Skip to content

All five questions

Will AI widen inequality?

The claim, as coded: “AI is likely to increase inequality unless policy intervenes”

The short answer AI-written

Will AI increase inequality absent policy? This claim has the broadest documented agreement in our network: automation has driven wage inequality and labor-share decline, and most expect AI to continue the pattern by default. New firm- and city-level evidence from China points the same way. The live dispute: could augmenting AI compress skill premiums instead?

A synthesis of the cited research, not a prediction or a representative consensus.

Why they disagree AI-written

Some studies find AI narrows gaps between workers in the same job, while others expect it to shift income from workers to the owners of capital.

Each row is one assumption and the published papers that make it (of 67 on this question). Bar: where their findings point.

Does AI help workers or replace them?

Helps workers4 papers
1 for
Replaces work5 papers
1 for, 4 mixed

Would settle it: Wage data by skill and experience inside firms that adopt AI, over several years.

Who receives the gains?

Owners of capital5 papers
2 for, 3 mixed
Depends on policy4 papers
2 for, 2 mixed

Would settle it: The labor share of income in sectors and countries with high and low AI use, under different policies.

What kind of evidence is the estimate built on?

Measured in workplaces19 papers
8 for, 3 mixed, 2 against
Models of the economy13 papers
4 for, 6 mixed

Would settle it: Studies that link workplace effects to changes in the national income distribution.

The case made for the claim

Acemoglu and Restrepo's work attributes a large share of rising US wage inequality to automation displacing routine task workers, and Acemoglu's recent writing extends the warning to generative AI. Korinek's research makes concentration of income a central default outcome absent policy. Prettner, Hemous, and Salomons document automation raising inequality and eroding the labor share in models and cross-country data. Brynjolfsson, Bengio, Del Rio-Chanona, Deming, and Koning all document uneven gains — across skill, gender, and geography. Outside our network, recent studies find AI adoption widening executive-worker pay gaps inside Chinese manufacturing firms, and wage stagnation in Beijing neighbourhoods most exposed to generative AI since 2023.

Still open

  • Do novice-biased experimental gains persist and generalize into actual wage compression at scale?
  • Does AI shift income mainly between workers (skill premiums) or from labor to capital (labor share) — and which matters more for policy?
  • How much of AI's distributional effect runs through geography — regions and countries — rather than individuals?
  • Do the early within-firm and within-city widening effects found in China appear in other economies with different labour institutions?

Sources

  1. Artificial Intelligence, Automation and Work
  2. The impact of generative artificial intelligence on socioeconomic inequalities and policy making
  3. Artificial Intelligence and Its Implications for Income Distribution and Unemployment
  4. Is Automation Labor-Displacing? Productivity Growth, Employment, and the Labor Share
  5. Building Pro-Worker Artificial Intelligence
  6. The Turing Transformation: Artificial Intelligence, Intelligence Augmentation, and Skill Premiums
  7. Artificial intelligence and within-firm wage gaps: Evidence from China’s manufacturing sector
  8. Generative AI impacts on intra-urban inequality and skill premium in Beijing

The numbers

Who gains more from AI at work

Studies that measured AI’s effect on output separately for less and more experienced (or lower- and higher-performing) workers. When the less experienced gain more, the gap between workers narrows; when the better performers gain more, it widens. 3 studies, each number with its quote; settings differ.

Customer support (Brynjolfsson et al.)+34% less experienced or lower performers
Consultants (Dell'Acqua et al.)+43% less experienced or lower performers; +17% more experienced or higher performers
Kenyan entrepreneurs (Otis et al.)−8% less experienced or lower performers; +15% more experienced or higher performers
−10%50%

Less experienced or lower performersMore experienced or higher performersGainLoss

All 12 numbers from 8 papers, in 3 charts, each with its quote

Who says what

Documented positions only, coded with cited evidence by an AI agent: 27 of 400 mapped researchers so far. Everyone else is unknown: no reviewed evidence, which is neither agreement nor neutrality.

23 agree3 it's complicated1 disagree27 of 400 coded

Agree (23)

It's complicated (3)

Disagree (1)

Economists and computer scientists

Documented positions on this claim of researchers on the co-authorship map, by field (from their OpenAlex topics). Researchers without a documented position are not counted.

FieldAgreeIt’s complicatedDisagree
Economists (241 on the map)21 of 243 of 240 of 24
Computer scientists (51 on the map)Too few to say (2 coded)

Each position with the researcher’s reasoning and cited evidence

What to read

Essentials on inequality

Read

Know

Follow

  • Daron Acemoglu agreesEconomist, Massachusetts Institute of Technology

    Argues AI could widen the gap between capital and labour income even without lower wages.

  • David Autor it's complicatedEconomist, Massachusetts Institute of Technology

    Author of work on the polarization of the US labour market and the growth of low-skill service jobs.

  • Paul KrugmanNewsletter · 5 posts in this guide

    Regular posts on AI, labour and the distribution of income.

All essentials on inequality, for 10 minutes to an afternoon

Where the papers point

66 agree42 it's complicated3 disagree111 of 313 coded

How each published paper’s own finding bears on the claim, read from its title and abstract by an AI model (AI-written). 111 of the 313 published papers address this question; the rest are not counted. This codes findings, not authors.

Findings that support the claim (66)

and 63 more in the research library

Findings that it depends (42)

and 39 more in the research library

Findings that go against the claim (3)

All research on this questionNews on this question