How New Jobs
Are Born
An empirical essay on how new occupations emerge after a general-purpose technology — across six revolutions, from steam to AI, with a particular focus on the labour market of 2026.
For two centuries the same argument has returned each time a major technology arrives. Machines will destroy work. Machines will create new work. The first claim is half-right — old occupations do disappear. The second is half-right — new ones eventually appear. What neither claim tells us is which new occupations, in which order, on what timescale, and what is missing right now.
This is the question I set out to answer. I studied five completed technological revolutions — steam, electricity, automobiles, computing, the internet — and one in progress, AI. I coded the new occupations each one created, dated their emergence from primary sources, and looked for a structure.
The structure that emerged is what I call a five-phase genesis sequence. After every general-purpose technology, new jobs appear in the same order: Operator (people who run the new machines), Interface (people who design the human-machine boundary), Orchestrator (people who coordinate complex systems built on top), Curator (people who interpret, audit, and assign meaning to outputs), and Social Repair (people who tend the social damage the diffusion produces). The order is empirical. The timing varies.
Applied to AI in 2026, the framework yields a clear and uncomfortable finding. Phases I and II are saturating. Phase III is forming. Phase IV — the audit, ethics, and model-trust functions — is the acute scarcity. Phase V is widely misread: about 1% of postings are AI-tagged directly, but a much larger 10–15% indirect pool is forming in reskilling, L&D, and care work and is not yet measured as “AI jobs.”
A recurring order — Operator, Interface, Orchestrator, Curator, Social Repair.
From Steam (1780) to AI (2018–present). Five completed; AI is open.
Average gap between Phase I emergence and Phase V stabilisation.
AI Phase V today — reskilling, L&D, care — uncounted as "AI jobs".
Five phases. One order. Six revolutions confirm it.
The argument of this essay rests on a single empirical claim: new occupations after a general-purpose technology emerge in a fixed order. I name the five phases and define what each one solves.
A general-purpose technology — steam, electricity, the combustion engine, the digital computer, the internet, and now AI — does not displace labour uniformly. It first creates a frontier of new tasks: a class of jobs that did not exist before because the technology had not yet existed. Across my six cases, these new tasks cluster into five functional categories that appear in a recurring sequence.
The pattern is not a metaphor. It is the result of bottom-up coding of occupational records — BLS series since 1900, secondary historical reconstruction before that — against the dates at which a recognisable category first appears in those records. Phase boundaries are coarse, and phases overlap; the order of their centre of gravity does not vary.
Each phase answers a scarcity that the previous one produced. Operators require interfaces. Interfaces require coordination. Coordination requires trust. Trust, once breached at scale, requires repair. This is why the sequence holds even across radically different technologies: the scarcities themselves recur.
Operator
People who run the new machines. Manual operation, maintenance, raw labor at the interface with the technology.
Phase 1 jobs absorb labor displaced from older systems. They are physically or cognitively close to the technology and create the first beachhead of employment.
- Steam (1830-1900): Firemen · Stokers
- Electricity (1880-1930): Linemen · Electricians
- Automobile (1908-1945): Assembly-line workers · Mechanics
- Computing (1945-1980): Keypunch operators · Tape librarians
- RLHF labelers
- Data annotators
- Prompt engineers (basic)
- AI model trainers
The pattern recurs.
From steam to the internet, the same sequence appears in every completed revolution. AI is the open case. This section sets out the evidence.
History does not repeat, but it rhymes — and the rhyme is more disciplined than one usually thinks. Every general-purpose technology since 1780 has produced the same five-phase progression. The bars below mark the empirical emergence window of each phase, drawn from contemporary occupational records, census categories, and trade publications of the period.
What changes is the pace. Steam needed eighty years. The internet did the same work in twenty. AI, on year eight, is roughly where the internet was in 2003 — Phases I to III dense, Phase IV emerging, Phase V largely absent from the visible labour market.
The order is empirical. The timing is variable. Phase V always arrives last — and always arrives.
AI
Phases 1-3 are forming rapidly. Phase 4 is the acute bottleneck. Phase 5 is structurally not yet formed at scale — but indirect labor (reskilling, L&D) is large.
AI Engineer postings on LinkedIn grew over 100% YoY in 2024-2025 — the fastest Phase 2 emergence on record.
2024-2025LinkedIn Economic Graph 2025Responsible AI postings tracked by Indeed represent ~1% of AI-tagged jobs (mid-2025).
2025Indeed Hiring Lab, Responsible AI trackerWEF Future of Jobs Report 2025 forecasts 170M new roles created and 92M displaced by 2030 — the Phase 5 cohort is the missing piece.
2030 (forecast)WEF Future of Jobs Report 2025
Dense at the base. Scarce at the top. Absent at the summit.
The visible AI labour market mid-2026, mapped against the five-phase framework. What is forming, what is bottlenecking, what is missing.
Treat the AI labour market as an aggregate and it looks like a boom: postings up, salaries up, headlines about AI Engineer being the hottest title of the decade. Disaggregate it by phase and the picture changes. The boom is concentrated in Phases I and II — the people who run and front the system. The phases that come next are emerging slowly, and one of them is in acute shortage.
The aggregate analysis hides the structure. Phase I (annotators, basic prompt work) absorbs about a third of all AI postings — labour that did not exist five years ago and that will, by the historical pattern, commoditise within another five. Phase II (AI engineers, designers, product managers) absorbs another quarter and is still expanding rapidly. Phase III (orchestration — MLOps, agent-system design) is forming and will reach maturity around 2028.
Phase IV is the operational story of 2026. Audit, ethics, red-teaming, and model-risk roles together represent only 5–8% of AI postings — yet the EU AI Act will compel firms above a threshold to staff exactly these roles starting 2026–2027. The supply of qualified hybrid technical-legal talent is small, slow to train, and not produced by any standard university pipeline. This is the cycle’s bottleneck, and the firms that solve it first will set the terms for the rest.
The case of Phase V — misread by almost everyone.
Casual readings of the labour market conclude that Phase V is empty. Indeed, fewer than 1% of AI-tagged postings carry « Responsible AI », « AI Safety », or « AI well-being » in the title. That number is real but misleading.
A much larger pool — somewhere between 10% and 15% of the AI-adjacent labour market, by my estimate — is doing what previous Phase V cohorts did: reskilling displaced workers, redesigning organisations around AI workflows, providing the care and support around an upheaval that productivity statistics cannot capture. These workers do not call themselves « AI workers ». Their employers do not tag the roles as AI. The trackers miss them.
What is structurally absent is the acute Phase V cohort — the clinical and policy specialists who, in previous cycles, eventually appeared to repair specific harms. AI-induced cognitive labour strain, organisational burnout, model-mediated trust collapse: these will produce occupations. They have not yet.
Phases I–III are dense — AI is rebuilding the labour base. Phase IV is acutely scarce. Phase V has a small direct cohort and a much larger indirect one not yet measured.
Phases I–II ranges are [estimated] — consistent across sources. Phase IV–V ranges are [illustrative] — there is real measurement gap, especially for indirect labour.
LinkedIn Economic Graph, Indeed Hiring Lab, OECD AI Index, Stanford AI Index, WEF Future of Jobs 2025, BLS occupational employment series.
What would prove the framework wrong.
A working hypothesis only earns trust if it can be refuted. I take seriously the four strongest objections, then commit to six predictions — each with an explicit falsification test.
- STEELMAN
Prior revolutions augmented physical labor while leaving cognitive work intact. AI directly targets the cognitive bottleneck that humans occupied. There is no historical analog. The job-creation phases that produced UX designers and traffic engineers may have no AI equivalent because AI itself can do those jobs.
RESPONSEPartly valid. We accept that Phase 5 timing may be faster (more disruption) or slower (humans aren't displaced from Phase 5 because AI can't do emotional repair work). But Phases 1-3 are already empirically present and following the historical sequence (annotators → engineers → orchestrators). The framework's value is the sequence, not the timing.
Six forecasts. Each with a falsification test.
If the framework is right, these should land. If they don’t, I revise.
- 011Y·high confidence
AI Engineer postings continue YoY growth above 50% in OECD countries.
If wrong: If LinkedIn AI Engineer YoY drops below 20% before June 2026 without macro recession, the thesis on Phase 2 timing is weakened.
- 023Y·medium confidence
Phase 4 (Trust/Audit) becomes the labor bottleneck of the cycle: postings grow 3-5x faster than Phase 1-2.
If wrong: If Phase 4 share remains ≤ 8% by end-2029 despite AI Act enforcement, either the framework is wrong or regulatory teeth are weaker than expected.
- 033Y·medium confidence
A formal 'AI Auditor' professional certification appears (ISACA, Big 4, or new body).
If wrong: If no certification crosses 5,000 holders by 2029, professional self-organization is slower than the historical pattern.
- 045Y·low confidence
Phase 5 (Social Repair) direct cohort emerges visibly: AI-induced burnout coaches, cognitive rehab specialists appear in BLS occupational categories.
If wrong: If no formal occupational code appears by 2031, my timing estimate may be too aggressive (or measurement systems lag reality).
- 055Y·medium confidence
Indirect Phase 5 labor (L&D, reskilling, HR for AI transition) crosses 5% of total labor force in OECD countries.
If wrong: If OECD reskilling labor remains below 3% by 2031, the indirect Phase 5 story is overstated.
- 0610Y·low confidence
By 2036, full 5-phase distribution looks like prior revolutions: Phase 1 (10-15%), 2 (25-30%), 3 (25-30%), 4 (20-25%), 5 (10-15%).
If wrong: If Phase 4+5 stay below 20% by 2036, AI may be a different kind of GPT than the prior five (more concentrated, less labor-rebuilding).
Pick your hat.
A framework is only useful when it changes a decision. Choose your role to read the actionable implications — for capital, for company, for regulator, for individual.
Buy Phase 4 ahead of consensus. Avoid pure-play Phase 2 at 2026 multiples.
- 01
Overweight Phase 4 infrastructure
AI audit firms, model-risk platforms, red-teaming services, compliance tooling. The talent bottleneck is the entry barrier — those who solve sourcing capture pricing power.
- 02
Hedge Phase 2 saturation risk
Prompt-engineering / generic 'AI Engineer' bootcamps will face commoditization 2027-2029. The historical pattern shows Phase 2 wages compress within 8-12 years of emergence.
- 03
Watch Phase 5 emergence signals
First-mover advantage in AI-induced burnout coaching, cognitive rehab, reskilling-at-scale. Look for ESCO/BLS occupational code creation as a leading indicator.
- Don't pay 2026 Phase 2 multiples for late entrants.
- Don't fund 'AI Ethics theater' (consulting without enforcement mandate).
Run it on yourself.
The playbooks above speak to four roles in general. The widget below speaks to you specifically. Pick what you actually do, and read the answer.
Find where you sit on the curve.
Pick your role, your age, and your horizon. The framework returns the phase you occupy today, your commoditisation risk, and a single concrete recommendation.
Translates capability into product. High demand, will commoditise 2027-2029.
Your role will face significant commoditisation pressure within your horizon.
- 1Move into orchestration / agent systems
- 2Own production reliability for one AI feature
- 3Learn the platform layer (vector DBs, evals, telemetry)
How I built this.
Five completed revolutions form the historical sample. Phases were induced bottom-up by coding new occupations and clustering by economic function. The AI mapping aggregates 28 labour-market datapoints (LinkedIn, BLS, Indeed, OECD, Stanford, McKinsey, WEF).
25 sources.
Academic literature (Autor, Acemoglu-Restrepo, Bessen, Perez, Polanyi, Susskind), industry reports (WEF, McKinsey), labour data (LinkedIn, BLS, OECD, Stanford, Indeed), and primary historical sources.