Redistributed

Paper: Daron Acemoglu. "Harms of AI." Cambridge Journal of Economics, 2021. Link: Cambridge

Why This Paper Matters

Daron Acemoglu is one of the most influential economists working on technology and labor. This paper provides the theoretical foundation for understanding why AI might harm workers — and why that outcome isn't inevitable.

The core argument: we're building the "wrong kind of AI." Current incentives push toward automation that replaces workers without productivity gains. Policy can redirect AI development toward augmentation instead.

This matters for redistribution because it shows that AI's labor impact isn't determined by technology alone — it's shaped by choices that can be made differently.

The Argument

Not All Automation Is Equal

Acemoglu distinguishes three types of automation:

1. Productive automation: Machines do tasks better than humans, generating genuine productivity gains.

Example: Automated manufacturing that produces more goods with fewer defects
Outcome: Workers may be displaced, but productivity growth can create new jobs and higher wages

2. Labor-complementing automation: AI augments human capabilities rather than replacing them.

Example: Diagnostic AI that helps doctors catch diseases earlier (see AI in Healthcare)
Outcome: Workers become more productive; their labor is more valuable, not less

3. "So-so" automation: Machines replace workers in tasks where they're only marginally better (or no better at all).

Example: Self-checkout kiosks that are slower than cashiers but cheaper
Outcome: Job loss without productivity gains; pure redistribution from labor to capital

The problem: current incentives favor so-so automation. It's profitable for individual firms even when it's bad for workers and society. The Distributional Effects of AI compound this: when so-so automation shifts returns from labor to capital, the gains concentrate among those who already own the most.

Why We Get the Wrong Kind

Tax incentives favor capital over labor: The US tax code treats capital investment favorably (depreciation, credits) while taxing labor (payroll taxes). This tilts the calculation toward replacing workers even when automation isn't more productive.

Shareholder pressure: Public companies face pressure to cut costs. Replacing workers with machines looks good on quarterly earnings, even if long-term productivity doesn't improve.

Research direction: AI research has focused on matching or exceeding human performance on specific tasks — the "beat the human" benchmark. This naturally leads to replacement rather than augmentation.

Market concentration: Dominant firms can afford so-so automation because they're not competing on efficiency. They automate to maintain margins, not to serve customers better. As Korinek and Vipra document in their analysis of AI Market Concentration, the cost structure of foundation models compounds this dynamic.

Path Dependence

A crucial insight: these choices are path-dependent. Once you build the labor-replacing system, switching is hard.

Infrastructure gets built around the automated approach
Skills for the old way of working atrophy
Complementary technologies develop around automation
Political coalitions form to protect existing investments

The direction set now will shape AI development for decades. Early choices matter disproportionately.

Policy Implications

Acemoglu argues for redirecting AI development:

Change tax incentives:

Remove favorable treatment for capital investment
Consider taxing automation directly (a "robot tax")
Reduce payroll taxes to make labor relatively cheaper

Direct R&D funding:

Fund research on labor-complementing AI
Create prizes for augmentation, not just automation
Evaluate publicly funded research on job impact

Strengthen worker power:

Collective Bargaining and AI gives workers voice in technology deployment
Works councils (European model) require consultation on automation
Sectoral bargaining sets standards across industries

Antitrust:

Break up concentrations that enable so-so automation (see AI and Antitrust)
Competitive markets force efficiency, not just cost-cutting

What the Paper Gets Right

Framework clarity: The distinction between productive, complementing, and so-so automation is analytically powerful. It cuts through the "technology is neutral" framing.

Incentive focus: Acemoglu shows how structural factors — taxes, governance, research funding — shape technology development. This identifies levers for policy.

Path dependence: Emphasizing that early choices lock in trajectories is important. It creates urgency without fatalism.

Political economy: The paper connects technology to power — who benefits from current arrangements and who would resist change.

Critiques and Limitations

Measurement challenges: Distinguishing so-so automation from productive automation empirically is hard. How do you measure counterfactual productivity?

Political feasibility: The policy recommendations (robot taxes, breaking up big tech) face enormous political resistance. The paper doesn't address how to build coalitions for change.

Global dynamics: The analysis is mostly US-focused. Automation incentives differ across countries; policy in one country may just shift activity elsewhere.

Labor agency: The paper focuses on policy; it says less about worker organizing and collective action as forces shaping technology.

Alternative views: Erik Brynjolfsson and others argue that automation's benefits are real and the problem is distribution, not the technology itself. Acemoglu doesn't fully engage this position. The debate between them — explored in depth in The Automation Debate — remains one of the most consequential disagreements in AI economics.

The Debate

Acemoglu's argument sits within a broader debate about automation and work:

Brynjolfsson's response: AI augments workers and increases productivity. The problem isn't automation — it's that we haven't updated institutions to share gains. Focus on redistribution (education, transfers, safety nets), not redirecting technology.

Benanav's critique: Both Acemoglu and Brynjolfsson overstate technology's role. The labor share has been declining for reasons unrelated to automation — globalization, financialization, weakened unions. Don't mistake a political problem for a technological one.

The synthesis: These views aren't fully incompatible. All agree:

Current outcomes are bad for workers
Policy matters
Worker power has declined
Status quo is unsustainable

The disagreement is over mechanisms and priorities, not goals.

Implications for Redistribution

The paper's redistribution lens:

Value: Automation without productivity gains is pure redistribution from labor to capital. Workers lose; shareholders gain.

Work: The "so-so" framing explains why jobs can disappear without the promised productivity benefits materializing.

Power: Concentration enables so-so automation. Breaking concentration is labor policy, not just competition policy.

Policy: Tax treatment, R&D direction, and labor standards are redistribution mechanisms. Current settings redistribute upward.

How to Use This Paper

When arguing that technology isn't neutral: Acemoglu provides rigorous economic reasoning for why AI's effects depend on deployment choices, not just technical capability.

When proposing policy: The framework suggests specific interventions — tax reform, R&D direction, antitrust — with clear mechanisms.

When countering fatalism: "Technology determines outcomes" is false. Acemoglu shows how path-dependent but policy-malleable AI development is.

When analyzing AI deployment: Ask: Is this productive automation, labor complementing, or so-so? The framework helps categorize and evaluate.

When looking for models that work differently: Some projects already embody the labor-complementing approach Acemoglu advocates. AI Coding Assistants that augment developers rather than replacing them, cooperative structures explored in AI Cooperatives, and public alternatives outlined in Public AI all suggest that the "wrong kind" of AI is not the only kind being built. The path Acemoglu describes is the default, not the destiny.

Source

Acemoglu, Daron. "Harms of AI." Cambridge Journal of Economics, Volume 45, Issue 4, July 2021, Pages 819–830. Cambridge

General

Paper Review: The Wrong Kind of AI?