Historical Context
On Tap Intelligence
Every great technology has transformed work. Here’s what history tells us about what comes next.
AI is doing for cognitive capabilities what electricity did for physical power: turning a scarce, expensive resource into an on-demand utility available to anyone.
Every GPT Follows the Same Arc
Every major general-purpose technology (GPT, not to be confused with Generative Pre-trained Transformer) follows a predictable five-phase arcThis framework synthesizes several academic models of how general-purpose technologies reshape economies. The closest single source is Carlota Perez (2002), Technological Revolutions and Financial Capital, which models technology waves in recurring phases. The definitive academic treatment of GPT theory is Lipsey, Carlaw & Bekar (2005), Economic Transformations: General Purpose Technologies and Long-Term Economic Growth. The specific five phases here (emergence, diffusion, displacement, reorganization, new equilibrium) are an editorial synthesis, not a direct citation from any single paper.. The names change, but the shape is the same. Steam, electricity, computers: each transformed the labor market through the same sequence of emergence, diffusion, displacement, reorganization, and new equilibrium.
Click or hover on a phase to learn more
Move your cursor over the timeline to discover hidden insights
Each successive technology has diffused faster than the last.
Steam took 90 years to fully reshape the labor market. Electrification did it in 50. Computers in 40. Each arc below spans a technology’s full transition from emergence to new equilibrium, and each is shorter than the one before. AI is on pace to be the fastest diffusion of a general-purpose technology in recorded history: 100M users in 2 months, majority adult adoption in under 3 years. Hover any arc to compare.
Sources: Allen (2009); Gordon (2016); David (1990); Autor et al. (2003); McKinsey, St. Louis Fed, Microsoft
vs. 20–60 yrs historically
No systematic unemployment yet
These projections extrapolate from adoption speed. If the diffusion phase that historically took 10–25 years is happening in 1–3, the displacement and reorganization phases may compress as well. That said, prior GPTs show that adoption speed does not reliably predict impact speed. The internet reached 50% of households by 2000 but didn’t show clear labor market effects for another decade. Organizational restructuring, education systems, and policy still operate at human speed. See live adoption data → See enterprise adoption data →
But what if this time is different?
Previous technologies automated tasks, sometimes many tasks, but none could emulate the full range of human cognition. If AI progresses toward general intelligence, the historical pattern may break down. The relevant question shifts from which tasks get automated to where humans retain comparative advantage: efficiency gaps, roles where the human element is the value, and complementarities between cognitive and physical work. Just because AI can do something doesn’t mean human labor involving it disappears, but it will probably look very different.
Adapted from Alex Imas
The Four Revolutions
Four technologies. Four massive disruptions. All eventually created more jobs than they destroyed, but the path was never smooth or quick.
ElectrificationClosest AI Analog
1880–1930This is the on-tap intelligence moment. AI transforms cognitive capabilities from scarce expert resources into utilities. The productivity gains will arrive later than expected, and through organizational redesign more than simple substitution. Brynjolfsson, Rock & Syverson (2021) call this the “Productivity J-Curve”: trillions in intangible investment are being made now but aren’t captured in measured output.
The On-Tap Intelligence Shift
Prior automation technologies had a consistent structure: they automated physical capabilities (steam, combustion, electricity) or rule-based cognitive tasks (computers). Each wave created a new protected domain, work the technology structurally couldn’t do, that workers could move toward. The task model of labor (Autor, Levy & Murnane, 2003) categorized this as the difference between routine tasks (codifiable, automatable) and non-routine tasks (requiring judgment, context, creativity).
AI breaks this pattern. Large language models and multimodal AI systems perform tasks that are simultaneously cognitive and ostensibly non-routine: legal analysis, medical reasoning, strategic synthesis, creative writing, code generation. This doesn’t mean AI equals human intelligence (it doesn’t), but it means the protected domain of prior automation waves is now being encroached on.
The most useful analogy is electrification. Before electricity, accessing significant mechanical power required physical proximity to a power source (a river, a steam boiler). Power was scarce, locationally fixed, and expensive. Electrification transformed power into a utility: standardized, reliable, available on demand anywhere on the grid, priced per unit of use. On-tap power.
AI performs this same transformation for cognitive capabilities. Legal analysis was previously accessible only to those who could pay $400/hour for a lawyer in a major city. Medical reasoning lived in academic medical centers. Strategic insight required expensive consulting firms. AI threatens to make these capabilities available to anyone with internet access, at marginal cost approaching zero. On-tap intelligence.
Crucially, when electricity made factory power cheap and ubiquitous, the result wasn’t fewer factories — it was dramatically more. Once factories adopted “unit drive” (individual electric motors per machine, replacing centralized shaft-and-belt systems), they could be single-story, lighter, and modular. Electrification accounted for half of all manufacturing productivity growth in the 1920s. Manufacturing output and employment grew for decades because cheaper power made previously unviable production economically feasible. The same dynamic may apply to cognitive work: when intelligence becomes on-tap, the question is whether there’s unmet demand for cognitive output. In sectors like software, creative services, and data analysis, the answer is emphatically yes. This is the demand elasticity effect: lower costs unlock new markets, potentially creating more total work than they eliminate.
From Scarce to On-Tap
| Physical Power | Cognitive Capability | |||
|---|---|---|---|---|
| Pre-Electricity | Post-Electricity | Pre-AI | Post-AI | |
| Access | Near a water wheel or steam engine | Any electrical outlet | Expensive credentialed experts | Any smartphone |
| Cost | High capital, fixed infrastructure | Per unit of use | $200–$700/hour | Near zero marginal cost |
| Geography | Mill towns, factory districts | Anywhere on the grid | Major metros with legal/medical/finance clusters | Anywhere with internet |
| Competitive advantage | Owning the power infrastructure | Using power intelligently | Accessing expert talent | Using AI intelligently |
What the Pattern Predicts
History doesn’t tell us the outcome. It tells us the shape. Here is what the pattern predicts, offered not as certainties but as the most historically-grounded expectations.
Near Term
- Labor market polarization accelerates — already visible in 2024–2025 hiring data for content, coding, and customer service roles
- Routine cognitive work faces displacement pressure first: document review, standard writing, basic code, data analysis, customer service
- Wage compression and employment decline appear in these categories as AI-assisted workers handle dramatically more volume
- New roles in AI oversight, training, and application emerge — but won’t immediately compensate for losses
- With 88% of organizations already adopting AI (McKinsey, 2025) and 39% of US adults using genAI within two years of ChatGPT’s launch (NBER, Bick et al.), the Solow Paradox may resolve in years rather than the decades it took for electricity and computers
Medium Term
- The Productivity J-Curve (Brynjolfsson, Rock & Syverson, 2021) resolves — trillions in intangible investment become visible as organizational complements to AI develop (new business models, new processes, new educational pathways, new regulatory frameworks)
- Aggregate productivity growth becomes visible and accelerates — potentially arriving by the early 2030s rather than the 2050s a historical baseline would predict
- Whether this growth translates to broadly shared wages depends entirely on institutional choices being made now
- AI’s low infrastructure requirements could spread gains more geographically than prior GPTs, which concentrated in industrial centers
Long Term
- A new occupational equilibrium emerges, dominated by human-AI collaborative roles
- Growth in human-specific services: caregiving, high-stakes physical tasks, relational work that requires accountability and presence
- Entirely new industries emerge, enabled by democratized access to cognitive capabilities — analogous to how electrification created the consumer appliance economy and how the GI Bill (8 million veterans educated by 1956) unlocked the postwar knowledge economy
- Average wages in the new equilibrium may be substantially higher — but the compressed transition means less time for workers and institutions to adapt
Historical calibration: In prior GPT transitions, the total timeline from meaningful diffusion to new equilibrium ranged from 40 to 70 years. AI’s diffusion phase is running at roughly 10x the speed of prior GPTs: enterprise adoption went from 33% to 88% in two years (McKinsey), and 39% of US adults used generative AI within two years of launch (NBER, Bick et al. 2024) — nearly double the PC adoption rate at a comparable stage. Deming, Ong & Summers (NBER, 2025) find that 1990–2017 was the least disruptive labor market period since 1880, suggesting the pace may now be accelerating sharply. The timelines above extrapolate from this 10x adoption speed. If the historical 40–70 year arc compresses proportionally, the full transition from diffusion to new equilibrium could play out in 7–15 years — a single generation rather than three.
Occupational Vulnerability Snapshot
Five vulnerability categories, grounded in the historical pattern of how general-purpose technologies reshape occupational structures.
| Category | Examples | Historical Analog | Timeline |
|---|---|---|---|
High vulnerability (near-term) | Paralegals, data analysts, basic copywriters, customer service, junior programmers | Clerical workers facing computerization (1970s–90s): 167,000+ telephone operators eliminated; travel agents fell from 270,000 to under 50,000 | 1–3 years |
High vulnerability (medium-term) | Financial analysts, radiologists, contract lawyers, accountants | No exact prior analog — the novel AI encroachment on non-routine cognitive tasks. Acemoglu & Restrepo find 50–70% of US wage changes over 4 decades trace to automation of routine tasks; AI now reaches non-routine work | 2–5 years |
Moderate (complementarity likely) | Managers, therapists, senior educators, executives, architects | Knowledge workers complemented by computers | 3–7 years, with transformation |
Lower vulnerability | Surgeons, plumbers, electricians, elder care, social workers | Skilled trades in the industrial era | 10+ years |
Emergent / new creation | AI trainers, AI governance specialists, human-AI collaboration designers | Electrical engineers in the electrification era; STEM employment has grown 50%+ since 2010 (Deming, Ong & Summers, 2025) | Growing now |
Note: Vulnerability does not mean elimination — it means transformation. Most “displaced” occupations in prior GPT transitions were reorganized, not eliminated. The exception: occupations where the technology directly substitutes for the core input (handloom weaving → power loom; data entry → AI processing). For these, the historical record offers little reassurance.
Want to see how your specific job breaks down? Explore the Task Visualizer to analyze task-level AI exposure for 100+ occupations.
What History Actually Proves
The technology doesn’t decide. We do.
Lesson 1: Invest in Complements, Not Preservation
Every successful institutional response to a GPT transition invested in workers’ capacity to participate in the new economy, not in protecting the old one. The Morrill Act (1862) created land-grant universities that provided the skilled workforce for industrialization. The GI Bill (1944) sent nearly 8 million veterans through education programs by 1956; by 1947, WWII veterans accounted for half of all college admissions. The VA estimated that increased federal income taxes from these better-educated workers paid for the program several times over. Community colleges, whose growth the GI Bill accelerated, now serve 44% of US undergraduates and are the nation’s primary workforce development engine. The equivalent for AI: radical investment in AI literacy, domain-expert + AI collaboration skills, and accessible retraining pathways.
Lesson 2: The Distribution Problem is Institutional, Not Technological
The computer era’s inequality was not technologically inevitable. Goldin & Katz document the college wage premium rising from 39% (1980) to 79% (2000), while real wages for men without degrees declined. Acemoglu & Restrepo (2022) find that 50–70% of US wage structure changes over four decades trace to automation displacing routine-task workers. But this reflected specific institutional choices: declining union density, wage policy, trade liberalization, corporate governance norms. Ford’s Five Dollar Day (January 5, 1914) is the counterexample: he raised the prevailing $2.34 to $5.00 per day because mass production requires mass consumers. The AI era’s distributional outcome will similarly reflect choices made now.
Lesson 3: The Gains Are Real. The Timeline is Not What You Think.
Every GPT ultimately created more jobs than it destroyed and raised average wages. This is true and important. It is also true that the prior pattern’s timeline (40 to 70 years) is politically and humanly unacceptable as a response to workers experiencing disruption today. “Eventually” is not a policy.
If on-tap intelligence enables the democratization of expertise, putting the equivalent of world-class legal, medical, educational, and financial guidance within reach of everyone rather than only the affluent, it could be among the most equalizing forces in human history. If it primarily displaces workers while concentrating gains among capital owners and a small elite of knowledge workers, it could be among the most destabilizing. The technology does not decide. We do.
Source Data
Where Does “5–12×” Come From?
The compression ratio depends on what you compare and which source you trust. Here are the three main comparisons, with ranges reflecting measurement uncertainty.
| Comparison | Historical | AI | Ratio | Source |
|---|---|---|---|---|
| vs. Electrification | ~40 yrs from dynamo (1880s) to productivity gains (1920s); only 3% of residences had electric lighting by 1900 | ~2–3 yrs: 39% of US adults using genAI by Aug 2024; ChatGPT reached 100M users in 60 days | 8–12× | David (1990); Bick, Blandin & Deming (2025); Epoch AI |
| vs. PC / Internet | PC: 20% adoption 3 yrs after IBM PC (1981); Internet: 20% two yrs after commercial opening (1995) | GenAI: 39% adoption 2 yrs after ChatGPT launch, nearly double the PC rate at a comparable stage | 2–5× | Bick, Blandin & Deming (NBER 32966, 2024); Census |
| vs. Enterprise adoption | 10–25 yrs historically | 33% → 88% in 2 yrs (McKinsey); but Census BTOS shows only 17.5% of firms actively using AI (Feb 2026) | 5–12× | McKinsey AI Survey (2023–2025); Census BTOS (2026) |
Note: Census Bureau’s rigorous sampling shows only 17.5% of US firms actively using AI (Feb 2026), a 5× gap with McKinsey’s 88%. The range reflects this measurement uncertainty. Adoption speed also does not guarantee proportional impact speed.
Key Sources
The History of Technological Anxiety and the Future of Economic Growth: Is This Time Different?
Joel Mokyr, Chris Vickers, Nicolas L. Ziebarth · Journal of Economic Perspectives, Vol. 29, No. 3 (2015)
The anchor reference for this page. Surveys 250 years of technological anxiety across three recurring themes: displacement fears, work dehumanization, and stagnation pessimism. Finds that predictions of widespread technological unemployment were 'by and large, wrong' but cautions against trivializing the costs borne by those actually displaced.
Technological Revolutions and Financial Capital
Carlota Perez · Edward Elgar Publishing (2002)
The theoretical framework behind the five-phase GPT arc (Section 01). Models technology waves as recurring surges with installation and deployment periods separated by a turning point.
The Computer and the Dynamo: An Historical Perspective on the Modern Productivity Paradox
Paul David · American Economic Review, Vol. 80, No. 2 (1990)
Origin of the electrification productivity paradox cited in the Electricity card. Showed that 40 years elapsed between the lightbulb and measurable productivity gains from electrification.
Economic Possibilities for our Grandchildren
John Maynard Keynes · Essays in Persuasion (1930)
Keynes predicted a 15-hour work week within 100 years, driven by compound productivity growth. He saw technological unemployment as 'growing-pains of over-rapid changes' rather than a permanent condition.