Wage Suppression and Inequality

The argument runs through bargaining power rather than headcount. If a system can perform a growing share of the tasks in a job, the employer's alternative to paying a given wage improves, even when nobody is dismissed. On this account pay growth slows first, then the share of national income going to wages falls, and the returns flow to whoever owns the models, the computing capacity and the data. The distributional question is separate from the employment question, and a labour market with stable employment and falling real wages would satisfy the fear entirely.

The historical basis is strong. Daron Acemoglu and Pascual Restrepo, writing in Econometrica in 2022, estimated that between 50% and 70% of the change in the United States wage structure over the previous four decades is accounted for by relative wage declines among worker groups specialised in routine tasks in industries that automated rapidly.^[1] That is a claim about robots, numerically controlled machinery and enterprise software rather than about generative AI, but it establishes that automation can reshape the wage distribution without producing mass unemployment.

A second strand concerns which workers AI complements. The International Monetary Fund's 2024 staff discussion note found that AI complementarity is positively correlated with income, so the workers best placed to gain are those already earning most, and modelled scenarios in which high earners capture a more than proportional increase in labour income.^[2] A third strand concerns algorithmic management. Veena Dubal, writing in the Columbia Law Review in 2023, described how on-demand platforms use granular behavioural data to set variable, personalised and unpredictable hourly pay, a practice she called algorithmic wage discrimination.^[7]

The clearest measured trend is the split between labour and capital, and it does not depend on AI. United States Bureau of Labor Statistics data put the labour share of nonfarm business output at 53.8% in the third quarter of 2025, the lowest reading in a series that begins in 1947, down from 54.6% the previous quarter against a decade average near 55.6%.^[8][9] The series has been falling since the early 2000s, long before generative models existed, so it describes the direction of the economy rather than proving an AI effect.

Those results do not line up neatly, and the disagreement is substantive rather than a matter of measurement error. Isabella Moder and Til Pommer at the European Central Bank found in 2026 that United States employment reallocated away from occupations at high risk of AI substitution between 2019 and 2025 while wage growth showed no significant relationship with substitution risk, which they read as a sign that adjustment is incomplete rather than absent.^[4] The finding by Erik Brynjolfsson, Danielle Li and Lindsey Raymond that generative assistance helps novices far more than experts points toward compression of within occupation pay gaps, not widening.^[5]

Usage data shows a different kind of concentration. Anthropic's economic index, drawing on a privacy preserving analysis of one million Claude.ai conversations and one million API transcripts across more than 150 countries, reported in November 2025 that Claude usage is geographically concentrated, with high income countries overrepresented in global usage relative to their working age populations.^[11] A June 2026 report in the same series found a positive but noisy relationship between an occupation's median wage and tokens per conversation, with top tercile occupations consuming 2.07 times as many tokens as bottom tercile occupations, and named pharmacists as a counterexample.^[12] On these measures adoption is following the existing income distribution, both between countries and between occupations.

The first pathway is a wedge between productivity and pay. If AI raises output per worker while the threat of substitution weakens wage bargaining, the gains accrue to firms rather than employees. That is what the falling labour share describes, and it does not require anyone to lose a job. Nothing in the technology prevents the pattern from continuing, and the mechanism Acemoglu and Restrepo documented for routine task automation applies to cognitive task automation without modification.^[1][8]

The second is geographic. If adoption stays concentrated in high income countries, as the Anthropic usage data indicates, then the countries that adopt AI fastest capture the productivity gains while countries whose comparative advantage is cheap labour lose it.^[11] The International Labour Organization measured exposure at 34% of employment in high income countries against 11% in low income countries, which cuts both ways: less exposure means less disruption but also less of the gain.^[16]

The third is inside firms. Algorithmic management makes pay setting continuous and individualised, so that two workers doing the same task at the same time can be offered different rates based on inferred willingness to accept. Dubal's study of on-demand drivers documented this in practice and argued that it removes the predictability that makes wage work a basis for economic stability.^[7] Extending the same techniques from platform work into salaried employment is a technical rather than a conceptual step.

The fourth is capital concentration. The IMF working paper by Emma Rockall, Marina Mendes Tavares and Carlo Pizzinelli found that when firms choose how much AI to adopt, the wealth inequality effect becomes more pronounced, because the cost savings from automating high wage tasks drive higher adoption rates.^[3] Models that ignore that choice, they argued, understate the trade off between inequality and efficiency that policymakers face.

Lawrence Mishel and Josh Bivens of the Economic Policy Institute have argued for years that automation is the wrong explanation for American wage stagnation. Their evidence is that measures of automation rose fastest in the late 1990s, the period of the strongest broad based wage growth in a generation, and that occupational employment patterns since 1999 have not polarised in the way the technology story predicts even as inequality rose.^[13] Their alternative account attributes wage suppression to policy choices about labour standards, union density, monetary policy and corporate governance. On that reading, blaming AI misidentifies the cause and misdirects the remedy.

The strongest empirical rebuttal is simply that the wage effect has not appeared. The European Central Bank analysis covering 2019 to 2025 found no significant relationship between an occupation's AI substitution risk and its wage growth.^[4] The payroll study by Brynjolfsson and colleagues, which did find employment effects, reported explicitly that adjustment came through headcount rather than compensation.^[17] Richard Audoly, Miles Guerin and Giorgio Topa at the Federal Reserve Bank of New York found that fewer than 10% of workers sit in occupations with high measured AI exposure at all, which bounds how large any wage effect can currently be.^[19]

There is also a serious argument that AI narrows rather than widens gaps. The field study of customer support by Brynjolfsson, Li and Raymond found the largest productivity gains among the least experienced workers, with almost none for the most skilled, a pattern that compresses output differences within an occupation.^[5] David Autor has built a broader case on that observation, arguing in 2024 that AI's distinctive property is extending the reach of expertise to workers who lack formal credentials, and that this could widen rather than hollow out the middle of the wage distribution.^[14] The IMF's own modelling reaches a similar conclusion for wages while warning about wealth.^[3]

Finally, Daron Acemoglu estimated that AI will add no more than 0.66% to total factor productivity over ten years, and argued the figure should be revised down to less than 0.53% once the limits of task level cost savings are taken into account.^[18] He called those macroeconomic consequences nontrivial but modest. A technology that produces aggregate gains that small cannot produce large distributional shifts through those gains, whatever it does to bargaining power.

The European Union has taken the most direct regulatory route. Annex III of the AI Act classifies as high risk any AI system intended for recruitment or selection, including targeted job advertising, filtering applications and evaluating candidates, and any system used to make decisions about the terms, promotion or termination of a working relationship, to allocate tasks on the basis of individual behaviour or personal traits, or to monitor and evaluate performance.^[20] Article 26 requires employers deploying such a system to inform workers' representatives and the affected workers before putting it into use.^[21]

United States action has been narrower. California's legislature passed Senate Bill 7, titled Employment: automated decision systems and known to its backers as the No Robo Bosses Act, which would have required notice before automated decision systems were used in employment decisions and barred employers from relying on them alone for discipline or dismissal. Governor Gavin Newsom vetoed it on 13 October 2025, objecting to unfocused notification requirements and overly broad restrictions rather than to specific misuse, and the veto was sustained on 2 March 2026.^[22] America's AI Action Plan, published in July 2025, directed the Department of Labor to fund rapid retraining for workers displaced by AI and to track wage effects through a new research unit, but proposed no wage or bargaining measures.^[23]

The IMF's 2024 note argued that whether AI raises or lowers inequality depends on policy choices rather than on the technology, and set out options across social protection, retraining and the taxation of capital income.^[2] Gallup's February 2026 survey of 23,717 United States employees found 18% expecting their own job to be eliminated within five years because of AI or automation, rising to 23% in organisations that had adopted AI.^[15] Policy has so far addressed that anxiety mainly through skills funding rather than through the distribution of gains.

Distinguish three claims that are often merged. The first is that AI lowers wages for exposed workers, which current studies do not support. The second is that AI shifts income from labour to capital, which is consistent with the falling labour share but not yet attributable to AI. The third is that AI concentrates gains geographically and among high earners, for which the adoption data is suggestive. The three rest on different evidence, run on different timescales, and imply different remedies.

Watch the labour share and the productivity to pay relationship rather than headline wage growth, because wage growth in a tight labour market can mask a shift in the split of output. Watch adoption by firm size and country income, since the distributional story depends on who deploys the technology rather than on what it can do. Where a study reports an AI skill wage premium, check whether it was measured in job advertisements, which reflect what employers offer to attract scarce skills, or in payroll data, which reflects what workers actually receive.

The disagreement between Mishel and Bivens on one side and the task based modellers on the other is not about whether inequality has risen. It is about whether technology or policy is the operative cause, and that question has a practical consequence: if policy is the cause, the remedies exist already and do not depend on regulating AI at all.^[13]