Sachs corroborated -- is a pronounced slowdown in hiring of entry-level workers into high-exposure roles.

Anthropic found that hiring of workers aged 22 to 25 into high-exposure occupations has slowed by approximately 14% since ChatGPT launched. Goldman's regression model found that in the most AI-substituted occupations, the unemployment rate gap between workers under 30 and workers aged 31 to 50 has widened sharply relative to pre-pandemic averages -- and that a one standard-deviation increase in AI substitution exposure widens the entry-level-to-experienced wage gap by roughly 3.3 percentage points. Cornell University research put the figure at a 13% reduction in junior hiring at US companies that adopted AI.

The reason is structural: senior workers in high-exposure roles carry accumulated judgment, context, and client relationships that AI does not yet replicate well. Entry-level workers, concentrated in the structured, rule-based tasks that form the on-ramp into knowledge careers, find themselves competing directly with AI for the same work -- and companies are choosing the cheaper option. The front door is narrowing without the back door being replaced.

The gender dimension is equally important and consistently underreported. Brookings Institution research found that 6.1 million US workers sit at the intersection of high AI exposure and low capacity to adapt, and 86% of them are women. The World Data analysis adds that 79% of employed US women work in jobs with high automation risk, versus 58% of men. This gap reflects occupational concentration: the administrative, clerical, and customer service roles AI is automating most aggressively are disproportionately held by women. The roles growing fastest -- AI engineering, cloud architecture, cybersecurity -- currently have among the lowest female representation in the industry.

The Global Picture: Destruction, Creation, and the Transition Gap

The macro-level projections from global institutions frame AI's labor impact as net positive over the medium term, but with a transition gap that will be unevenly and acutely felt. The World Economic Forum's Future of Jobs Report 2025 projects 92 million roles displaced globally by 2030 and 170 million new ones created, for a net gain of 78 million jobs. The IMF estimates that approximately 40% of global jobs carry some AI exposure, rising to 60% in advanced economies -- and that roughly half of the exposed roles in advanced economies are likely to benefit through productivity gains while the other half face reduced labor demand.

The jobs being created are real and pay well. AI engineering roles now average $170,750 per year in the US, 17.7% above non-AI peers. Machine learning engineers average $186,067. AI job postings are 134% above 2020 levels. Demand for AI governance skills is up 150%, AI ethics roles up 125%, and prompt engineering positions up 90%. But the WEF's own analysis shows that the fastest-growing occupations globally by raw headcount are farmworkers, delivery drivers, care workers, and educators -- roles driven by demographic shifts and the green transition, not by AI itself.

The 81,000-person survey published by Anthropic's economics team adds the worker perspective. One in five respondents expressed concern about economic displacement. People in highly exposed occupations -- as defined by actual Claude usage data -- were measurably more nervous. Early-career workers expressed significantly more concern than senior workers, aligning with the hiring slowdown data. Yet the same survey found that the mean productivity impact of AI was rated as 'substantially more productive' across the user base, and that productivity gains were most pronounced in high-wage roles -- the same roles that are most exposed.

What the Data Actually Tells Workers and Employers

The convergence of 2026 research produces a set of conclusions that are more specific and less panicked than the headlines suggest. First, mass unemployment from AI has not arrived. Anthropic's framework found no statistically significant spike in unemployment among the most AI-exposed occupations. Goldman's estimate of 16,000 net jobs lost monthly is real and growing, but against a US labor force of roughly 168 million, it represents a manageable rate of change -- for now.

Second, exposure does not equal displacement. Anthropic's chief economist Peter McCrory has been consistent on this point: the gap between what AI theoretically could automate and what is actually automated in practice remains wide in most sectors. What high exposure signals is that growth will slow, entry hiring will thin, and the composition of work will shift. That is not the same as immediate job loss, but it is a serious structural change that compounds over years.

Third, the protection strategy is clearer than it has ever been. Goldman Sachs's framework explicitly categorizes roles high on 'augmentation potential' as the durable positions: lawyers, construction managers, physicians, senior engineers -- jobs where AI handles portions of the workload while human judgment, physical presence, and accountability remain essential. Building deep domain expertise in a physically present or ethically complex field, then moving toward roles that direct AI rather than compete with it, is the path the data consistently supports.

Fourth, for anyone entering the workforce, the signal from the hiring data is unambiguous: the traditional entry-level on-ramp into knowledge careers is getting narrower. The roles that historically provided early-career workers with structured repetitive tasks -- tasks that built the experience base for advancement -- are precisely the ones AI is automating first. That does not mean knowledge careers are closed, but it does mean the path into them is changing faster than academic curricula or career advice has caught up.

Sources