A difference-in-differences study from Apollo Global Management's Torsten Slok and Sania Edlich, built on observed AI usage data rather than theoretical exposure scores, finds real wage growth in highly AI-exposed occupations slowed by 6.7 percentage points after 2023 with no significant employment effect. That reconciles two camps that have spent two years talking past each other.
The statistic that spread was 29%. That is the share of employees who told researchers they have deliberately undermined their employer's AI strategy — rising to 44% among Gen Z workers. It is a good number for a headline and a bad number to reason from on its own, because a survey of self-reported bad behaviour tells you about mood, not mechanism.
The reason serious economics accounts amplified Fortune's July 30 piece was the second half of it: a new working paper from Apollo Global Management arguing that AI is already reshaping the labour market, but not in the way either side of the debate has been claiming. Not through layoffs. Through pay.
The 29% figure comes from an April 2026 survey by the enterprise AI firm Writer and the research group Workplace Intelligence, covering 2,400 knowledge workers across the US, UK and Europe, including 1,200 C-suite executives. It is worth noting that one of the two organisations that produced it sells AI software, which does not make the finding wrong but does mean it should be cited rather than asserted.
More usefully, sabotage in that survey is a broad category. It covers pasting proprietary information into public chatbots, using unapproved tools, flatly refusing to use mandated ones, and — at the sharper end — deliberately producing low-quality output so the AI system looks ineffective. Some of that is resistance. A good deal of it is ordinary shadow IT that gets a more dramatic label when the tool is a language model. Anyone running a rollout should separate the two before concluding their workforce is in revolt.
The same body of survey work contains a finding that complicates the resistance narrative: heavy AI users report being promoted and given raises at roughly three times the rate of slow adopters. That is worth stating, and worth not over-reading. It almost certainly reflects who adopts — ambitious people in visible roles — rather than adoption itself protecting anyone's pay. Selection, not treatment.
The research is by Sania Edlich and Torsten Slok, Apollo's chief economist. Slok is a notable author for this particular finding because he spent most of the past year on the other side of it. In late May he published a note titled, without much ambiguity, that there was zero evidence of AI-related job losses — and in April argued via Jevons paradox that AI would produce more lawyers and accountants, not fewer.
He has not reversed on employment. He has reversed on whether employment was the right thing to be looking at.
The methodological change is the substance of the paper. Most AI-and-jobs research uses exposure scores: an assessment of how much of an occupation's task list a model could theoretically perform. Those scores are guesses about capability, and they have been wrong in both directions for three years. Edlich and Slok instead use observed usage data from the Anthropic Economic Index — what people are actually asking AI systems to do — and match it to occupational wage and employment series from the Bureau of Labor Statistics.
They then run a difference-in-differences model with occupation and year fixed effects across 321 matched occupations, comparing high-exposure and low-exposure work before and after 2023. The finding: real wage growth in the most AI-exposed occupations slowed by roughly 6.7 percentage points relative to less-exposed occupations after 2023, with no statistically significant effect on employment levels. The effect is largest among the lowest earners.
Read carefully: this is a gap in wage growth, not a wage cut. Highly exposed workers are still getting raises. They are getting smaller ones than comparable workers who are not exposed.
Three limitations deserve to sit in the body of the article rather than a footnote, and the authors are open about the first two.
None of that makes the finding uninteresting. It makes it a well-constructed piece of evidence rather than a settled result, which is a category the AI-labour debate has been unusually bad at recognising.
In late July, Anthropic's head of economics posted a long argument on X that the labour market has not yet taken a visible hit from AI. It read as a rebuttal to the Apollo paper, and it was treated as one. It is not really in conflict with it.
Unemployment is a level. Wage growth is a rate. An economy can hold employment flat while the returns to labour in a set of occupations quietly decay, and every aggregate labour-market indicator will look fine throughout. The two claims — no jobs crisis, and workers in exposed occupations are being squeezed — are answers to different questions, and the public debate has been treating them as a single binary for two years.
“The question was never whether AI takes your job. It is whether AI takes your leverage.”
That framing also explains the sabotage finding better than fear of unemployment does. Employees notice when their negotiating position weakens well before it shows up in a layoff notice. A worker whose output is now partly reproducible by a tool their employer pays for has less to threaten with at review time, and knows it. Resistance to a rollout is a rational response to a change in bargaining power, whether or not anyone involved would describe it in those terms.
None of this is a new shape. Mechanisation and industrial robotics both produced episodes where output per worker rose and the worker's share of it did not. Labour's share of national income in advanced economies has been drifting down since roughly 1970, through several waves of technology that were each expected to lift wages broadly and did so unevenly at best.
What would have to be true for this cycle to break the pattern is reasonably specific: the productivity gains would need to be hard to capture at the firm level — diffuse, portable, and available to workers as bargaining leverage rather than to employers as a substitute. Current evidence points the other way. The gains sit inside enterprise licences and internal tooling, which is exactly where they are easiest to keep.
If you run an AI rollout that is not landing, the useful diagnostic is not an engagement dashboard. It is whether your people believe the productivity gain is being shared. The Writer survey's sabotage cases cluster around mandated adoption, which is the deployment style that most clearly signals the gain is being captured rather than shared.
If you are an individual trying to work out what to worry about: on this evidence, the near-term risk in a highly exposed knowledge-work occupation is a smaller raise, not a redundancy. That is a less dramatic threat and a more corrosive one, because it compounds silently and never produces a moment where anyone has to explain it to you.
And it is worth holding the whole thing loosely. This is one working paper, on one company's usage data, covering fewer than half of the occupations in the economy, from an author who held the opposite position ten weeks ago. That last fact is a point in his favour — updating on evidence is the job — but it is also a reminder of how fast the read on this changes.
On this evidence, not yet in measurable numbers. The Apollo paper found no statistically significant employment effect across 321 occupations. What it did find was slower real wage growth — about 6.7 percentage points — in the most AI-exposed occupations after 2023. Flat employment and compressed pay are not contradictory findings; they are answers to different questions.
DrafterDaily reads the papers behind the headlines and tells you where the evidence stops. One briefing, every weekday.
The headline was the 94% drawdown from a $3.44 billion peak. The more useful arithmetic is the cash sitting inside the thing being bought — and the fact that everyone except the angels was paid in another private company's paper.
The biography is the least interesting part. What the board actually did was unbundle the CEO job — keeping Tim Cook on with a defined policymaker-engagement remit — and that separation says more about Apple's next decade than the appointment does.
Every outlet reported the size of Meta's teen-safety settlement and the screen-time limits. Almost nobody read the payment structure — where roughly $5.3bn is released only if two competitors act.