The Age of AI: Compared to the Industrial Revolution and the Age of the Internet

13.1  What a general-purpose technology is

Economists call a technology a general-purpose technology — a GPT, in a usage that long predates the language-model sense of the acronym — when it satisfies three conditions: it is pervasive, spreading across many sectors; it keeps improving over decades; and it spawns complementary innovation, enabling inventions that would not otherwise exist. Steam, electricity, and information and communications technology qualify. Nicholas Crafts examines whether AI does, from an explicitly historical standpoint.1

The framework is more useful than any analogy, because unlike an analogy it comes with a testable prediction: GPTs produce large productivity effects, but with a substantial lag.1 That is the forecasting tool. It tells you not to expect the effect on the schedule the excitement suggests, and equally not to conclude from an absent effect that nothing is happening.

13.2  The lesson of the dynamo

This is the segment worth building the whole episode around, and it comes from a 1990 paper by Paul David that has aged extraordinarily well.2

Electric motors were available to American manufacturers from the 1880s. Factory productivity did not respond for roughly forty years. The reason was not that the technology was bad; it was that firms installed it wrongly, and installing it wrongly was the rational thing to do.

Steam-powered factories were built around a single enormous engine driving a system of overhead shafts and belts. Machines had to be arranged by their power requirements and their proximity to the driveshaft — the building's layout was dictated by the physics of power transmission. When electricity arrived, the obvious move was to replace the steam engine with a large electric motor driving the same shafts. This worked, and it delivered almost nothing.

The gains came only when firms grasped that electricity permitted individual unit drive — a small motor on each machine. That removed the driveshaft, which removed the constraint on layout, which allowed factories to be arranged around the flow of work instead of the distribution of power. Buildings could be single-storey and wide rather than multi-storey and cramped. That reorganisation is where the productivity lived.2

The pattern recurs. The first barcode was scanned in the mid-1970s; it took roughly thirty years of complementary investment across the supply chain — scanners, inventory systems, standards, reorganised logistics — before the payoff arrived.3

FIG-13-01

Same power source, two factories

A two-panel cutaway of a factory floor. Left — "electrified, unchanged": a multi-storey building with one large electric motor sitting exactly where the steam engine was, still driving overhead shafts and belts; machines crowded along the driveshaft; work path drawn as a tangled line. Right — "re-architected": a wide single-storey building, a small motor on each machine, no overhead shafting, machines arranged in a clean production line; work path drawn as a straight arrow. Annotate: same technology · same decade · the productivity is in the right-hand panel.

Aspect 16:9 · 3200×1800 px · Brand palette: navy #00040c, silver-blue #a8c0d8, column gold #a87848 · Style: technical cutaway illustration, period-appropriate but clean

13.3  The Solow paradox and the productivity J-curve

Robert Solow's 1987 remark — that the computer age was visible everywhere except in the productivity statistics — named the same puzzle for information technology. It resolved in the 1990s, when measured productivity growth finally accelerated, roughly on the schedule the dynamo precedent would have predicted.3

Brynjolfsson, Rock, and Syverson give the mechanism a formal shape and a name: the productivity J-curve.4 Their argument runs as follows. A GPT requires large complementary investments — business-process redesign, co-invention of new products and business models, and human capital. Most of those investments are intangible and are poorly captured in national accounts. So while firms are accumulating that intangible capital, real output is being produced and not measured, which makes productivity growth look worse than it is. Later, when the benefits are harvested, measured productivity growth overstates the true rate. The result is a J-shaped path.4

They quantify it for the previous cycle: adjusting for intangibles associated with computer hardware and software puts total factor productivity about 15.9% higher than official measures by the end of 2017.4

13.4  Where the analogies break

Three differences worth naming, because a framework used carelessly is as misleading as an analogy.

Speed and reach

Previous GPTs spread sector by sector, gated by physical infrastructure. Generative AI propagated across essentially all industries and geographies at once, through digital infrastructure that already existed.5 If diffusion is faster, the lag may be shorter — though re-architecting an organisation is a human and institutional process, and there is little evidence that that has got faster.

Target

Steam and electricity automated physical work; ICT automated clerical and computational work. This one targets cognitive work directly. Whether cognitive work re-architects like a factory floor is an open empirical question, not a settled one.

Distribution of the gains

Here the historical record is least comfortable and most relevant. Economic historians caution that the essence of the first Industrial Revolution was not short-run productivity growth but the invention of a new method of invention — and that its early decades saw output per worker rise substantially while typical wages stayed largely flat. Robert Allen named this period Engels' Pause: between roughly 1780 and 1840, British output per worker rose about 46% while real wages rose about 12%. The gains accrued to profits, which financed the capital accumulation that the new technology required; wages only rose in line with productivity after the middle of the nineteenth century.6

Both halves of that matter. If AI raises the productivity of research itself, the compounding argument is real — that is the strongest version of the transformation case from Lesson 12. And the distributional warning is equally real: a sixty-year gap between rising output per worker and rising typical wages is not a footnote. It was most of a working life.

Demonstration protocol

A static graphic rather than a screen recording.

1. Three GPTs on one timeline, with the lag between introduction and measured productivity effect marked on each as a shaded span.
2. Add AI's introduction date and leave the effect end of the bar deliberately open, with a question mark. Do not fill it in — the honesty of the empty bar is the point.
3. Overlay the Engels' Pause figures — 46% against 12% — as a second register on the earliest bar.
4. Close on the practical question: what would organisational re-architecture look like where the viewer works? That is the transferable output.