The conventional wisdom on AI and work is this: automation is a job-killer, and the real question is how many positions will vanish before society adapts. But a new white paper from Apollo Global Management suggests the conventional wisdom has the wrong target entirely — because while employment levels remain largely unchanged, real wages for the most AI-exposed workers have already fallen by an average of 6.7% since 2023.
The Thesis
I believe the AI wage compression story is more dangerous than the AI job elimination story — and that we are collectively under-prepared for it. Here is why: job losses are visible, countable, and politically galvanizing. Wage suppression is slow, diffuse, and almost impossible to attribute to a single cause. When a factory closes, workers march. When a software engineer’s salary growth quietly flatlines for three years running, it barely registers as a news story. That asymmetry in visibility is exactly what makes the wage-compression dynamic so insidious.
Apollo’s paper, authored by analyst Sania Edlich and chief economist Torsten Sløk, tracked wage and employment data across 321 occupations in the United States. Its core finding: AI’s effects on overall employment were statistically “not detectable,” yet roles with high AI exposure saw that 6.7% average drop in real wage growth after 2023 — roughly the moment ChatGPT entered the mainstream public consciousness. The two variables decoupled. Jobs survived. Pay did not.
Supporting Argument: The Numbers Reveal a Class Divide, Not Just an Industry Story
If the wage compression were spread evenly across income levels, it might be manageable — a modest adjustment as the economy recalibrates around new tools. It is not spread evenly. Apollo’s data shows service workers suffered an average 24.3% decline in earnings growth since 2023, while workers in the bottom quartile of earners saw wages decline by 10.7% over the same period. Among the highest-paid workers, no “significant effect” was observed.
Read that again: the workers with the least financial cushion absorbed the sharpest hit. This is not a random distribution of economic disruption. It is a regressive shock — one that widens inequality at precisely the moment when AI’s productivity gains are inflating the valuations of the companies deploying it. AI spending is already warping GDP measurements in ways that make the broader economy look healthier than working households feel, and the wage data from Apollo helps explain why that gap between macro statistics and lived experience keeps growing.
The specific occupations with the steepest real-wage declines are instructive. Computer programmers saw a 6.1% drop in real wages between 2022 and 2024, despite — or perhaps because of — having a 0.75 AI exposure score on Anthropic’s Economic Index, the highest in Apollo’s comparison table. Statistical assistants fell 5.4%. Software quality assurance analysts and testers dropped 2.9%. These are not low-skill roles. They are the white-collar knowledge-work positions that the previous generation of automation left largely untouched. AI has reached them now, and it is showing up in compensation before it shows up in headcount.
Supporting Argument: The Mechanism Is Leverage, Not Replacement
To understand why wages compress before jobs disappear, it helps to think about what AI actually does to a workforce. It does not, in most cases, replace a human wholesale. It makes each remaining human more productive — or, from an employer’s perspective, it reduces the number of humans needed to achieve the same output. That productivity gain flows primarily to capital, not to labour, because the supply of workers willing to do AI-augmented versions of these jobs has not shrunk. If anything, it has grown.
There is a striking parallel here between the Apollo findings and a data point from a Goldman Sachs analysis cited in the same research thread: workers displaced from “technology-disrupted occupations” historically took an average real pay cut of around 3% when finding new employment, and saw real earnings grow an average of 10 percentage points less over the following decade. Combining these two observations — Apollo’s current-period wage suppression and Goldman’s long-run earnings penalty for displacement — suggests that the workers absorbing AI-era wage compression today are not just losing ground now; they are likely setting a lower baseline from which future earnings will compound. The short-term dip and the long-run trajectory are the same wound at different stages.
University of Pennsylvania economist Ioana Marinescu has put a threshold on this mechanism: once approximately 37% of “intelligence tasks” within a role are automated, wages begin to take a measurable hit. Apollo’s data suggests we have already crossed that threshold for several occupations. The question is not whether the threshold exists but how many more roles are approaching it. Apollo estimates 5.8 million U.S. workers currently hold positions that qualify as highly exposed to AI — a figure Edlich and Sløk describe as likely to “grow substantially” as adoption deepens.
This dynamic also reframes the debate around Goldman Sachs’s projection of 15 million displaced workers. Displacement and wage compression are not the same event, but they share a root cause: AI increases employer leverage in compensation negotiations by reducing the cost of each unit of cognitive output. Whether that leverage shows up as layoffs or as salary freezes depends on labour market conditions, union density, and the replaceability of specific skills — but the underlying pressure is identical.
Supporting Argument: Invisibility Is the Feature, Not the Bug
Policymakers and the public have spent enormous energy debating job loss because job loss is easy to measure and easy to dramatise. Entire legislative frameworks — from trade adjustment assistance to universal basic income proposals — are structured around the assumption that the primary harm of automation is unemployment. Sam Altman’s much-discussed proposal for an AI-funded American dividend is predicated on the same framing: that the danger is workers being left out of the workforce entirely.
Wage compression breaks this model. A worker whose real pay has declined 6% over two years is not unemployed. They do not show up in jobless-claims data. They do not qualify for retraining programmes designed for displaced workers. They are, in every administrative sense, fine — even as their purchasing power erodes and their sense of economic security quietly deteriorates. Big Tech’s recent reversal on the AI jobs-wipeout narrative — with executives now emphasizing augmentation over replacement — may be technically accurate while still being economically misleading if wages are the real casualty.
The invisibility of the harm is, in effect, a political subsidy for the status quo. It means the pressure to act remains low even as the damage accumulates.
The Strongest Counterargument
The most serious objection to this thesis comes from labour economists who note that wage data over a two-year window is inherently noisy, and that correlation between AI exposure and wage stagnation does not establish causation. Apollo’s own paper acknowledges “challenges” in its methodology, including shifts in how the Bureau of Labor Statistics classifies occupations over time and changes in data collection practices. Critics could reasonably argue that what looks like AI-driven wage suppression is actually a post-pandemic normalization of wages that had inflated artificially during the labour shortage of 2021–2022. On this reading, computer programmers and software testers are not being squeezed by AI — they are simply returning to a more typical wage trajectory after an anomalous boom.
There is also the counter-evidence within Apollo’s own data: personal finance advisors, despite having more than a third of their tasks exposed to AI, saw wages grow 8.4% over the same period. Administrative law judges saw wages surge 17.5%. If AI exposure were a reliable predictor of wage decline, these outliers would not exist.
I think this objection is worth taking seriously — but it does not ultimately weaken the core argument. The post-pandemic normalization thesis cannot fully explain why the wage declines are so strongly concentrated in the bottom earnings quartile rather than spread evenly across high-exposure roles. If it were purely mean reversion, we would expect the correction to hit hardest where wages had risen fastest — which was, in fact, higher up the income distribution during 2021–2022. The regressive pattern in Apollo’s data is not what a normalization story would predict. As for the outliers: personal finance advisors and administrative law judges both involve high interpersonal trust, regulatory accountability, and contextual judgement that resist straightforward AI augmentation. They are instructive exceptions that prove the rule rather than disprove it.
Why It Still Holds
The wage-compression story does not require AI to be uniquely powerful or uniquely malevolent. It only requires that AI make knowledge workers more substitutable at the margin — and the evidence suggests it is already doing exactly that for a defined class of occupations. The 1,000 AI researchers and executives who recently called for a more measured pace of AI deployment were focused on existential and safety risks. Wage compression is a subtler harm, but it is happening now, not in some speculative future, and it is happening to people who did everything “right” — they got technical degrees, they learned to code, they built careers in high-value knowledge work.
The policy implication is not to slow AI. It is to stop designing labour protections exclusively around the unemployment model and start building instruments that can detect and respond to wage suppression in real time. That means indexing wage-growth data to AI exposure by occupation, expanding the earned income tax credit to cover workers whose wages are stagnating in AI-disrupted fields, and requiring that productivity gains from AI deployment be reported alongside headcount changes in corporate disclosures.
The Prediction
Within the next three years, I expect at least one major economy — most likely the United Kingdom or Germany, given their stronger labour-data infrastructure — to formally classify AI-correlated wage suppression as a distinct policy category, separate from unemployment, and to introduce wage-floor adjustments tied to measured AI exposure in high-risk occupations. The United States will lag, held back by the framing dominance of the jobs-destruction narrative. What would prove me wrong: evidence that the wage declines Apollo identified fully reverse by 2027 as workers upskill into AI-augmented roles, suggesting the suppression was genuinely transitional rather than structural.
