The “Simple” Automation Model

The intuitive story (the one behind most AI anxiety today):

A machine does what a worker did → that worker loses the job → total employment falls.

Same logic the Luddites used in the 1810s when smashing power looms.

Why It Breaks Down — Three Hidden Assumptions

• Fixed amount of work (“lump of labor”). ATMs made branches cheaper → banks opened more of them → teller employment kept rising.
• Tech only destroys jobs, never creates them. “Software developer” and “data analyst” didn’t exist in 1900. Most jobs today were invented by past tech revolutions.
• Same tech, same outcome everywhere. Germany and the U.S. adopt robots at similar rates — yet inequality looks very different.

The Task-Based Framework

┌────────────────────────────────────────────────────────┐
│                    JOB DECOMPOSITION                   │
│                                                        │
│    "Accountant" = { data entry, analysis, client       │
│                     communication, judgment calls }    │
│                                                        │
│    Automation affects these tasks differentially:      │
│    • Data entry: HIGH substitution potential           │
│    • Analysis: PARTIAL augmentation                    │
│    • Communication: LOW substitution potential         │
│    • Judgment: VERY LOW substitution potential         │
└────────────────────────────────────────────────────────┘

Autor’s Task Taxonomy

The Acemoglu–Restrepo Framework

Technology hits labor through three forces pulling in opposite directions:

• Displacement (−) — machines take over tasks workers used to do. A robot replaces a welder on the assembly line.
• Productivity (+) — automation lowers costs → output expands → more demand for labor in the tasks machines don’t do. Cheaper cars → more jobs in design, sales, and repair.
• Reinstatement (+) — new tasks emerge where humans still hold the advantage. MRI technicians, software developers, AI trainers — jobs that didn’t exist before the technology.

Displacement-Reinstatement Model

%%{init: {'theme': 'base', 'themeVariables': {'fontSize': '20px', 'primaryColor': '#4a7c6f', 'primaryTextColor': '#1e293b', 'primaryBorderColor': '#334155', 'lineColor': '#64748b', 'secondaryColor': '#f9fafb', 'tertiaryColor': '#f9fafb'}, 'flowchart': {'useMaxWidth': true, 'nodeSpacing': 40, 'rankSpacing': 50, 'width': 1150, 'height': 650}}}%%
flowchart LR
    A["Technological<br/>Change"] --> B["Displacement<br/>Effect"]
    A --> C["Reinstatement<br/>Effect"]
    A --> D["Productivity<br/>Effect"]
    B --> E["Automates existing tasks<br/>(− Labor Demand)"]
    C --> F["Creates new tasks<br/>(+ Labor Demand)"]
    D --> G["Expands output<br/>(± Labor Demand)"]
    E --> H{"Net Employment<br/>Effect"}
    F --> H
    G --> H
    style A fill:#1e293b,color:#f9fafb,stroke:#334155
    style B fill:#b44527,color:#f9fafb,stroke:#334155
    style C fill:#4a7c6f,color:#f9fafb,stroke:#334155
    style D fill:#b7943a,color:#1e293b,stroke:#334155
    style H fill:#64748b,color:#f9fafb,stroke:#334155

Source: Based on Acemoglu & Restrepo (2019)

The Causal System

%%{init: {'theme': 'base', 'themeVariables': {'fontSize': '20px', 'primaryColor': '#4a7c6f', 'primaryTextColor': '#1e293b', 'primaryBorderColor': '#334155', 'lineColor': '#64748b', 'secondaryColor': '#f9fafb', 'tertiaryColor': '#f9fafb'}, 'flowchart': {'useMaxWidth': true, 'nodeSpacing': 50, 'rankSpacing': 50, 'width': 1150, 'height': 650}}}%%
flowchart LR
    A["Institutions<br/>(taxes, unions,<br/>education)"] -->|shape| B["Technological<br/>Change"]
    B --> C["Task<br/>Reconfiguration"]
    C --> D["Labor Market<br/>Adjustment"]
    D --> E["Distributional<br/>Outcomes"]
    E --> F["Political<br/>Responses<br/>(policy, parties)"]
    F -->|reshape| A

    style A fill:#4a7c6f,color:#f9fafb,stroke:#334155
    style B fill:#1e293b,color:#f9fafb,stroke:#334155
    style F fill:#b44527,color:#f9fafb,stroke:#334155

Causation is recursive: institutions shape technology and are reshaped by its consequences.

What Each Mechanism Explains

These are not competing theories. They explain different parts of the same labor-market transformation — each answers a different observed pattern:

Mechanism 1: Skill-Biased Technological Change → Rising College Premium

U.S. College Wage Premium, 1963–2022

Figure 1: College vs. high-school earnings gap

U.S. Job Polarization, 1980–2005

Figure 2: Employment share change by skill percentile

Mechanism 3: Geographic Concentration → Regional Decline

Mechanism 4: Productivity–Wage Decoupling → Stagnant Pay

Figure 3: Productivity grew far faster than compensation

The Institutional Mediation Thesis

Same Technology, Different Outcomes

%%{init: {'theme': 'base', 'themeVariables': {'fontSize': '18px', 'primaryColor': '#4a7c6f', 'primaryTextColor': '#1e293b', 'primaryBorderColor': '#334155', 'lineColor': '#64748b', 'secondaryColor': '#f9fafb', 'tertiaryColor': '#f9fafb'}, 'flowchart': {'useMaxWidth': true, 'nodeSpacing': 30, 'rankSpacing': 45, 'width': 1150, 'height': 650}}}%%
flowchart TD
    A["Industrial Robots<br/>(Same Technology)"] --> B["Germany"]
    A --> C["United States"]
    A --> D["Sweden"]
    B --> E["Strong unions<br/>Vocational ed<br/>Codetermination"]
    C --> F["Weak unions<br/>Limited training<br/>Market-led"]
    D --> G["Strong unions<br/>Active labor<br/>market policy"]
    E --> H["Moderate inequality<br/>Job security"]
    F --> I["Sharp inequality<br/>Job loss"]
    G --> J["Low inequality<br/>Smooth transitions"]
    style A fill:#1e293b,color:#f9fafb,stroke:#334155
    style H fill:#b7943a,color:#1e293b,stroke:#334155
    style I fill:#b44527,color:#f9fafb,stroke:#334155
    style J fill:#4a7c6f,color:#f9fafb,stroke:#334155

Source: Conceptual diagram based on Hall & Soskice (2001); Thelen (2014)

Path Dependence and Lock-In

Source: Arthur (1989); Pierson (2004)

Mokyr: Institutions Come Before Technology

So far: institutions shape the consequences of technology. Mokyr flips this — institutions also shape which technologies emerge in the first place.

• Useful knowledge spreads only when institutions let people share ideas openly.
• The Industrial Enlightenment — open science, patents, education — turned ideas into machines.
• Britain had this package; Song China and imperial Rome didn’t.
• Today: an AI revolution needs institutional innovation, not just better algorithms.

The China Shock — A Window on How Labor-Market Shocks Actually Work

The China Shock — A Window on How Labor-Market Shocks Actually Work

Case Evidence: The China Shock in Three U.S. Regions

Source: https://www.youtube.com/watch?v=SreJlZGd1c0

1. Auto Alley survived where Silicon Valley collapsed — what made the difference?
2. Why didn’t displaced workers move to growing regions or retrain into new jobs?
3. The video calls it “decimation” — what benefits does that frame leave out?

Expert Voice: David Autor on the China Shock

Source: https://www.youtube.com/watch?v=lUngEbyiaFs

1. What does Autor mean by “pie vs. slice” — and why does it matter?
2. Tech or trade? How does Autor’s “1980 vs. 2000” split reshape the debate?
3. Autor says we could have slowed it down — what policy tools were available?

Frey & Osborne (2017): Automation Risk

Headline finding:

“47% of U.S. employment at high risk of automation.”

• They labeled whole jobs “automatable” — but every job is made of many different tasks. A nurse gives injections (maybe automatable) AND comforts scared patients (not). Replacing one task doesn’t replace the nurse.
• “A machine could do it” is not the same as “a company will replace the worker.” Robots are expensive, customers often prefer humans, and regulators are slow. Just because something is technically possible doesn’t mean it will happen.

Frey & Osborne (2017): Automation Risk

Headline finding:

“47% of U.S. employment at high risk of automation.”

Why the 47% is misleading:

• When researchers fixed the method, the scary number collapsed. The OECD redid the study task-by-task instead of job-by-job: only 9% of jobs were at high risk — five times smaller than 47% (Arntz, Gregory & Zierahn, 2016).

What Does AI Do to the Three Forces?

AI extends automation into non-routine cognitive work — tasks once thought safe. The same A–R three-force framework still applies:

• Displacement (−): ChatGPT writes first drafts, summaries, and customer-service replies → fewer hours for paralegals, copywriters, and junior analysts.
• Productivity (+): Doctors, lawyers, and engineers using AI tools deliver more work per hour → demand for their services can grow.
• Reinstatement (+): New roles emerge — prompt engineers, AI safety researchers, model evaluators.

Platform Labor and the Gig Economy

• Independent contractor status avoids labor protections
• Algorithmic management replaces human supervisors
• Task-level labor markets atomize work
• Surveillance and control without employment rights

Key Takeaways

1. Jobs are bundles of tasks — automation hits tasks, not whole jobs.
2. Three forces, not one — displacement (−), productivity (+), and reinstatement (+); the net effect depends on which dominates.
3. Polarization is real — middle-skill jobs hollow out while both the top and bottom grow.
4. Same tech, different outcomes — institutions, not technology alone, decide who wins and who loses.
5. Headline numbers mislead — “47% of jobs at risk” collapses to 9% under task-level analysis.
6. AI is the new frontier — same framework, but faster speed and cognitive target. Policy is still a choice.