Apollo's 6.7-point AI wage gap rests on 11 of roughly 800 BLS occupations

Counted forward from the first quarter of 1947, the second-quarter reading sits at the bottom of 318 quarters, and the BLS productivity report does not attribute the decline to a cause ^[1][17]. CNBC reports that some researchers put it down to decades of automation that AI may accelerate ^[16]. August's payroll beat came a quarter later, with wage growth lagging the latest inflation readings ^[3].

The evidence carrying the AI attribution is one paper. Torsten Slok, Apollo Global Management's chief economist, and his co-author Sania Edlich found real wage growth 6.7 percentage points slower after 2023 in occupations classified as highly exposed to AI than in less exposed ones, and no statistically significant effect on employment ^[4][5]. Of roughly 800 BLS occupations, 321 were usable and 11 cleared the high-exposure threshold ^[7]. That sets 11 occupations against 310 ^[20], with the exposed side at 3.4% of the usable set and about 1.4% of the full classification ^[18][19]. Apollo called the result "early evidence" and said the work accounted for occupational differences and annual labor-market trends ^[7][9]. The authors wrote that the study mattered as a demonstration that "AI research has entered a new phase, one in which labor market impacts can be measured from observed adoption rather than predicted from theoretical exposure" ^[8].

Ben Zipperer, senior economist at the left-leaning Economic Policy Institute, said the sample was too small to be convincing ^[11]. "It's absolutely the case that AI could be affecting the demand for certain types of jobs," Zipperer said ^[10]. His objection is about where the saved money goes: if AI makes software cheaper to build, the savings get spent, including on hiring in other occupations. "That makes the highly exposed jobs look worse by comparison, even though some of that measured loss is just income increases for other workers," he said ^[12]. He also said: "There was a relative slowdown in labor demand for computer programmers and related jobs in the wake of pandemic rehiring that had nothing to do with AI" ^[13].

The composition of the job gains cuts the same way. Tech and professional services, both high paying, are losing jobs while hospitality and health care lead the gains, and that mix pushes down average pay on its own ^[14]. Pandemic-era wage growth reflected an extremely tight labor market, and current gains sit closer to the recent historical norm ^[15].

I would expect pay to register an AI effect before employment does, because a firm that reassigns a worker keeps the headcount and takes the saving at the next merit cycle, which is the pattern Apollo's authors describe when they say productivity gains may be captured through wage compression instead of workforce reduction ^[6]. Zipperer's counter-case uses the same numbers: tech unwinding its pandemic over-hiring produces slower wage growth and weaker employment in the same occupations the exposure classification flags ^[13]. Eleven occupations cannot separate the two ^[7].

So the differential is the number to test. Widen the high-exposure set past 11 occupations and the 6.7 points either holds or shrinks ^[7][4]. If labor's share climbs back while tech hiring recovers, the series low belongs to the pandemic cycle ^[1]. If the employment effect turns statistically significant, the compression framing is the wrong one ^[5][6].