Both Things Are True

In July 2026, more than 200 economists — including 16 Nobel laureates — signed a joint statement called “We Must Act Now,” warning that AI could reshape the economy at a speed and scale exceeding the Industrial Revolution, with risks that include large-scale job displacement. Among the signatories: Daron Acemoglu and Simon Johnson, two 2024 Nobel laureates whose earlier skepticism about AI’s disruptive potential made their names on that list striking.

That same season, the employment data told a quieter story: overall employment kept growing. Not despite AI — around it. The World Economic Forum projects that between 2025 and 2030, AI-driven restructuring will create 170 million jobs and displace 92 million, for a net gain of 78 million — 7% of today’s global workforce. That’s real churn — 22% of the 1.2 billion formal jobs in the dataset get touched — but it isn’t collapse.

Both of these things are true at once. The warning is real. So is the growth. What’s actually happening is narrower and stranger than either headline suggests: displacement concentrated at specific rungs of specific ladders, growth concentrated in specific rooms of the house. Eighty-six percent of employers now expect AI to transform their business by 2030 — the highest of any technology trend they were asked about. Almost nobody is planning around a world where nothing changes.

This piece exists to help you find your rung. Not to reassure you, not to alarm you — to show you what the data says about your specific corner of the map, and what to do about it.

If you only read one section, read this. Six moves the research backs, roughly in order of how fast they compound:

Jump to a section: What the data actually shows · The resilience formula · Careers on each side of the line · The jobs being born · What stays, what shrinks, what’s born · Where the experts disagree · Your playbook · How seriously to take this · Sources

The interactive below breaks this down by career, industry, and emerging job — find your field or browse by category.

  • Exposed
  • Pressured
  • Transforming
  • Resilient
  • Durable

Resilience score: our synthesis of OECD capability-gap, IMF complementarity, WEF growth projections, and Microsoft applicability data — a reading guide, not a prophecy.

What the Data Actually Shows

An amber canary holding a lantern at the mouth of a dark tunnel built from office cubicles
The Stanford payroll study is literally titled “Canaries in the Coal Mine” — the entry level is the labor market’s early-warning system.

Start with the clearest signal in the data: it’s the entry level that’s coughing, not the whole mine.

A Stanford Digital Economy Lab study of ADP payroll records — covering roughly 4.6 million workers across more than 730 occupations, about one in six American workers — found that since generative AI went mainstream, early-career workers aged 22-25 in the most AI-exposed occupations have experienced a 16% relative decline in employment, even after controlling for firm-level shocks. (An earlier cut of the data, covered by CNBC, put the figure at 13%; the published paper settles on 16%.) Employment for more experienced workers in those same exposed occupations — software development, accounting, customer service — has stayed stable or kept growing.

The decline isn’t slowing down. Stanford’s live “Canaries” dashboard showed the year-over-year drop for young workers in AI-exposed occupations accelerating from about 2.8% in April 2024 to more than 4% a year since, while the same age group in low-exposure occupations grew roughly 2% annually. As Stanford’s Erik Brynjolfsson put it: “Whatever it is, it’s not going away.”

The pattern shows up beyond payroll data, too. SignalFire’s 2026 State of Talent Report found entry-level tech hiring down roughly 65% at major tech companies and 76% at early-stage startups since 2019; computer science graduates from top-20 programs are 45% less likely to land a job at a major tech company than earlier cohorts. Meanwhile, US white-collar payrolls have contracted for dozens of consecutive months — a streak former Glassdoor chief economist Aaron Terrazas calls unprecedented outside a recession.

And yet the aggregate numbers flatten the story out. ADP chief economist Nela Richardson: “In the aggregate, AI’s impact on jobs remains modest. But when AI’s impact is measured by career stage, dramatic differences emerge.” Anthropic’s head of economics, Peter McCrory, said in March 2026 there’s “at least no larger material difference” in unemployment rates between AI-exposed and less-exposed occupations — no economy-wide collapse yet. And where adjustment happens, it runs through hiring, not pay: Stanford’s researchers found the market responding mostly by not hiring, rather than cutting wages.

One more distinction matters more than any other in this data: where AI augments work, employment holds. Where AI automates it outright, employment declines. That line — augmentation versus automation — turns out to be the closest thing to a unifying theory in everything that follows.

The Resilience Formula

A person and a robotic arm stacking blocks together on one workbench
Augmentation, not automation, is the throughline of everything below.

Four factors, showing up independently across multiple research groups, explain why some careers are holding the line while others erode.

Tacit knowledge beats codified knowledge. The Stanford researchers behind the entry-level findings attribute young workers’ vulnerability to a specific mechanism: AI is good at replacing “codified knowledge” — the book-learning from formal education — and much worse at replacing knowledge built from years of hands-on experience. That’s why employment for older workers in the same exposed occupations keeps growing while their 22-year-old colleagues’ doesn’t. Experience isn’t just a resume line. It’s currently the moat.

Augmentation beats automation. Anthropic’s most recent Economic Index (January 2026) shows augmented use — AI collaborating with a person — rising to 52% of Claude.ai conversations, overtaking fully automated use, which fell to 45%. This matters because Stanford’s payroll data shows employment declines concentrated in occupations where AI is used mostly to automate, with no clear negative relationship where it’s used mostly to augment. The question worth asking about your own job isn’t “does AI touch this” — it’s “does AI do this, or work alongside whoever does?”

Physical work in messy environments is still hard for AI. The OECD’s AI Capability Gap Index — how far current AI falls short of what a job actually requires — finds the largest, most stubborn gaps in manipulation (0.7) and robotic intelligence (0.6), the two most physical domains. Microsoft’s own applicability research, built from 200,000 anonymized Bing Copilot conversations, ranks orderlies, roofers, surgical assistants, massage therapists, and water-treatment plant operators among the occupations AI is least equipped to touch. Both groups converge: healthcare and hands-on blue-collar work are best positioned to withstand this wave, while the jobs “most at risk” are built on “providing information and assistance, writing, teaching, and advising.”

Responsibility and human relationships are still off-limits. The IMF’s framework adds a second axis to “does AI touch this job”: complementarity — whether AI enhances the role or replaces it. Surgeons, lawyers, and judges score high on both, because society isn’t going to accept unsupervised AI making those calls, however capable the model gets. The OECD finds the same pattern from a different angle: the biggest average capability gaps sit in social interaction, problem solving, and metacognition — the parts of work that are fundamentally about other people, not information.

Put those four together and you get a rough test for your own job: how much depends on experience you’ve already banked, on judgment rather than execution, on a body doing something in a room, or on someone trusting you specifically? The more boxes checked, the more resilient your position — for now.

Careers on Each Side of the Line

An office desk dissolving into pixels on one side of a dividing line; a nurse, a construction worker and a teacher standing solid on the other
Routine information work dissolves first; hands-on, human-facing work holds the line.

Fading. WEF’s declining-jobs list for 2030 reads like a catalog of codified, repeatable information work: postal service clerks, bank tellers, data entry clerks, cashiers and ticket clerks, administrative assistants and executive secretaries, graphic designers, telemarketers, legal secretaries, and accounting, bookkeeping, and payroll clerks. David Autor treats translation as already effectively displaced, and notes clerical and administrative support work has been shedding millions of jobs for decades — AI is accelerating a trend, not starting one. Customer service is following the same arc as chatbots absorb routine contact-center volume.

Transforming. Software engineering is the most interesting case in the dataset: simultaneously the fastest-growing job category by percentage (WEF) and one of the hardest-hit at the entry level (Stanford). It’s bifurcating: SignalFire’s data shows engineering hiring at large tech companies down only 11% since 2019, versus a 25% drop in overall tech hiring — and engineers now make up 55% of all new hires, up from 46% in 2019. Within engineering, front-end roles are down about 25% since ChatGPT’s 2022 launch, while AI/ML roles are up 39% and forward-deployed engineering roles up 30%. Finance shows a similar split — accountants and auditors on WEF’s declining list, fintech engineers on its fastest-growing one. Creative work is contracting hardest of all: design down 48% at major tech companies, marketing down 36%.

Enduring. The care economy tops every resilience ranking here: personal care, social work, and community services score the OECD’s single highest capability gap (6.4), and WEF projects nursing, personal care aides, and social work among the largest absolute-growth occupations through 2030. Skilled trades — construction, repair, installation — hold up even in the OECD’s forward-looking scenario, alongside healthcare, protective services, community/social service, and legal work. Autor names trades explicitly as resilient, “where there’s a lot of expertise involved” that AI can’t easily absorb. Education rounds out the list, propped up as much by demographics as by AI-resistance.

The Jobs Being Born

Server racks of increasing height with circuit-board sprouts growing from their tops while a person waters them
1.3 million AI jobs and 600,000 data-center jobs that didn’t exist five years ago — and the buildout is still being watered.

New professions are appearing at a pace that’s easy to miss if you’re only tracking what’s disappearing. According to LinkedIn data, the global economy added 1.3 million new AI-related jobs in the two years before January 2026 — roles like AI Engineers, Forward-Deployed Engineers, and Data Annotators that mostly didn’t exist five years ago — plus more than 600,000 new AI-enabled data-center jobs. AI Engineer (machine learning engineer) is LinkedIn’s #1 fastest-growing role for 2026, built on LangChain, retrieval-augmented generation, and PyTorch.

Some of the growth is physical, not digital: data-center technicians and commissioning managers are among 2026’s fastest-growing roles, concentrated in Washington D.C., Atlanta, Columbus, Houston, and Dallas — the AI buildout needs people who can wire and cool buildings, not just train models.

Inside existing companies, Microsoft’s 2025 Work Trend Index found 78% of business leaders planning to hire for new AI-specific roles (95% at “Frontier Firms”), with top emerging positions being AI trainers (32%), data specialists (32%), security specialists (31%), AI agent specialists (30%), and ROI analysts (29%). US roles requiring AI literacy rose 70% year over year, and “Head of AI” postings have surged across Australia, Canada, India, Germany, the UK, and the US.

There’s a pattern underneath it: new jobs cluster around building AI, deploying it, feeding it training data, supervising it, and monetizing it. Microsoft’s framing for where this is heading is blunt — every employee becomes an “agent boss,” managing AI agents the way a manager once managed people. Eighty-one percent of leaders expect agents moderately or extensively integrated within 12-18 months, and Microsoft’s telemetry shows active agents scaling 15x year over year (18x at large enterprises). If you’re early in your career, “supervising an agent” is a real skill line now, not a punchline.

Industries: What Stays, What Shrinks, What’s Born

Built to last. Healthcare and the broader care economy sit at the top of nearly every list here — Autor notes healthcare already accounts for roughly one in five US dollars and is the fastest-growing broad employment sector, with room for AI to upskill nurses, nurse practitioners, and technicians rather than replace them. Construction and the skilled trades hold up under every scenario the OECD modeled, including the forward-looking one. Education keeps growing on demographics as much as AI-resistance. And the AI buildout itself is now an industry: data centers and the physical trades that support them are hiring at pace with the software side.

Transforming or shrinking. Business process outsourcing and call centers, data-entry-heavy back offices, translation services, and print and postal work all sit inside the shrinking category the WEF, IMF, and Autor independently point to — cheap, codified, repeatable information work. Traditional back-office banking roles like tellers show the same pattern. Media and content work is transforming rather than vanishing outright, but Microsoft’s applicability research flags “providing information and assistance, writing, teaching, and advising” as the highest-overlap tasks of all.

Being born. AI infrastructure and data centers; AI security and compliance roles (31% of leaders’ hiring plans, per Microsoft); agent operations and AI agent specialists (30%); the training-data industry that annotators and content analysts now work in; and a fast-growing AI deployment-consulting layer — the forward-deployed engineers and AI consultants and strategists who help other companies actually put this technology to work.

Where the Experts Disagree

Strip away the headlines and the experts aren’t actually arguing about direction. They’re arguing about speed, and about how much pain the transition causes along the way.

Anthropic CEO Dario Amodei has made the most alarming public prediction in this space: AI could eliminate half of all entry-level white-collar jobs and push unemployment to 10-20% within one to five years, concentrated in technology, finance, law, and consulting. By May 2026 he’d added a nuance worth sitting with. Invoking the Jevons Paradox, he argued that automating 90% of a job doesn’t necessarily eliminate it — “the 10% kind of expands to be 100% of what people do” — because cheaper, faster output raises demand and workers shift to what’s left. But he qualified his own optimism: “AI is moving faster than all these previous technologies. And so when you strain a system more than it’s usually strained, it’s possible you get these weird behaviors and this big disruption.” His concern isn’t that the old logic is wrong — it’s that AI might be moving too fast for the labor market to rebalance the way it always has.

Nobel laureate Daron Acemoglu sits at the skeptical end on magnitude: he estimates AI will deliver only about 0.55% in total factor productivity gains over the next decade — roughly a 1-1.5% GDP bump — and that only about 5% of tasks are profitably automatable near-term, because current models “struggle with complex professional environments where it cannot ‘read a room’ or ‘connect non-obvious dots across domains.’” And yet Acemoglu still signed “We Must Act Now,” warning that if AI compresses the disruption manufacturing robots caused into a much shorter period, “that would be really disruptive, really costly for people’s livelihoods.” Low near-term automation and real long-term alarm aren’t a contradiction — they’re the same forecast at two different timescales.

David Autor reframes the whole question: the risk isn’t running out of jobs, it’s the devaluation of expertise. When AI automates something you spent years learning to do, its market value can evaporate even while the job title survives.

Goldman Sachs lands closest to reassurance, with real teeth in the caveats: it estimates AI exposes the equivalent of 300 million full-time jobs worldwide, and that roughly two-thirds of US occupations carry some automation exposure — but most are only partially exposed (25-50% of workload), which historically means complemented rather than replaced. Goldman’s historical anchor: 60% of today’s workers hold occupations that didn’t exist in 1940, and more than 85% of employment growth over the past 80 years came from technology creating new work, not just destroying old work.

The honest summary: nobody serious is arguing AI leaves the labor market unchanged. The argument is over whether the next five years look like a hard landing or a fast, survivable one — and how much of the cost falls on people who are 22 right now.

Your Playbook

A person climbing a staircase of glowing blocks while a small friendly robot places the next block ahead of them
Keep climbing, and let AI place blocks — don’t race it down the stairs.

The evidence above isn’t just diagnosis. It points fairly directly at what to do, whether you’re choosing a career, deep into one, or watching someone in your life navigate the transition.

Use AI Daily on Your Real Work

The data doesn’t reward AI-curiosity in the abstract — it rewards people already using it as part of how they work. Microsoft’s 2026 Work Trend Index identifies a “Frontier Professionals” segment (16% of AI users) whose habits are worth copying: 86% treat AI output as a starting point, not a final answer; 53% deliberately pause to decide what should be human versus AI; and 43% intentionally do some work without AI to keep their skills sharp. That last habit matters — Stanford’s data shows the employment penalty concentrates where AI automates a role outright, not where it augments the person doing it.

Build AI Literacy Formally

Microsoft’s 2025 Work Trend Index names AI literacy the single most in-demand skill of the year, and WEF ranks AI and big data the fastest-growing skill category overall, with a 17-percentage-point rise in employers naming it a core skill since 2023. This isn’t a suggestion to poke around with a chatbot — LinkedIn data shows US roles requiring AI literacy grew 70% year over year. Treat it like a credential you’re building, not a hobby you’re dabbling in.

Climb the Judgment Stack

As AI absorbs execution, the human value concentrates upstream, in judgment. WEF ranks analytical thinking the single most sought-after core skill, essential at seven in ten companies. Microsoft’s 2026 survey of 20,000 knowledge workers found quality control of AI output (50%) and critical thinking and objective analysis (46%) are the fastest-rising skills workers themselves report needing. The job increasingly isn’t doing the work — it’s deciding whether the work AI did is actually right.

Bank Tacit Knowledge Fast

Experience is the one asset AI can’t fast-forward through, and Stanford’s data proves it directly: workers with more tenure in exposed occupations keep growing employment while their less-experienced peers don’t. If you’re early-career, SignalFire’s advice is blunt and practical — bypass the traditional hiring gatekeepers and ship a verifiable portfolio of real work rather than waiting for a credential to open a door. New graduates are now twice as likely to be a “founder” than they were at the 2022 market peak; treat that as a legitimate path, not a consolation prize.

Anchor in Human-Facing, Physical, or High-Responsibility Work

If you’re choosing or switching fields, the OECD, IMF, and Microsoft research point at the same territory from three different angles: care work, skilled trades, protective services, education, and roles where responsibility and physical presence can’t be delegated to a model. None of this means “avoid technology” — it means anchoring your work in the parts of the job description AI still can’t do.

Make Reskilling a Habit, Not an Event

WEF projects 39% of workers’ existing skills will transform or become outdated by 2030, and estimates 59 out of every 100 workers will need training by then — with 11 of them unlikely to get it. Microsoft’s number is even starker: 70% of the skills used in most jobs today will change by 2030. The people this research treats kindly aren’t the ones who did one big retraining push. They’re the ones who never really stopped.

By Career Stage

Stage What the data says to do
Student / early career You’re standing on what LinkedIn’s Aneesh Raman calls “the bottom rungs of the career ladder” — the ones being broken first. Build a portfolio before you need one, use AI on everything you make, and weight resilient fields (care, trades, hands-on health) more heavily than you would have five years ago.
Mid career Your experience plus AI fluency is real leverage — this is the “agent boss” transition in practice. Microsoft found organizational factors account for 67% of whether AI actually benefits a worker, versus 32% for individual effort, so weigh whether your employer’s leadership is genuinely aligned on AI, not just talking about it.
Late career Your judgment and institutional memory are the scarcest resource in this whole picture. Lean into mentorship, supervision, and the calls only someone with your track record gets trusted to make.

How Seriously to Take All This

A note on humility, because this piece is built almost entirely on data that’s younger than it looks. Microsoft’s own researchers, whose applicability scores are cited throughout this piece, explicitly warn against equating high AI-activity overlap with actual job or wage loss — their data doesn’t capture the downstream business effects of the technology, which are “very hard to predict and often counterintuitive.” The Stanford Canaries research behind the entry-level findings hasn’t been peer-reviewed yet, however carefully it controls for confounders. And every forecast in this article — the WEF’s 2030 projections, Amodei’s five-year window, Acemoglu’s decade-long productivity estimate — is a bet on a trend line, made by people who have been wrong about technology before and will be again.

None of that is a reason to ignore the data. It’s a reason to hold it the way you’d hold any honest forecast: as the best current read, not a verdict.

Sources

Primary reports

Real-world data

Expert voices