AI and Jobs: A Moving Target
For students choosing a path, parents trying to advise them, and the deans and provosts who make this happen for students.
The kid in my office last week
Last week a junior sat in my office and asked, almost apologetically, whether he should switch majors or add more courses to take. He is a strong CS student.
I get some version of this conversation almost every week now. Parents. High school juniors ask me on tour. The question is always close to the same.
What should my kid study to get a job? And keep it? And earn enough to live on?
Five years ago, I had a clean answer. Computer science. Maybe engineering. Pick up some math. Learn to write and how to effectively talk about your work. Done.
I do not give that answer anymore. Not because any field is dead. Computing and engineering is definitely not. The question itself has changed. We are really asking three things at once. Which jobs is AI absorbing? Which is it amplifying? Which will it leave alone long enough for a young person to build a career on top of? And tucked underneath: which will pay enough to make the path worth it?
That is harder to answer. But it is answerable. You have to be honest about three things. What kinds of work humans actually do. What kinds of work AI is good at. And what the past tells us about how new technologies reshape jobs.
So that is what this article does. With the technical details, because the answers hide there. In plain language, because you should not need a PhD to plan your own life.
One warning before we start. I am giving you scenarios, not predictions. Anyone offering a confident prediction about AI right now is probably guessing it. What I can offer is a way to think about your own bets. And a sense of what the people building this technology actually believe. They disagree openly and the disagreement matters.
The short version, by who you are
You probably do not have time for six thousand words. So here is the distilled version. The reasoning is underneath.
If you are a student
The labor market has started shifting under your feet. US programmer employment dropped about 27.5 percent between 2023 and 2025. Big tech hiring of fresh graduates is down about 50 percent over three years. Software developers ages 22 to 25 are down about 20 percent from their 2022 peak (but not software engineer or software architect). Customer support shows the same pattern. None of this is computing dying. It is the bottom rung of the ladder being sawn off. Senior engineers and computing experts are doing better than ever. The same compression is starting to show up in junior accounting, junior consulting, junior legal work, and commodity creative work.
Eight things to do.
Build a portfolio of real work. Not credentials. Not GPA. Three or four real things you made. Code on GitHub. A research write-up. A working prototype. A piece you published. An internship that produced something concrete. Show people what you can do.
Get fluent with AI tools, on your own terms. Use Claude, ChatGPT, Gemini, and Cursor every week. Build something with a model after you have a solid thoughtful work. Tune one. Then practice doing hard work without AI, so your judgment develops. Both halves matter.
Pair your major with depth in something else. CS plus biology. Finance plus quantitative skill. Liberal arts plus coding. Engineering plus AI fluency. Single-track resumes do less work than they did five years ago.
Look hard at the program before you choose. The gap between a strong program and a stale one is now bigger than the gap between majors. Curriculum quality. Faculty doing real work. Project-based capstones. Hands-on. Real industry partnerships. That is what determines outcomes.
Treat your first job as skill acquisition, not salary maximization. A 90 thousand dollar offer near a senior mentor doing real work compounds faster than a 150 thousand dollar offer where you do AI-replaceable tasks alone. Five years later, the first earner is often ahead.
Pick up one serious hands-on skill outside your major. Chip Design. Electrical work. EMT certification. Building robots. CNC machining. Sailing. It is a hedge against every scenario in this article.
Build relationships on purpose. Three professors who know your work. Five classmates you stay in touch with. One mentor a decade ahead of you. AI does not call your old boss for a job lead.
Take healthcare, the trades, and skilled technical paths like cybersecurity seriously. Not as a fallback. The earnings are real. For some students they are the strongest bet on the board.
If you are a parent
Your kid is entering a labor market that does not look like the one you entered. The doom headlines are overblown for some fields and understated for others. Your kid will hear all of them. They need help reading the noise.
Five things to do.
Do not push them into a major because it sounds safe. None of the safe choices from 2010 are clearly safe now. And do not push them away from a field because of headlines. The right answer is depth, engagement, and a serious program in something they love.
Let them try real work before they commit. A research lab with hands-on. A weekend with an electrician. A summer at a hospital. An internship at a startup. Career fit is mostly discovered, not chosen.
Talk about debt and time-to-income explicitly. Two hundred thousand dollars of debt for a 55 thousand dollar starting job is a different bet than zero debt and an apprenticeship at 80 thousand. Both can be right. Neither is automatically right.
Help them build a portfolio, not a resume. Ask what they have shipped, not what classes they have taken. If the answer is nothing yet, that is the work.
Hold two questions in your head, separately. Will my kid get a job in five years? Will my kid be okay in twenty? The first is about the labor market right now. The second is about whether they build deep skill, durable judgment, and human relationships across decades.
If you teach, advise, or run a program
The students you graduate in 2030 will work alongside AI systems dramatically more capable than today’s. Some of the curriculum you teach has not been seriously updated since 2015. The gap between strong programs and stale ones has widened to the point where it matters more than the choice of major. Your graduates’ outcomes will reflect the choices you make in the next two years.
Six things to do.
Rebuild the curriculum and understand the AI abundance vs AI replacement. What your graduates can do something employers will pay for in 2030.
Protect the foundational years from AI shortcuts. First-year writing, math, and core technical courses must build cognition before AI is allowed in. Use a tiered model. Restrictive in foundations. Progressive as competence is demonstrated.
Build a real AI literacy curriculum across every major. Not a workshop. A coherent sequence. Tier 1 foundations for everyone. Tier 2 applied work by discipline. Tier 3 specialized for graduate and advanced learners. Tied to demonstrated competence, not seat time.
Update curricula in every reshaped discipline, not just computing. Engineering. Business. Journalism. Design. The sciences. Project-based learning. Real systems work. Capstones tied to current problems. Shipped work as a graduation requirement.
Build serious applied pathways alongside the four-year degree. Bridges to technical work, trades, and healthcare. Stackable credentials. Apprenticeship-style internships. Regional public institutions especially have an obligation here.
Invest in faculty AI fluency. Faculty who are not themselves AI-fluent cannot teach the literacy students need. That is a faculty development priority, not an IT issue.
That is the action list. The rest is the reasoning behind it.
A useful map of jobs
Most “future of work” essays split jobs into high-skill and low-skill, or creative and routine. Those splits are useless here. AI does not respect them. It is happily eating routine creative work like graphic design templates and stock copy. And it is sweating over routine physical work, like folding a fitted sheet or troubleshooting a 30-year-old furnace in someone’s attic.
A better map has four buckets.
Hands work. The constraint is what your body can do in the physical world. Four kinds. Fine motor: surgery, dentistry, microelectronics rework, lab bench work, fiber optic splicing. Gross motor: framing, roofing, warehousing, mechanics, line cooks. Unstructured environments: plumbing in a hundred-year-old building, EMT calls, oilfield service, wildland firefighting. Structured environments: assembly lines, packaging, warehouse picking, fast food. Unstructured is the hardest category for AI and robots to touch. Structured is where humanoid robots are arriving first.
Minds work. The constraint is what your brain can produce in symbols. Four kinds again. Structured generation: boilerplate code, standard reports, template contracts, junior accounting, marketing copy. The bucket LLMs are eating fastest. Open-ended reasoning: research, novel design, real legal strategy, complex system architecture. AI assists. AI does not replace. Retrieval and synthesis: literature reviews, claims processing, document review. AI is reshaping this work significantly. The work itself does not vanish, but the time it takes a human has dropped five to ten times. Judgment under uncertainty: clinical decisions, executive calls, courtroom strategy, investment-committee decisions. AI is an input. Accountability stays human.
Hybrid work. The messy middle. You need both hands and minds. A surgeon reasoning about anatomy and sewing a vessel by feel. A field engineer diagnosing a failing turbine and climbing up to fix it. A teacher planning a lesson and reading a room of fourteen-year-olds. A controls engineer at a chemical plant. An aircraft mechanic with avionics training. The most AI-resistant of all. Both loops have to close inside the same person.
People work. Management, mentorship, sales, persuasion, care, trust-building. A nurse manager. A construction foreman. A founder. An engineering manager. A college dean. A pastor. People work has a hands component (presence, eye contact) and a minds component (planning, reading social situations). Its irreducible core is another human accepting you as someone worth listening to. Every senior role in every field is, eventually, mostly people work.
Hold this map in your head. Now we ask which AI capabilities threaten which buckets.
What AI is actually good at
The biggest mistake people make is treating AI as one thing. It is not. It is a family of different systems, each chewing through a different part of the job map.
Large language models. ChatGPT, Claude, Gemini. Trained to predict the next word on roughly the readable internet, then tuned to follow instructions. Extraordinarily good at structured generation in symbol space. Writing a Python function. A contract clause. A literature review. A SQL query. They improve fastest on tasks the model can check itself against, like code and math. They are weak at long-horizon planning. Weak at acting in the physical world. Prone to making things up in domains they were not trained on. They threaten structured minds work.
Reinforcement learning agents. The family that gave us AlphaGo, AlphaZero, AlphaFold, and the systems that beat the world at chess, poker, and StarCraft. RL is what you reach for when there is a clear goal, a fast simulator, and a closed loop between action and reward. It is also the engine inside the new “reasoning” models that take a hard problem and chew on it before answering. And it is what trains humanoid robot policies, by simulating billions of stumbles before the robot ever touches the floor.
Vision and multimodal models. Models that see, read, and listen, often inside one architecture. They can read a chest X-ray, transcribe a meeting, navigate a website, look at a factory line and identify a bad weld. They threaten retrieval and synthesis in domains we used to assume needed eyes. Radiology. Pathology. Inspection. Claims processing.
Generative media models. Image, video, music, voice. They threaten the structured-creative end of minds work. Stock photography. Illustration. Marketing video. Voiceover. They are not yet replacing the top of any creative field. But they have devastated the bottom and middle.
Embodied AI and humanoid robots. Tesla Optimus. Agility Robotics’ Digit. 1X NEO. Apptronik Apollo. Boston Dynamics Atlas. As of early 2026 the picture is real but narrow. Tesla has reportedly deployed over a thousand Optimus units inside its own factories. Figure has units at BMW. Digit moves totes at Amazon. Pricing has fallen from a 2023 norm of 50 to 250 thousand dollars per unit toward Tesla’s stated 20 thousand dollars target at scale. They handle structured tasks. They cannot yet match a human at a varied factory line, let alone an unstructured plumbing call. The cost curve is bending fast.
Map all of this back. Structured minds work is being compressed now. Open-ended minds work and judgment work get a strong assistant but no replacement. Structured hands work is approaching pressure at speed. Unstructured hands work is years away. People work is the most resilient of all.
What the labor market is actually doing
Numbers anchor the abstractions. Here are the ones that matter.
The Stanford Digital Economy Lab paper “Canaries in the Coal Mine?” used anonymized payroll data through September 2025. It found that employment for software developers ages 22 to 25 dropped nearly 20 percent from its 2022 peak. The same data shows that employment for older software developers in the same fields kept rising. The pattern held across the broader category of computer occupations and across customer support. Stock clerks and home health aides, rated as much less AI-exposed, showed nothing of the kind.
Read that asymmetry carefully. The 22-to-25 cohort doing entry-level structured work is being compressed. The senior cohort doing higher-judgment work is growing. This is what we would expect from how LLMs behave. It is not a story about computing collapsing. It is a story about the specific tasks that used to occupy the bottom rung of the career ladder.
Other data points line up the same way. IEEE Spectrum, citing US BLS data, reported overall programmer employment fell 27.5 percent between 2023 and 2025. The broader category of “computer occupations,” which includes hardware, ML, security, and systems engineers, kept growing. SignalFire found fresh-graduate hires at major tech companies globally are down more than 50 percent over three years. Indian IT services have cut entry-level roles 20 to 25 percent. EU job platforms saw a 35 percent decline in junior tech postings in 2024. Wall Street banks are cutting roughly 200,000 roles concentrated in entry-level analyst positions. Challenger, Gray and Christmas attributed nearly 55,000 US layoffs in 2025 directly to AI.
So what is this, despite the headlines? Not a collapse of any field. The same period has seen senior engineering wages reach all-time highs. Demand for ML, AI, hardware, and security engineers far exceeds supply. The career ladder is not broken. The entry rung is significantly higher than it used to be.
Outside of knowledge work the picture is different. Programmer employment is down 27.5 percent. Nursing-aide employment is up. Skilled-trades demand is rising, partly because AI itself needs trades to exist. Every gigawatt of AI compute requires electricians, HVAC technicians, controls technicians, plumbers, and the data-center construction crews behind them. The International Energy Agency projects global data-center electricity use more than doubling to about 945 terawatt-hours by 2030. That is a tailwind for the people who wire, cool, and maintain those buildings. Nurse practitioner employment is projected to grow about 40 percent by the early 2030s.
What the people building AI actually believe
The people closest to AI are openly disagreeing about what happens next. The disagreement is the news. Anyone offering your kid a single confident answer is hiding it.
Dario Amodei, CEO of Anthropic. In May 2025 he warned Axios that AI could eliminate up to half of entry-level white-collar jobs and push US unemployment to 10 to 20 percent within one to five years. Concentrated in finance, consulting, law, and tech. He repeated the warning in a January 2026 essay. His core claim: AI’s labor-market shock will be larger and faster than past technology shocks, because earlier waves automated narrow capabilities while AI threatens broad ones. He has called for a “token tax” on AI firms to fund redistribution.
Mustafa Suleyman, CEO of Microsoft AI In early 2026 he argued publicly that AI is on a path to automate most or all white-collar tasks within 12 to 18 months, naming law, accounting, project management, and marketing. His tone is calmer than Amodei’s. His timeline is shorter.
Demis Hassabis, CEO of Google DeepMind. He places true human-level AGI 5 to 10 years out. He has said the prospect of systems smarter than humans keeps him up at night, and that society, including most economists, is not ready. He paints a “radical abundance” best case alongside a darker downside.
Sam Altman of OpenAI. The cost of intelligence is dropping by more than 10 times per year. Intelligence will eventually be “too cheap to meter.” Increased productivity historically creates more demand, not less. He has floated “universal basic wealth” as a policy response. At the BlackRock summit in March 2026 he warned that capitalism has historically depended on a labor-capital balance that breaks if it is hard to outwork a GPU.
Elon Musk takes the most absolute abundance position. Within 10 to 20 years, AI plus robotics will produce so much that work becomes optional. The right response is “universal high income,” not basic, paid for by the productivity surplus. Musk says saving money will eventually be unnecessary. The deepest open question, in his framing, is meaning, not economics. Whether you believe him or not, he is acting on it. Optimus is being deployed inside Tesla factories. He just announced the largest single semiconductor commitment in years.
Now the other side of the spectrum. Geoffrey Hinton, who helped train modern neural networks, warns that AI-driven disruption will concentrate wealth and is calling for redistribution. Yann LeCun, Meta’s chief AI scientist, thinks current LLM-based approaches have fundamental architectural limits, and that real autonomous intelligence is further out. Gary Marcus argues the field consistently over-promises. Yale’s Budget Lab, looking at 2022 to 2025 data, found AI had not yet caused widespread job losses across the broad US labor market.
The honest summary. The people building this technology think the labor-market shock is real and large. They disagree about timing. They disagree about magnitude. They disagree about whether AGI arrives in 2026 or 2040. They disagree about whether the eventual settling-out is utopia, dystopia, or another awkward middle. That uncertainty is the actual situation.
The buildout is enormous, and it needs people
A minute on the physical and capital buildout. It is the single biggest determinant of how capable AI gets between now and 2030. And it is creating massive demand for engineers and tradespeople of every kind.
In March 2026, Elon Musk announced “Terafab,” a 20 to 25 billion dollar semiconductor compound on the north campus of Giga Texas. A joint venture among Tesla, SpaceX, and xAI. The stated target is one terawatt of AI compute capacity per year. In April 2026, Intel publicly joined as the foundry partner. Tesla’s own filings describe a parallel buildout. Cortex 1, with over 100,000 H100-equivalent GPUs, is in production. Cortex 2 is in early ramp. Custom AI5 inference chips are taping out for 2027 production.
OpenAI, Oracle, SoftBank, and partners under the “Stargate” banner have committed roughly 400 billion dollars of AI infrastructure investment. Sam Altman publicly stated an aspiration of building roughly one gigawatt of new AI capacity per week. Microsoft, Google, Meta, and Amazon are running parallel buildouts at similar scale. China is building furiously. The Gulf is building nation-scale AI infrastructure.
Now think about what that buildout actually requires. Chip designers. Distributed systems engineers. Networking engineers. Power systems engineers. Mechanical and HVAC engineers for cooling. Civil and structural engineers for the buildings. Environmental engineers for permitting. Construction trades by the tens of thousands. Embedded engineers for the robots that go inside. Security engineers. Compiler engineers. Materials scientists for the new chips. Chemical engineers for the new battery chemistries. Nuclear engineers for the small modular reactors being announced to power it all.
The dominant story in tech-industry headlines is that AI is replacing programmers. The dominant story in actual industry capital allocation is that AI is creating the largest hiring wave for engineers and tradespeople in a generation. Both stories are true. They are about different parts of the work. And the students who recognize this distinction, and orient toward the second story, are walking into one of the strongest job markets that any group of technical fields has ever seen.
Three scenarios for the next five years
The next five years are mostly determined by AI capabilities that already exist and are scaling, plus the labor-market response.
Scenario 1. Slow but steady. AI capabilities improve roughly on the current curve. Reasoning models get better year over year. Agents get reliable enough for narrow business workflows. Humanoid robots stay confined to a handful of structured industrial deployments. Junior structured knowledge work compresses sharply. Mid-career knowledge workers become two to four times more productive. Trades and the care economy quietly absorb a lot of displaced labor. Wages bifurcate. People who use AI well get a real premium. People who do not drift toward physical and care work. My estimate: roughly 50 percent probability.
Scenario 2. Agent breakthrough. Autonomous agents make a real generational leap, the way image generation did between 2022 and 2024. The story we are already living in entry-level CS spreads everywhere. Junior analysts in finance. Junior associates in law. Junior consultants. Junior auditors. Whole departments shrink not through layoffs but because nobody backfills. This is the world Amodei and Suleyman warn about. Probability: maybe 30 percent. Painful for new graduates outside specialized technical fields. Better for graduates who can build, deploy, or audit those agents.
Scenario 3. Embodied AI inflects. Cheap humanoid platforms reach product-market fit in structured industrial environments. Auto plants, fulfillment, light assembly, food processing. Humanoid manufacturing costs reportedly dropped 40 percent from 2023 to 2024 per Goldman Sachs. If this hits inside five years, you would see millions of structured-environment hands jobs come under pressure for the first time. Probability: maybe 20 percent, weighted toward later in the window.
These scenarios are not mutually exclusive. The world, I expect is mostly Scenario 1, with patches of Scenario 2 in legal, healthcare admin, software, design, and marketing, and early signs of Scenario 3 in some factories.
Three scenarios for the next ten years, and the AGI question
The five-to-ten-year window depends on things that are genuinely uncertain. Whether scaling continues to deliver gains. Whether humanoid robotics inflects. Whether the data and energy buildout actually gets built. Whether the political response is fast or slow. And whether AGI, defined as systems that can do essentially any cognitive task a human can do, actually arrives.
That last one matters. Amodei: 2026 to 2027 for systems “broadly better than all humans at almost all things.” Hassabis: 5 to 10 years for true human-level AGI. Suleyman: most office work automatable in 12 to 18 months. Altman: deliberately fuzzy. Musk: keeps sliding the date. Independent researchers are more conservative. The 2025 prediction-tracking community puts a roughly 50 percent chance of “minimal AGI” around 2030, with researcher surveys still clustering around 2047. Reasonable people place this anywhere from 2027 to 2047.
The takeaway for students. The 10-to-15-year window is exactly when most credible forecasts say this question gets answered. Whatever you study, you will spend at least the second half of your career working alongside systems dramatically more capable than today’s. Plan for that.
Three scenarios for 2031 to 2036.
Scenario A. A new equilibrium that looks like the IT revolution did. Past transitions led to robust growth in employment and real wages after a painful adjustment. Think about what the 1990s software boom did to clerical work, secretarial work, middle management, and physical record-keeping. It demolished entire categories. Then it created, over a decade and change, web designer, data scientist, product manager, DevOps engineer, cloud architect, growth marketer, and dozens more. In Scenario A, AI does the same. New categories emerge. AI-systems orchestrators. Agentic-workflow designers. Model evaluators. AI auditors. Robot-fleet operators. AI-and-domain hybrids in every field. Painful in the middle. Fine at the end.
Scenario B. Abundance, with a distribution problem. AI plus robotics drives the marginal cost of intelligence and a lot of physical production toward zero. Goods get cheaper. Healthcare gets cheaper. Education gets nearly free. The optimistic version is more jobs, not fewer, because people seem to want unlimited stuff and have a deep drive to express creativity and be useful. Demand stretches to fill supply. Personal trainers. Micro-restaurants. Indie game studios. Custom education. Musk’s version is more radical. Most work becomes optional. Universal high income replaces wages. Hassabis describes it as radical abundance and asks the harder question: what happens to meaning when most of us are not economically necessary? The pessimistic version of B is that abundance happens but most of the surplus accrues to a small number of firms that own the model weights and data centers, and the political fight of the 2030s is over distribution.
Scenario C. A two-tier economy. Some humanoid robotics. Some agent breakthroughs. Lots of disruption. A sluggish institutional response. Top tier: people who own AI, deploy it, build with it, train it, govern it, plus professionals whose human accountability is irreducible. Middle tier: the white-collar middle gets thinner. Bottom tier: trades and the care economy hold up well in volume but do not pay enough to replace what was lost in the middle.
Numbers, with wide error bars. A is about 40 percent. B is about 25 percent and rises later in the window. C is about 35 percent.
The historical argument
Pause for a second on the abundance argument and the case that this time is different. Students and parents hear both. They do not know what to do with them.
The historical case, in its strongest form. Every previous wave of automation destroyed entire categories of work and was, in its moment, terrifying. Each ended up creating more work than it destroyed, because human wants are unbounded.
Five examples.
Agriculture. In 1790, about 90 percent of the US workforce worked on farms. By 1900, 41 percent. By 1930, 21.5 percent. Today, well under 5 percent. The total US population multiplied. We did not run out of demand for food. Farm labor was liberated into other work, much of it better paid.
The Industrial Revolution. Manufacturing as a share of US workers rose from about 15 percent in 1880 to peak at about 38 percent in 1944. Then automation, robotics, and globalization started the same cycle on the manufacturing workforce. By 2019, manufacturing had fallen to about 8.5 percent of US employment, even as manufacturing output kept growing. Service-sector employment rose from about 31 percent of the US workforce in 1900 to 78 percent by 1999.
Electrification. Between 1900 and 1940, the grid was built out. Lamplighters vanished. Ice harvesters and ice deliverers. Switchboard operators. Stenographers in their original form. And electrified factories created whole new categories. Electrical engineer. Electrician. Line worker. Power plant operator. Appliance repair. The entire consumer electronics industry. Electrical engineering as a field was not threatened by electrification. It was born of it.
The IT and internet wave. Between 1980 and 2020, hundreds of clerical job categories declined or vanished. Typing pools. Travel agents. Paper records clerks. Bank tellers. Video rental clerks. In their place: software engineer, data analyst, web designer, growth marketer, cloud engineer, DevOps. Net jobs grew. Computer science went from a small academic specialty to one of the largest majors in the country.
The pattern. Displaced workers and their children did not, in aggregate, end up unemployed. They ended up doing different work, in jobs that did not exist when the transition started. People bought iPhones they did not strictly need. They bought a second one for their kid. They bought yoga classes and craft beer and Peloton and TikTok subscriptions. Demand stretched.
Now the pushback, in its strongest form. This time may actually be different. Earlier waves automated muscle and left the brain alone. Earlier waves automated narrow skills and left flexible reasoning alone. AI threatens both. Amodei has written exactly this argument. Earlier shocks affected only a small part of the range of human ability. AI’s effects are broader and faster. There may be much less room left to expand into.
There is also a distribution concern that history does not settle. Real wages for the median American worker have grown much more slowly than productivity since the late 1970s. If AI’s surplus accrues mostly to firms that own the model weights and data centers, the historical “more jobs eventually” outcome could be true and the median worker could still be worse off.
Both arguments are right about something. Abundance is plausible. The transition will be brutal. The distribution of the abundance is the real political fight. The right response, at the level of an individual choosing what to study, is to plan for the brutal transition while staying open to the abundance. Build skills that compound regardless of which version arrives.
One observation worth keeping. In every previous transition, the people who designed the new technology came out best. Electrical engineers in the early 20th century. Computer scientists in the late 20th. There is no reason to think the AI era will be different. And the same observation applies more broadly. The people who deeply understood whatever the new economy needed (electricity, software, biology, finance, design) did well. The same will be true here.
A 10-year job map: replaced, assisted, untouched
Three lists. Each starts with the rule for what puts a job there.
Most likely to be largely replaced or compressed by 70 percent or more
The work is overwhelmingly structured generation or structured retrieval in symbol space. There is a clear evaluation signal. Legal accountability is modest. Examples: junior software engineers writing standard CRUD code, glue code, and routine bug fixes. Tier-1 SOC analysts and helpdesk first-line support. Junior paralegals doing document review and contract templating. Junior accountants and auditors doing transaction-level work, basic tax return prep, and routine reconciliations. Insurance claims processors and underwriters at the routine tier. Mid-tier financial analysts doing standard models and pitch decks. Customer support agents handling structured queries. Translators and transcriptionists for non-specialist content. Copywriters and SEO writers for commodity content. Stock photographers, stock illustrators, and template-driven graphic designers. Voiceover artists for commercial work. Data-entry clerks and routine billing specialists. Telemarketers. Executive assistants whose work is mostly scheduling and travel booking. Standard CAD drafters and routine BIM modelers. Inspection-line factory workers in structured environments. Warehouse pickers in structured fulfillment. Long-haul truckers on highway segments (partial, not full).
The pattern. Eighty percent of the day is structured, repeatable, and evaluable.
Significantly assisted but still anchored to humans
The largest category. The quality of your AI fluency determines your career outcome. AI changes how you work but does not remove the job. Examples: senior software engineers, architects, and principal engineers. AI engineers, ML engineers, MLOps engineers, AI product managers. Hardware engineers, chip designers, semiconductor process engineers. Cybersecurity engineers, threat hunters, incident responders, CISOs. All licensed engineering disciplines (electrical, mechanical, civil, chemical, materials, biomedical, industrial, environmental, aerospace, petroleum, nuclear). Power systems and controls engineers. Doctors of all specialties. Lawyers doing strategy, litigation, deal negotiation, and trial work. CPAs and senior auditors doing complex tax and M&A work. Financial advisors and wealth managers. Senior consultants and partners. Marketing strategists, brand managers, creative directors. Investigative journalists and opinion writers. Senior editors and acquisitions editors. Architects (the building kind) and senior interior designers. Pharmacists. Veterinarians, dentists, surgeons. Researchers across academia and industry. Teachers, professors, and instructional designers. HR business partners and talent leaders. Project, program, and engineering managers. Salespeople in consultative or relationship-driven fields. Air-traffic controllers, pilots, ship captains. Police officers, social workers, public defenders.
Least impacted
Three properties, any of which earns a spot. Unstructured physical environments. Fundamentally interpersonal work. Physical care of vulnerable bodies. Examples: skilled trades in unstructured environments (electricians, plumbers, HVAC, welders, carpenters, masons, roofers, locksmiths, automotive mechanics, heavy-equipment operators). Industrial and field-service technicians (controls techs, instrumentation techs, fiber-optic splicers, oilfield service, wind-turbine techs, solar installers, marine engineers, building-automation field engineers). Bedside and critical-care nursing. Physical and occupational therapy, speech-language pathology. Mental-health counselors, therapists, psychologists, psychiatrists. Home health aides, CNAs, eldercare workers. Veterinary technicians. Dental hygienists. Early-childhood and K-12 teachers, especially elementary and special education. Coaches, personal trainers, fitness instructors. Hairstylists, barbers, estheticians. Chefs and line cooks in unstructured kitchens. Emergency responders (firefighters, paramedics, EMTs, search and rescue). Clergy, chaplains, hospice workers. Diplomats, negotiators, mediators. Judges, trial attorneys, lawmakers. Senior leadership across every industry. Specialized artisans and high-craft creative work, including top-tier writers, directors, musicians, master sommeliers, and chefs at the highest level. Researchers in physical sciences who do bench or field work.
Two patterns to notice. First, the jobs most likely to be replaced cluster in the 22-to-30-year-old demographic. They are the rungs of the ladder that recent graduates have used since the 1980s. Second, these lists are not a hierarchy of value or pay. Career resilience and earning power are different axes.
What to study, by major area
A walk through the major fields the way I think about them. What is happening to each. What to do about it. Same logic across fields. No favorites.
Engineering and computing
A broad family. The reshaping inside it is uneven.
Software, AI, ML, and cybersecurity at the senior level are doing extremely well. AI engineer median total compensation in 2026 is around 185 thousand dollars. ML engineer around 165. Senior bands run 200 to 260. Frontier labs and FAANG go well past 400. Demand for strong senior engineers far exceeds supply. The bar for entry has risen. The path of “graduate, get a junior job, learn on the job” is mostly closed at large companies. Students who want this work need to arrive with shipped projects, real systems experience, and demonstrable AI fluency.
Hardware and chip design is in scarce supply and well paid. The Terafab project, the Stargate buildout, and every AI accelerator company need electrical and computer engineers who can do real silicon work. Power systems engineers. Controls engineers. Embedded engineers. RF engineers. Some of the best-paid and most stable specialties on the board. Few new graduates pick them.
Mechanical engineering has tailwinds from the AI infrastructure buildout. Data center cooling. Robotics. Reshoring of manufacturing. Aerospace and defense. Electric vehicles. The structured CAD work is being augmented by AI. The judgment-and-physics work is not.
Civil and structural engineering has the largest demand surge in decades. Every gigawatt of AI compute needs substations, transmission, water systems, site work. Aging US infrastructure needs rebuilding. Bridges. Roads. Airports. Water systems. The PE stamp matters. AI cannot legally sign drawings. A human engineer must.
Chemical, materials, and biomedical engineering all sit in a strong position. Battery chemistry. Semiconductor process. Drug discovery. Biomaterials. AI helps with simulation. Lab work and process expertise are not replaceable. Pharmaceutical and biotech demand is growing.
Industrial, environmental, nuclear, and aerospace engineering each have specific tailwinds. Manufacturing reshoring. Grid modernization. Permitting demand. Defense. Small modular reactors.
The pattern across all of engineering. AI is making structured generation work cheaper. AI is making the judgment work more valuable. And AI is creating massive demand for engineers because building AI itself is the largest engineering project of the decade. Strong engineers from any discipline will do well. Stale programs in any of these disciplines will produce graduates who struggle. Quality of program matters more than choice of major.
Sciences
Biology, chemistry, physics, math, statistics, geosciences, environmental science. AI is reshaping these fields without replacing the people in them.
Hypothesis generation stays human. Experimental design stays human. Bench work stays human. Field work stays human. Reading the literature, choosing what to study next, knowing what is real versus what is artifact. All human. AI accelerates the cycles. A modern biology lab can generate and test 10 times the hypotheses it could a decade ago. The scientists are not interchangeable.
Domain scientists with serious computational skill and AI fluency are getting offers from industry at levels that would have been unthinkable a decade ago. Biotech pays. Materials companies pay. Quantitative finance pays.
For an undergraduate, the question is what comes next. Med school. Grad school. Industry. Each has its own logic. The portfolio principle still applies. Research experience. Published or near-published work. Real lab work. Show the skill. Do not just claim it.
Health and care professions
The strongest tailwinds of any field for the next decade. The driver is demographics, not AI. The aging US population needs more care than the workforce can supply. Nurse practitioner employment is projected to grow about 40 percent by the early 2030s.
Medicine, nursing, physician assistant, occupational therapy, physical therapy, mental health counseling, social work, dentistry, veterinary medicine, pharmacy. AI helps with diagnosis, charting, billing, and routine documentation. Humans hold the patient relationship, the physical care, and the legal accountability.
Pay varies. Doctors and dentists earn at the top. Nurse practitioners and physician assistants typically earn 120 to 140 thousand. Nurses earn 80 to 100, with strong management ladders. Counseling and social work pay less but the demand is large and growing. For students who want stable demand, hands-on work, and meaning, this family of fields is hard to beat.
Trades and skilled technical work
The most underrated path on the board for the right student. Electrician. Plumber. HVAC technician. Welder. Controls technician. Lineman. Building automation. Solar and wind installer. Industrial maintenance. Fiber optic technician. Telecom field tech.
Demand is strong and growing. The AI buildout is one of the largest construction booms in modern US history. Every gigawatt of new compute needs trades on site for years. Pay is real. Experienced trades workers in many regions earn 80 to 120 thousand. Master-level and self-employed regularly clear 150. Debt is low. Time to income is fast.
This is not a fallback path. For some students it is the strongest bet they could make.
Business, finance, and organizations
The most internally divided category.
Junior analyst work in finance, junior consulting, junior accounting, junior auditing, basic corporate development. All being compressed. The deck-building, modeling, and comp-table work that used to occupy first-year analysts is exactly what AI does well.
Senior advisory work, deal-making, complex tax strategy, M&A diligence at the partner level, wealth management built on long-term client trust, enterprise sales, executive leadership. All hold up well. Some get more valuable as AI compresses the work below them.
For students drawn to business, the playbook is to build technical depth alongside the business education. A finance degree paired with serious quantitative skill. Accounting paired with data and AI fluency. Marketing paired with creative direction and analytics. Consulting paired with a real domain like healthcare, energy, or technology.
Entrepreneurship is more open than it was five years ago. AI lowers the cost of building things. A two-person company can now do what 20 used to.
Arts, communication, and creative work
The field most cleanly split into two markets. The top is doing fine. The middle and bottom are taking heavy damage.
The compressed end. Stock photography. Stock illustration. Template-based graphic design. Commodity copywriting and SEO writing. Voice work for commercial purposes. Most translation. Routine video editing.
The intact end. Top-tier writing, both fiction and journalism with original reporting. Top-tier visual art with a real point of view. Senior creative directors. Senior editors. Live performance. Theater. Documentary work. Creative work whose value depends on a specific human point of view.
For students drawn to creative work, the realistic frame is this. Paths to entry are harder than they were. Compensation at the top of the market may go up, because the bottom is getting cheaper and the audience for distinctive work grows. Build a real point of view. Ship things. Develop a voice. Treat AI as a tool that handles the parts you would have hated anyway.
Performance arts (music, dance, theater, live entertainment) hold up well. The product is presence. AI cannot supply that.
Humanities and social sciences
History, philosophy, English, languages, religious studies, classics, anthropology, sociology, psychology, political science, economics, geography.
The honest picture. Direct labor markets for most humanities and social science PhDs were soft before AI. AI does not help. The reshaping at the undergraduate level is more interesting. Humanities skills (writing, reasoning, reading hard things carefully, building arguments, understanding people) are exactly the skills employers say they need from AI-era workers. The challenge is that students in those majors often graduate without specific applied skills employers can recognize.
The playbook for a humanities or social science undergraduate. Pair the major with one applied skill. Coding. Statistics. Design. Business. Languages. Build a portfolio of writing, research, or applied work that an employer can read in five minutes. Take seriously the option of professional school after the degree. Law. Medicine. Education. Public policy. Do not assume the degree alone will be read as a credential the way it was in 1990.
Psychology and economics specifically have stronger direct paths than the rest of this category, especially with quantitative depth.
Education
K-12 teaching, especially in math, science, and special education. Higher education. School counseling. Educational technology. Curriculum design.
Strong and stable demand at the K-12 level. The driver is demographics and irreducible human work. AI is going to change how teaching is done. AI tutors are real and useful. Classroom presence and human mentorship are not going away.
Pay is modest in absolute terms. Stability is high. For students who genuinely love teaching, the path is durable.
Public service and law
Government service, policy work, public defense, prosecution, judicial work, military service, law enforcement at the supervisory level, social work, public health.
Mostly durable. The reshaping happens inside. Junior paralegal work compresses. Junior policy analyst work compresses. Senior strategy and judgment work holds up. Trial work, courtroom argument, judging, legislating. None of those are about to be done by AI.
Law specifically has the same internal split as business. The bottom of the pyramid is being compressed. The senior partner work is not. Students considering law should look hard at which kind of legal work draws them, and at the program quality and placement record of the schools they consider.
To grow over the next 15 years
The 15-year horizon is leadership. People work, whatever your technical background. Running a team. Running a department. Running a company. Running a hospital wing. Running a school. Running a portfolio. The senior-most jobs in every industry are about judgment, accountability, persuasion, and the trust to hold scarce resources for a group. None of those is anywhere close to being automated. The people who grow into them in the late 2030s are the ones who, in 2026, started building three things. Verification skills (the ability to catch AI when it is wrong, in a domain where you have real expertise). Systems thinking (the ability to design how tasks compose into a working organization). And relationships (the people who know your work, trust you, and call you when something hard comes up).
Earnings at 15 years. The spread between paths is widest here. People who reach senior leadership levels (VP or director in tech, partner in professional services, attending physician in medicine, principal engineer in industry, master in trades, founder of a successful firm) earn many multiples of their starting salaries. The peak is highest in tech, finance, and entrepreneurship. The median in nursing administration, engineering management, school principalship, or running a successful trades business is also genuinely strong. And the path is more predictable.
The deepest answer about long-term success. The skill that compounds most over a career in a high-disruption era is the ability to keep learning under uncertainty. Build that skill in college. Struggle honestly with hard material. Do not let AI do your thinking for you when you are supposed to be developing it. Read widely. Write in your own voice. You will adapt to whatever AI looks like in 2041. If you do not build it, no specific major will save you.
Closing
Back to the question I get very often. The most honest answer I have:
Study something hard enough that becoming actually good at it changes who you are. Make sure that thing has at least one of three properties. It requires deep human judgment under accountability. It requires presence in the physical world. Or it builds with AI rather than competing against it. Layer on the durable skills that no specific tool obsoletes. Verification. Systems thinking. The ability to learn under uncertainty. The patience to build relationships. And separate the question of whether you will get hired from the question of whether you will earn enough. Both matter. Neither answers the other.
You are entering a labor market in the middle of a transition we have not finished naming. The people building AI think the labor-market shock is real and large. They disagree about timing, magnitude, and outcome. They do not disagree that it is happening. The capital being deployed (Terafab, Stargate, Cortex, the gigawatt-per-week buildout) is what determines how capable AI will be in 2030. The answer is much more capable than today.
Here is the good news. The people who think clearly about all of this and act early have an enormous advantage over the people who do not. The transition is not a tide that sweeps everyone equally. It is a long, uneven, occasionally violent reorganization of work. The people who understand its shape will do well inside of it.
If you found this useful, share it with someone in college, someone advising a kid in college, or someone running a college. The conversation we need to have about the next fifteen years is too big for any one of us to figure out alone.