What Does the Research Actually Say About AI and Jobs?
Goldman Sachs estimates generative AI exposes the equivalent of 300 million full-time jobs worldwide to some degree of automation. Exposure does not mean replacement. Anthropic's own usage data shows 57% of the work people bring to Claude is augmentation, with AI working alongside a person, versus 43% where AI completes a task alone.
The most widely quoted number comes from Goldman Sachs. In March 2023, its economists published "The Potentially Large Effects of Artificial Intelligence on Economic Growth," putting the global figure at 300 million full-time-equivalent jobs exposed to some degree of AI automation, alongside a projected 7% boost to global GDP if the technology delivers. The exposure was uneven. Office and administrative support topped the list at 46% of task-time potentially automatable, followed by legal at 44% and architecture and engineering at 37%, while construction sat at 6%, installation and repair at 4%, and building and grounds cleaning at 1%. The US average was 25%.
Which jobs have the most tasks AI could automate?
Source: Share of task-time exposed to generative-AI automation, by occupation category (US). Exposure is what AI could plausibly perform, not projected job losses. The US average is 25%. Source: Goldman Sachs, The Potentially Large Effects of AI on Economic Growth (Mar 2023), via CNBC.Exposure gets flattened into "AI will take these jobs" more often than the research supports. Anthropic's Economic Index, built on millions of real Claude conversations, found roughly 36% of occupations use AI in at least a quarter of their tasks, and about 4% use it across three-quarters. Anthropic also split usage into automation, where AI completes a task alone, and augmentation, where it collaborates with a person, and the split came out 57% augmentation to 43% automation. More often than not, then, AI is working alongside someone.
The OECD has been tracking a related question: when a job is highly exposed to AI, what happens to the skills it demands? Its working paper "Artificial Intelligence and the Changing Demand for Skills in the Labour Market" found AI-exposed occupations increasingly ask for management, project coordination, and communication skills alongside the technical work. Using panel data from real postings, the authors also found early evidence that this skills premium is starting to fade as adoption matures, a sign that employers are settling into what AI replaces inside a role while the role itself stays.
| Profession | Projected employment change, 2024-2034 | All-occupation average |
|---|---|---|
| Accountants and auditors | +5% | +3% |
| Lawyers | +4% | +3% |
| Physicians and surgeons | +3% | +3% |
| High school teachers | -2% | +3% |
Will AI Replace Lawyers?
No. The Bureau of Labor Statistics projects lawyer employment to grow 4% through 2034, in line with the rest of the economy. Legal work is heavily exposed to AI, 44% of task-time by Goldman's estimate, because so much of the job is document review and drafting, but the licensed, liability-bearing part of the profession is still adding people.
On paper, legal work looks like a textbook automation case, because contract review, discovery, research, and first-draft memos are exactly the structured, text-heavy tasks language models handle well. That is why legal ranked second highest in Goldman's exposure analysis at 44%. Yet BLS projects 31,500 lawyer job openings a year through 2034, with employment growing 4%, slightly ahead of the 3% average. Every licensed profession in this piece shows the same combination of high task exposure and positive employment. In law, AI absorbs the drafting and research, and the lawyer's day shifts toward counsel, negotiation, and the judgment calls a license exists to protect.
Will AI Replace Accountants?
No. BLS projects accountant and auditor employment to grow 5% through 2034, faster than the economy-wide average, with roughly 124,200 openings a year. Bookkeeping and reconciliation are among the most automatable tasks that exist, yet the number of people signing off on the results is still going up.
Business and financial operations sat at 35% task exposure in Goldman's breakdown, since categorizing transactions, reconciling accounts, and drafting routine filings are close to the ideal AI use case. We build these automations for clients today, matching invoices to purchase orders, flagging anomalies, and drafting first-pass summaries. What changes is how an accountant spends the day, with less manual entry and more judgment calls around tax strategy, audit risk, and advising a client on what the numbers mean. BLS's 5% growth projection, ahead of the 3% average, points the same way.
Will AI Replace Doctors?
No. BLS projects physician and surgeon employment to grow 3% through 2034, right at the all-occupation average, and the broader category of healthcare diagnosing and treating practitioners is projected to grow 8%. An aging population needing more care is adding medical jobs faster than AI is removing them.
Medicine has some of the same shape as law and accounting, since much of the day goes to documentation, note-taking, and pattern recognition, which are tasks where AI tools help. But "doctor" covers dozens of specialties bundled under one BLS category, and none of the growth projections point down. Physicians and surgeons overall are projected to grow 3%, matching the all-occupation average, while the wider bucket of healthcare diagnosing and treating practitioners, including nurse practitioners and physician assistants, is projected to grow 8%. Radiology, covered next, is the clearest case study for a specialty people assumed AI would gut first.
Will AI Replace Radiologists? What Did the 2016 Prediction Get Wrong?
No, and radiology is the cleanest natural experiment we have. In 2016 Geoffrey Hinton said radiologists would be obsolete within five to ten years. A decade later, the US radiologist workforce is up about 10%, case volume is up 25% since 2018, and average pay reached $571,000 in 2025, up 9% year over year.
In 2016, Geoffrey Hinton, one of deep learning's foundational researchers, told a Toronto machine learning conference that people should stop training new radiologists, calling it "completely obvious" that within five years deep learning would outperform them at reading scans. "If you work as a radiologist," he said, "you're like the coyote that's already over the edge of the cliff but hasn't yet looked down."
The 2016 prediction did not happen
Source: US active radiologists over the last 10 years and radiology case volume since 2018, a decade after Geoffrey Hinton called the job obsolete. Average radiologist salary reached $571,000 in 2025, up 9% year over year. Source: Fortune, May 2026, citing physician staffing and workforce data.That did not happen. Fortune's decade-later retrospective reports the US radiologist workforce grew roughly 10% over the last ten years, while radiology case volume climbed 25% between 2018 and early 2025, outpacing the supply of radiologists to read them. Average radiologist salary hit $571,000 in 2025, up 9% from the year before, and open radiologist job listings sat around 4,333 as of March, taking an average of 130 days to fill. A radiologist quoted in the piece put it plainly: "We actually have a huge shortage of radiologists."
Nvidia CEO Jensen Huang made the same point on the Dwarkesh Podcast, saying radiology doomers conflate one task, reading a scan, with the entire job. Netflix cofounder Reed Hastings said on the Possible podcast: "We're drawn to these scenarios of AI wiping out things, and again, it hasn't happened in radiology." There are structural reasons for this. Medicare and Medicaid only reimburse a study if a licensed physician performs the final read, and nobody has settled who is liable when AI misses a diagnosis. Reading images is also just one part of the day, since radiologists consult with physicians, monitor patients, and, in the interventional subspecialty, perform procedures. If the scan-reading is automated, their time shifts to everything else. Hinton has since said he was wrong about timing, though he still expects AI to make radiologists "a whole lot more efficient."
Will AI Replace Teachers?
Teacher employment is projected to decline, but the data ties that decline to enrollment and budgets rather than AI. BLS projects high school teacher jobs to fall 2% through 2034, driven by shrinking school-age enrollment and tight state and local budgets. In classrooms, AI shows up as a grading and lesson-planning tool.
Teaching is the one profession here where BLS projects a decline: high school teacher employment is projected to fall 2% through 2034, and all 66,200 annual openings come from replacing people who leave. It would be easy to read that as an AI story, but BLS attributes the decline to falling school-age enrollment and the risk that state and local budget shortfalls lead to layoffs, which makes it a matter of demographics and municipal finance rather than AI grading essays. Where AI does show up in the classroom, it is as an augmentation tool for differentiated worksheets, first-pass grading feedback, and lesson plan scaffolding, with the teacher working alongside it.
Are Artificial Intelligence Jobs in Demand?
Yes, and demand is rising fast. LinkedIn's 2026 Jobs on the Rise analysis ranked AI engineer as its top fastest-growing role, ahead of chief risk officer. While AI automates specific tasks inside existing professions, it is also creating new, fast-growing job titles that barely existed five years ago.
AI is creating jobs as well as automating tasks. LinkedIn's 2026 Jobs on the Rise report, built from millions of jobs started between 2023 and mid-2025, put AI engineer at the top of the list, ahead of chief risk officer and director of artificial intelligence. That fits with "artificial intelligence engineer" being one of the highest-volume career search terms right now, which reflects people watching a new job category open up and wanting in. This growth is concentrated in a narrow band of technical roles, though, and it does not offset a decline in paralegal or bookkeeping headcount one for one.
So, Honestly, Will Artificial Intelligence Take Away Your Job?
AI rarely takes your job outright. More often it replaces specific tasks, and the role changes around them. We automate tasks such as intake forms, scheduling, and first-draft replies inside small businesses every week, and the people who did that work mostly do more of what remains. The exposure is concentrated in entry-level, high-repetition work.
We are an AI research and implementation lab that builds AI systems for small and mid-size businesses, so we see this from the inside. The pattern has been consistent across every client: we automate a task rather than a role. A front-desk coordinator stops manually entering intake forms and starts handling the phone calls and exceptions that need a person. A marketing team stops writing first drafts from scratch and starts editing, positioning, and deciding what to publish. Headcount rarely drops on day one, but the job itself shifts toward less repetition and more judgment.
The risk is not evenly distributed, though. Anthropic's own June 2026 Economic Index report, drawn from interviews with 81,000 Claude users, found displacement worry "concentrated among early-career workers and occupations where we observe Claude doing the most work." That matches Pew's earlier finding that the most AI-exposed jobs skew toward data entry and routine analysis, and it matches what we see directly: the entry-level, high-repetition rung of almost every career ladder is shrinking, while the senior, judgment-heavy rungs above it are getting busier. If you are worried about AI and your job, the question to ask is not whether your profession survives but whether the specific tasks you do today will still be worth paying a person to do in three years. For most licensed, judgment-driven professions, the answer through the next decade is still yes.
