Technology and Social Change

Lecture 15: Artificial Intelligence and the Political Economy of Technological Change

Bogdan G. Popescu
bgpopescu@tec.mx

Tecnológico de Monterrey

Part 1: The Puzzle

AI and the Political Economy of Technological Change

Will AI create shared prosperity—or concentrated gains?

From steam engines to large language models: the same political-economy questions

The same AI can lift wages or replace workers — depending on institutions and choices about deployment

Same fundamental questions as the Industrial Revolution

A modern car assembly line: automation has reshaped manufacturing for decades. AI now extends this transformation to cognitive work. Source: Marek Slusarczyk / Wikimedia Commons (CC BY 3.0).

The Productivity-Pay Gap

Figure 1

Why This Matters

The chart is not just a number — it is the lived experience of a generation of workers.

  • Cashiers replaced by self-checkout
  • Paralegals worried about generative AI
  • Manufacturing workers displaced by robots

Now AI raises the same question on a much larger scale. Will it lift wages, or widen the gap further?

Three Guiding Questions

1. Why did productivity rise while wages stagnated after 1980?

2. Will AI raise worker productivity — or replace workers?

3. Will AI’s gains be broadly shared — or concentrated among a few firms and workers?

Learning Objectives

By the end of this lecture, you should be able to:

  1. Explain the productivity bandwagon and the two conditions it requires
  1. Use historical evidence (the Engels Pause) to understand AI’s potential trajectory
  1. Evaluate empirical evidence on whether AI complements or replaces workers
  1. Identify policy levers that can redirect AI toward shared prosperity

Part 2: The Framework — Then and Now

The Productivity Bandwagon

The popular belief:

Technology raises productivity → firms expand → labor demand rises → wages follow.

Wages “ride along” on the productivity bandwagon.

But Acemoglu & Johnson (2023) ask: Is this automatic? Or does it depend on conditions that can fail?

Two conditions must hold for the bandwagon to work.

Condition 1: Does Technology Need Workers?

Marginal productivity: does the new technology raise the value of an additional worker, or replace workers altogether?

Complement → workers more valuable

  • Spreadsheets made accountants more productive — demand for accountants grew

Substitute → workers less valuable

  • Self-checkout kiosks replaced cashiers — demand for cashiers fell

Question for AI

Which tasks does it complement, and which does it substitute?

Condition 2: Do Workers Have Bargaining Power?

Even when technology needs workers, gains may not reach them. Workers need power to capture them.

Strong bargaining power → workers share gains

  • Postwar US (1948–1979): unions, tight labor markets, rising minimum wage — productivity and wages rose together

Weak bargaining power → owners keep gains

  • Post-1980: union decline, globalization, weakened labor standards — productivity rose, wages didn’t

Question for AI

Who controls AI deployment — workers or just firms?

When the Bandwagon Breaks

The bandwagon needs BOTH conditions. If either fails, productivity rises but wages stagnate.

  • Condition 1 fails → technology substitutes for labor
  • Condition 2 fails → workers lack the power to capture gains
  • Both can fail at once → maximum inequality

The post-1980 chart you just saw shows exactly this pattern.

The Engels Pause: 1790–1840

During early industrialization (Britain, 1790–1840):

  • Productivity rose substantially as factories spread
  • But real wages stagnated for 50+ years
  • Living conditions worsened: child labor, urban slums, long hours

Friedrich Engels documented this gap. Economic historians now call it “the Engels Pause.”

Friedrich Engels (1820–1895): his observations on workers’ conditions during industrialization gave the “Engels Pause” its name. Source: Wikimedia Commons (public domain).

Why the Bandwagon Broke Then

Both conditions failed simultaneously:

Condition 1 failed: early factory machinery substituted for skilled artisans (handloom weavers replaced by power looms)

Condition 2 failed: workers had almost no bargaining power

  • Unions were illegal (Combination Acts, 1799–1824)
  • Surplus rural labor flooded cities — workers were easy to replace
  • No vote, no welfare state, no factory inspections

Result: productivity gains went to industrialists, not workers — for half a century.

Why the Bandwagon Eventually Worked

After 1840, both conditions were gradually restored — through deliberate institutional change:

  • Trade unions legalized (1824, 1871) — restored Condition 2
  • Mass public schooling (Forster Act 1870) — raised worker skills, restored Condition 1
  • Suffrage extended (1867, 1884) — gave workers political voice

By 1900, productivity gains were broadly shared.

The lesson: the bandwagon does not run automatically. Institutions decide whether it carries everyone.

Part 3: AI Today — What Does the Evidence Show?

What Does AI Do to Tasks?

Recall from Lecture 11: jobs = bundles of tasks; technology affects tasks, not whole jobs.

AI is different from earlier automation in important ways:

Dimension Past Automation AI
Tasks affected Routine manual Cognitive + creative
Workers displaced Factory, farm Professionals, analysts
Initial effect Skill destruction Skill compression
Speed Decades Months to years

The Engels Pause took 50 years. AI may compress that timeline — because cloud deployment is near-instant, AI affects cognitive tasks across many sectors at once, and adoption requires no new factories.

Two Big Questions about AI

We have a framework. Now we ask the same two questions of AI:

Question 1 (Condition 1): Does AI complement workers — or substitute for them?

Question 2 (Condition 2): Are workers gaining bargaining power — or losing it?

The empirical evidence so far gives us partial, sometimes contradictory, answers.

Study 1: Brynjolfsson, Li & Raymond (2023)

Setting: One Fortune 500 company

  • 5,172 customer-service workers
  • About 3 million customer chats
  • AI tool: GPT-based chat assistant suggesting real-time responses

Design: AI assists workers, does not replace them

Outcome: Resolutions per hour (problems solved per hour)

Study 1: Results—Biggest Gains for Novice Workers

Figure 2: AI productivity gains are largest for lower-skill workers

Study 2: Noy & Zhang (2023)

Setting: Randomized experiment (online)

  • 453 college-educated professionals
  • Tasks: realistic writing (emails, reports, marketing)
  • AI tool: ChatGPT

Design: Treatment group uses AI; control group does not

Outcomes: Time to complete + quality of output (graded by humans)

Study 2: AI Compresses the Performance Gap

Figure 3: AI flattens the quality gap between strong and weak writers

Interpreting the Productivity Evidence

What do these studies tell us?

  • AI can raise worker productivity (Condition 1 may hold)
  • Current AI often complements rather than substitutes
  • Gains are largest for lower-skill workers (skill compression)

But:

  • These studies test AI as tool (worker + AI), not replacement
  • Distribution question remains: Do workers capture gains? (Condition 2)

Study 3: Robots and Jobs (Acemoglu & Restrepo, 2020)

U.S. local labor markets and industrial robots

  • +1 robot per 1,000 workers leads to:
    • Employment rate falls 0.20 pp
    • Average wages fall 0.42%
  • Effects strongest in manufacturing regions

Implication: Displacement was NOT offset by task creation locally

Source: Acemoglu & Restrepo (2020, Fig. 6)

Video: Acemoglu on AI and Inequality

Source: Acemoglu on AI and Inequality

  1. What historical example shows technology can increase inequality?
  2. What two main ways is AI likely to be used against workers?
  3. Why can we not rely on “the market” alone for AI outcomes?
  4. How is AI linked to threats to democracy?

The Evidence in Summary

How does each study map onto the two conditions?

Study Complement or Substitute? Tests Condition Bottom Line
Brynjolfsson, Li & Raymond (2023) Complement C1: marginal productivity AI raises productivity, especially for novices
Noy & Zhang (2023) Complement C1: marginal productivity AI compresses skill gaps in writing
Acemoglu & Restrepo (2020) Substitute C1: marginal productivity Robots reduced employment and wages locally

All three speak to Condition 1. None directly tests Condition 2 — whether workers actually capture the gains.

AI Exposure: Who Is Affected?

Eloundou, Manning, Mishkin & Rock (2024):

  • ~80% of U.S. workers have at least 10% of tasks exposed to LLMs
  • ~19% have at least 50% of tasks exposed
  • Unlike robots, AI hits high-skill cognitive workers hardest

This inverts the usual automation pattern

A modern data center: the physical infrastructure powering AI. The concentration of computing resources in a few firms raises questions about who controls AI’s economic gains. Source: CERN / Wikimedia Commons.

AI Exposure by Occupation Group

Figure 4: AI exposure is highest for cognitive occupations

Capability vs. Reality: The AI Adoption Gap

Figure 5: Theoretical AI capability far exceeds actual observed usage across all occupational categories

Early Warning: Young Worker Hiring Slowdown

Figure 6: Job finding rates for young workers in AI-exposed occupations show early decline

Caveat: This is early warning evidence, not definitive proof of AI displacement. The pattern is correlational — post-ChatGPT, in exposed occupations, for one age group — and consistent with multiple causes.

Macro Projections: Massive Disagreement

Source GDP Impact Method Horizon
Acemoglu (2025) +1.0% General equilibrium model; task-level substitution ~10 years
McKinsey (2023) +6.0% Productivity potential across use cases ~decade
Goldman Sachs (2023) +7.0% Macro projection (level effect on output) ~10 years

These estimates differ by an order of magnitude because of fundamentally different modeling choices, not disagreement about AI’s capabilities.

Takeaway: Be skeptical of any single headline number. The range itself is the information.

Who Is Actually Exposed? The Measurement Problem

Boix et al. (2026): Three popular AI exposure indexes yield contradictory sector rankings

Sector Felten Brynjolfsson Webb
Finance & Insurance 1 1 7
Agriculture 13 13 1
Manufacturing 9 8 2

Cells show each sector’s rank on each index (1 = most exposed to AI). The three measures use different methods — task-capability matching (Felten), suitability for machine learning (Brynjolfsson), patent-task overlap (Webb) — and disagree sharply on who is most affected.

If we cannot agree on who is exposed, forecasts about AI’s winners and losers remain deeply uncertain

New Task Creation: Grounds for Cautious Optimism

Autor, Chin, Salomons & Seegmiller (2024):

  • ~60% of employment in 2018 is in job titles that barely existed in 1940
  • New work creation has historically been the primary engine of employment growth
  • But the pace of new work has slowed since 1980, especially for less-educated workers

Hope for AI: If AI creates valuable new tasks, reinstatement can offset displacement—but this requires deliberate institutional effort

Part 4: The Choice

Institutional Responses: What Might Work?

Policy Condition Targeted
Education/retraining Condition 1: maintain marginal productivity
Collective bargaining Condition 2: restore rent sharing
Income support (UBI) Cushion displacement transitions
Antitrust/competition Reduce market power concentration
AI governance Shape technology development direction

No single policy suffices—combinations are needed, as during industrialization

Source: Author’s illustration based on Acemoglu & Johnson (2023)

Garment workers’ strike march, New York, 1958. Collective bargaining has historically been a key mechanism for workers to share in productivity gains. Source: Kheel Center, Cornell University / Wikimedia Commons (CC BY 2.0).

AI Concentration: Economic and Political

The economic risk — “winner-take-most”:

  • Foundation models cost billions to build — only a handful of firms can compete
  • Increasing returns to scale: more data → better models → more users → more data
  • Monopsony risk in AI-adjacent labor markets: few employers, many displaced workers — Condition 2 weakens

The political risk — regulatory capture (Boix et al., 2026):

  • Big Tech capex surged from $36B (2015) to $400B (2025) — over a third of all US private IT investment
  • Data centers and chips are immobile assets — AI firms now have strong incentives to lobby and shape regulation

Concentration is not just an economic problem — it shapes who decides what AI becomes.

Redirecting AI: From Displacement to Augmentation

Acemoglu & Johnson (2023): The direction of AI is a choice, not a given

  • Current incentives favor automation (tax-advantaged capital, cheap compute)
  • Alternative: human-complementary AI that creates new tasks
  • Policy levers: equalize capital/labor taxation, fund augmentation R&D, give workers voice in deployment

Will we create shared prosperity, or repeat the Engels Pause?

Summary: Key Takeaways

  1. The productivity bandwagon is not automatic — it requires marginal productivity AND bargaining power
  1. History warns: productivity can rise while wages stagnate for generations (the Engels Pause)
  1. Current AI evidence is mixed: AI often complements workers, but also displaces them, with uneven exposure
  1. Concentration is political: who controls AI shapes who captures the gains
  1. The direction of AI is a choice — institutions and policy can redirect it from displacement to augmentation

References (1/2)

  • Acemoglu, D. (2025). The simple macroeconomics of AI. Econ. Policy, 40(121), 13–58.
  • Acemoglu, D., & Johnson, S. (2023). Power and Progress. PublicAffairs.
  • Acemoglu, D., & Restrepo, P. (2019). Automation and new tasks. J. Econ. Perspect., 33(2), 3–30.
  • Acemoglu, D., & Restrepo, P. (2020). Robots and jobs. J. Polit. Econ., 128(6), 2188–2244.
  • Acemoglu, D., & Restrepo, P. (2022). Tasks, automation, and the rise in U.S. wage inequality. Econometrica, 90(5), 1973–2016.
  • Anthropic. (2025). The labor market impacts of AI: Evidence from real-world usage. Anthropic Research.
  • Autor, D. H. (2015). Why are there still so many jobs? J. Econ. Perspect., 29(3), 3–30.
  • Autor, D., Chin, C., Salomons, A., & Seegmiller, B. (2024). New frontiers. Q. J. Econ., 139(3), 1399–1465.

References (2/2)

  • Boix, C., Becher, M., González-Rostani, V., & Stegmueller, D. (2026). AI’s economy and its political and institutional consequences. APSA Task Force on AI and Political Science.
  • Boix, C. (2022). AI and the economic and informational foundations of democracy. In J. B. Bullock et al. (Eds.), The Oxford Handbook of AI Governance. Oxford UP.
  • Brynjolfsson, E., Li, D., & Raymond, L. R. (2023). Generative AI at Work (NBER WP 31161).
  • Brynjolfsson, E., Rock, D., & Syverson, C. (2021). The productivity J-curve. AEJ: Macro., 13(1), 333–372.
  • Dell’Acqua, F., et al. (2023). Navigating the Jagged Technological Frontier (HBS WP 24-013).
  • Eloundou, T., Manning, S., Mishkin, P., & Rock, D. (2024). GPTs are GPTs: Labor market impact potential of LLMs. Science, 384(6702), 1306–1308.
  • Goldin, C., & Katz, L. F. (2008). The Race between Education and Technology. Harvard UP.
  • Noy, S., & Zhang, W. (2023). Experimental evidence on the productivity effects of generative artificial intelligence. Science, 381(6654), 187–192.
  • Peng, S., Kalliamvakou, E., Cihon, P., & Demirer, M. (2023). The impact of AI on developer productivity: Evidence from GitHub Copilot. arXiv:2302.06590.

Appendix

Appendix: Automation and Inequality (Acemoglu & Restrepo, 2022)

U.S. economy, 1980–2016:

  • 50–70% of U.S. wage structure changes explained by automation
  • Wage declines largest in routine-task occupations
  • Concentrated in middle-skill occupations (job polarization)
  • Disproportionately affects non-college workers

Historical parallel: Artisan displacement during the Industrial Revolution

Appendix: AI and Coding Productivity (Peng et al., 2023)

Peng, Kalliamvakou, Cihon & Demirer (2023): GitHub Copilot RCT

  • Developers completed coding tasks 55.8% faster with AI
  • Randomized controlled trial with professional developers

Implications:

  • Large productivity gains in well-defined coding tasks
  • Consistent with Brynjolfsson and Noy & Zhang findings
  • Human judgment remains critical for code review and design

Appendix: Macro Projections Bar Chart

Figure 7: Estimates of AI’s macroeconomic impact vary by an order of magnitude