• General-purpose technology (Bresnahan & Trajtenberg, 1995): why some technologies restructure the entire economy rather than just one task

• Factor prices (“Why Britain?”) & Engels’ Pause (Allen, 2009): why factor prices shape where and whether mechanization happens, and why growth need not raise wages

• The task-based framework (Acemoglu & Restrepo, 2019): why technology displaces tasks, not jobs, and why net employment can rise or fall

• The productivity bandwagon (Acemoglu & Johnson, 2023): the two conditions under which technology raises wages

• Legibility and privacy as power (Scott, 1998; Cohen, 2019): why technology that “sees” society shifts power toward the state

• Frey & Osborne (2017) vs. Arntz et al. (2016): counting whole jobs as automatable (47%) vs. counting tasks (9%) — the unit of analysis changes the conclusion entirely.

• AI as complement vs. substitute: Brynjolfsson et al. (2023) and Noy & Zhang (2023) find AI helps workers; Acemoglu & Restrepo (2020) find robots replace them. The same family of technology can do either.

• Realism vs. Liberalism: the security dilemma predicts an AI arms race, while complex interdependence predicts cooperation — the US–China nuclear pledge shows both forces operating at once.

• Germany vs. USA vs. Sweden: identical robots, opposite labor-market outcomes, showing that institutions, not the technology, decide who wins.

• Why did the Industrial Revolution begin in Britain? Allen’s (2009) factor prices: high wages + cheap coal made machines profitable there and nowhere else.

• Why did ATMs increase the number of bank tellers? The productivity effect in the task-based framework: cheaper branches –> more branches –> more tellers.

• Why did trade-shocked US regions turn to populism? Distributional conflict plus non-adjustment: displaced workers did not relocate, and the losses concentrated and persisted (Autor, Dorn & Hanson, 2013).

• Why does predictive policing reproduce discrimination? Performative prediction: training on arrest data creates a self-fulfilling feedback loop (Lum & Isaac, 2016).

• Identified a measurement error: Frey & Osborne (2017) conflated technical feasibility with economic adoption; switching to task-level analysis cut the “47% at risk” headline to 9% (Arntz et al., 2016).

• Spotted deep uncertainty: three leading AI-exposure indexes (Felten, Brynjolfsson, Webb) rank the same sectors contradictorily, so claims about “which jobs AI will hit” rest on shaky measurement (Boix et al., 2026).

• Compared competing projections: estimates of AI’s GDP impact diverge by an order of magnitude — Acemoglu (2025, +1%) vs. Goldman Sachs (2023, +7%) — because they rest on different modeling assumptions, not different data.

• L8 — The Industrial Revolution: the first general-purpose technology shock

• L9 — How Technologies Emerge: technology as a self-building, evolving system

• L11 — Labor & Automation: how technology reshapes work and wages

• L13 — AI, Surveillance & Human Rights: technology and state power

• L14 — AI & International Relations: technology and great-power competition

• L15 — AI & the Political Economy: will AI deliver shared prosperity?

• General-Purpose Technology (GPT) (Bresnahan & Trajtenberg, 1995): a technology that is pervasive, improvable, and spawns new innovations. Steam, electricity, and digital/AI are the three canonical GPTs.

• Why Britain? (Allen, 2009): innovation follows factor prices, not genius. High wages + cheap coal (i.e., expensive British labor but abundant, cheap domestic coal) made the spinning jenny profitable in Britain (30–40% return) but not elsewhere (2–3%). The knowledge existed everywhere; only Britain had the incentive.

• The factory system (Mokyr, 2002): an organizational innovation, not just a technical one — centralized power, supervision, time discipline (Thompson, 1967).

• Technology = “the programming of phenomena to our use.” A phenomenon must first be observed, then understood (science), then exploited (engineering). Magnetism was known for 2,000 years before Faraday (1831) made it usable.

• Combinatorial innovation: new technologies are assembled from existing ones (the 1886 Benz Motorwagen = engine + wheels + steering + transmission).

• Recursive innovation: “technology creates itself out of itself” — each new technology becomes a building block for the next, so innovation accelerates.

Net employment effect = (Productivity + Reinstatement) − Displacement

• Displacement (−): machines take over existing tasks
• Productivity (+): cheaper output expands demand (ATMs –> more bank branches –> more tellers)
• Reinstatement (+): new tasks emerge that didn’t exist before

• This is why the “lump of labor” fear is wrong: the amount of work is not fixed
• Frey & Osborne (2017) claimed 47% of jobs were at risk
• Using task-level analysis, Arntz et al. (2016) corrected this to 9%

• Germany: strong unions + codetermination + vocational training –> moderate inequality, job security
• USA: weak unions + market-led –> sharp inequality, job loss
• Sweden: active labor-market policy –> low inequality, smooth transitions

The China Shock (Autor, Dorn & Hanson, 2013):

• ~2.4 million US jobs lost, concentrated in specific regions
• Workers did not relocate; effects persisted >10 years
• Shifted those regions toward populism
• Same shock, but Auto Alley recovered (Japanese FDI) while furniture towns did not

• Legibility (Scott, 1998): states simplify society to control it. AI is “Seeing Like a State 2.0” — it tracks movement, networks, and emotions at near-zero cost.

• Privacy as a power resource (Cohen, 2019): privacy is not secrecy; it limits the information asymmetry between citizens and institutions. AI widens that asymmetry.

• Performative prediction (Lum & Isaac, 2016): predictive policing trains on arrest data (not crime), then sends police where it predicts — a self-fulfilling feedback loop that reproduces past discrimination.

• An arms race no one wants (the security dilemma, Realism):
  • The US limits chips to feel safer
  • China sees a threat and races to catch up
  • The US clamps down harder
  • Each side’s defense looks like aggression to the other

• Whoever controls the chips has the power: the AI supply chain runs through three places —
  • Designs — US (Nvidia)
  • Factories — Taiwan (TSMC, 92% of advanced chips)
  • Machines — Netherlands (ASML, 100% of EUV)
  • And unlike oil, you can’t stockpile chips

• You can’t tell if a rival is cheating (the verification problem):
  • Missiles can be counted, inspected, and verified
  • But AI is just software — invisible, easy to copy, deployable in secret
  • So no one can prove a state is breaking the rules

• The result:
  • 166 countries voted to regulate killer robots (LAWS — lethal autonomous weapons)
  • But a few (e.g. Russia, North Korea) opposed
  • And no binding treaty exists — when cheating can’t be caught, defection wins

• AI can complement workers (makes them more productive):
  • Call-center output +34% for novices (Brynjolfsson et al., 2023)
  • The writing-quality gap between weak and strong writers shrinks (Noy & Zhang, 2023)

• But AI can also substitute for workers (replaces them):
  • +1 robot per 1,000 workers –> employment −0.20pp, wages −0.42% (Acemoglu & Restrepo, 2020)
  • ~80% of workers have ≥10% of tasks exposed to LLMs — and this time it hits cognitive jobs hardest (Eloundou et al., 2024)

• The Engels’ Pause recurs. L8 (Britain), L11 (the China shock & polarization), and L15 (the post-1979 decoupling, and perhaps AI) are the same story: productivity can rise while wages stagnate when institutions don’t share the gains.

• Institutions are the hinge. The same technology de-democratizes or re-equilibrates (L13), starts an arms race or enables cooperation (L14), and concentrates or shares gains (L11, L15). Technology is never destiny.

• Every technology creates winners and losers, and losers fight back — Luddites, the Stasi’s victims, trade-shocked towns, the Global South. Polanyi’s (1944) double movement (L9) is the general pattern.

1. Takes a clear, qualified position in the first sentence (“Technology raises wages only when institutions share the gains; AI is no exception…”).

2. Defines key terms precisely with an author (GPT, legibility, productivity bandwagon, security dilemma).

3. Uses specific cases that do analytical work — name the study, the number, the mechanism (not “AI is dangerous”).

4. Engages authors with each other — e.g., Brynjolfsson et al. (2023, complement) vs. Acemoglu & Restrepo (2020, substitute), then resolves the tension.

5. Signposts so the reader can follow the argument paragraph by paragraph.