Study: AI Hasn't Taken Jobs—But It's Stealing Wages

According to Fortune China, people are already exhausted by AI: tired of the 'AI will take our jobs' narrative, and tired of the fact that even the smartest economists still can't explain the current reality. Just at this moment, a new theory has emerged that elegantly connects all the dots: AI isn't eliminating jobs—it's stealing wages. This is the core reason why employees are rising up.

A new study by Apollo Global Management shows that the earliest measurable damage from AI technology is not unemployment, but wage compression. This finding comes amid the most heated debates in the economics community. Even the developers of AI systems cannot agree on what their own data is telling them.

An Economist Changes His Mind

For most of 2026, Apollo's chief economist, Slock, argued that the macroeconomic impact of AI on the labor market was nearly impossible to quantify. In April this year, he wrote: 'AI is everywhere, except in the upcoming macroeconomic data.' In other words, there is no visible evidence of AI's impact in employment, productivity, or inflation data.

Known for his 'Daily Spark' blog and the 'Daily Chart' series previously launched at Deutsche Bank, this analyst has long predicted an 'industrial renaissance' and forecasted that AI would spark a wave of small business startups. On May 29, he published an article on 'Daily Spark' titled 'There Is No Evidence That AI Is Causing Unemployment,' asserting that AI creates more jobs than it destroys. He also cited the 'Jevons Paradox,' which he had been promoting since April, popularizing the idea that 'increasing resource efficiency does not reduce labor demand—in fact, it increases overall demand for that resource.' Shortly after, Anthropic CEO Amodei began using the term and retracted his earlier prediction that AI would cause mass unemployment.

In mid-July, Slock publicly stated that economists' interpretations of the relationship between AI and employment were vague. He pointed out that 'experts simply cannot reach a consensus' on AI's specific impacts at the corporate and employment levels. On July 30, Slock co-published a paper with Sanya Edelich, attempting to unify the conflicting theories. Unlike the theoretical 'exposure' scores widely used in AI labor research over the past few years, the research team used actual usage data from the Anthropic Economic Index—real interaction logs with Claude—to measure how employees actually use AI, rather than what it could theoretically do. The problem they uncovered was not job loss, but 'wage compression.'

Slock wrote: 'Analysis of actual Claude usage data shows that in occupations highly exposed to AI, employee wage growth is slowing, while employment levels remain stable. This suggests that companies are capturing AI-driven productivity gains by suppressing wages rather than through layoffs.' This conclusion also explains the widespread resistance, even open opposition, to AI adoption across industries. Employees seem to realize that these machines will reduce their income.

Employees All Feel the Pressure

In June 2026, an independent survey by Software Finder of 1,005 employed Americans captured this real-world anxiety that falls outside academic models. Half of the respondents said they actively resist new AI tools.

Some of the survey's findings contradict Slock's paper to some extent: Apollo's data suggests that wages in AI-impacted occupations are compressed regardless of whether employees adopt AI, while Software Finder's survey shows that employees currently using AI earn more than those resisting it. This gap likely stems from the characteristics of AI adopters (e.g., managers, higher earners, and those with greater job stability are more willing to use AI), but it does not prove that using AI itself protects wage levels.

For example, Software Finder's data shows that employees resisting AI have an average annual salary of $65,645, about 20% lower than the $81,526 earned by active AI users. 45% of respondents resisted AI due to fears of being replaced; only 16% believed their company adopted AI for genuine business value, rather than following trends or competitive pressure.

Both phenomena can coexist: according to Slock's research, even as wages in AI-affected roles gradually decline, employees resisting AI may still face wage 'discounts.' AI might just be a 'wage-eating machine.'

Moreover, a phenomenon known as 'AI shame' is widespread in the workplace: 13% of employees admitted to pretending to use AI—appearing to leverage tools while manually completing tasks; only 6% believe managers can accurately track how frequently employees use company-deployed AI tools.

An exclusive Fortune report indicates this resistance may go beyond passive avoidance and escalate into deliberate sabotage. In April 2026, Writer and Workplace Intelligence jointly surveyed 2,400 knowledge workers (including 1,200 corporate executives) in the U.S., U.K., and Europe.

The survey found that 29% of employees admitted to actively sabotaging their company's AI strategy, rising to 44% among Gen Z employees. Sabotage behaviors vary: inputting company proprietary information into unauthorized public AI tools; violating rules by using 'shadow AI' systems; outright refusing to use designated AI tools; some even tampering with performance evaluation data or deliberately lowering work quality to prove AI's ineffectiveness. Among employees who admitted sabotage, 30% said their primary motivation was fear that AI would replace their jobs—aligning perfectly with the core concern of AI-resistant workers in the Software Finder survey.

What the Data Reveals

Apollo's paper uses a 'difference-in-differences' model, matching 321 occupations with U.S. Bureau of Labor Statistics data from 2015 to 2025.

The study found that after 2023, employees in high AI-exposure occupations experienced a 6.7 percentage point slower real wage growth rate compared to those in low-exposure occupations, but no statistically significant impact on employment data.

This is the core argument of the study: AI-driven productivity gains are real, but these benefits flow to corporations, not employees. This conclusion also confirms Fortune's March report: AI reduces the time required for individual tasks, meaning companies can assign more tasks to employees.

The impact is concentrated mainly among middle- and low-income groups:

- Lowest income quartile: 10.7% lower wages compared to low-exposure occupations - Second quartile: 5.4% lower wages - Third quartile: 4.0% lower wages - Highest income quartile: no statistically significant impact. High-income individuals appear better able to absorb AI's shocks or benefit from AI adoption - Service industry occupations: 24.3% lower wages. However, authors caution this result is based on a small sample and should be interpreted carefully - Management and professional roles: 4.1% lower wages - Blue-collar roles: no significant impact

Currently, about 5.8 million employees in the U.S. (approximately 3.7% of the total workforce) work in high AI-exposure occupations and continuously feel this wage pressure. This implies an estimated annual labor income loss of at least $28 billion in the U.S. The paper's authors expect this number to keep rising.

Anthropic Economist Offers a Different View

Complicating matters further, the underlying data relied on by Slock's paper comes directly from Anthropic. Yet, Peter McAllory, the company's head of economics, posted a long thread on X in late July offering a completely different interpretation.

Based on 18 months of internal research, he concluded that the U.S. labor market 'has not yet experienced a clear impact from AI.'

McAllory pointed out that the current 4.2% unemployment rate is at the Federal Reserve's definition of full employment, job openings roughly match the number of unemployed, and employment rates for prime working-age adults (25–54 years) are near multi-decade highs.