Editorial Team
7 min read
Apollo’s analysis estimates 6.7% slower real-wage growth after 2023 in 11 occupations with high Claude-use scores. But its result falls to 1.89% under a broader cutoff, disappears under a stricter one, and cannot by itself show that AI caused the gap.
Apollo Global Management’s new paper offers an early warning about how generative AI could affect pay before payrolls. Its authors estimate that real-wage growth was 6.7% lower after 2023 in a small set of occupations with the highest scores in Anthropic’s Claude-use index.
That is a consequential estimate, but a conditional one. Apollo’s own alternative tests shrink it sharply, its exposure measure comes from one AI provider’s interaction logs, and the design cannot establish that AI caused the relative wage gap. The paper is better read as evidence of a pattern worth testing than as a measure of a nationwide pay cut.
Cover of Apollo Global Management’s July 2026 white paper, “The Impact of AI on the U.S. Labor Market.” Source: Apollo Global Management.
Torsten Slok, Apollo’s partner and chief economist, and Apollo analyst Sania Edlich wrote the analysis. It uses annual Bureau of Labor Statistics Occupational Employment and Wage Statistics data from 2015 through 2025 and treats 2023—the first full year ChatGPT was publicly available—as the break between its pre- and post-periods.
The authors match 321 occupations to Anthropic’s Economic Index. That index scores an occupation by the share of its tasks, weighted by task importance, observed in Claude interactions. An occupation qualifies as highly exposed at a score of at least 0.5. Eleven occupations clear that line, including computer programmers, customer-service representatives, data-entry keyers and financial and investment analysts; the other 310 matched occupations form the comparison group.
In a difference-in-differences regression with occupation and year fixed effects, the high-exposure group has a -0.0667 coefficient for log real wages after 2023. Apollo describes that as wage growth about 6.7% slower than in lower-exposure occupations. The employment coefficient is also negative, -0.0639, but is not statistically significant. It is therefore not evidence that employment stayed unchanged or that AI preserved jobs.
Apollo estimates that the 11 occupations employ 5.8 million people, or 3.7% of its roughly 145 million CPS-weighted workforce measure. Its $28 billion annual labor-income figure is a modelled aggregate based on the estimated wage effect, CPS employment counts and mean wages—not a payroll total observed to have disappeared.
Apollo’s company-reported Table 6 log(real wage) coefficients under three high-exposure thresholds. Source: Apollo Global Management white paper.
The 0.5 threshold does a great deal of work in the headline finding. When the authors rerun the wage model with a broader high-exposure cutoff of 0.4, the estimated effect is -1.89% and statistically significant. At a stricter 0.6 cutoff, it is -1.74% and not statistically significant. Fewer occupations above the stricter cutoff may leave that estimate less precise, but the sensitivity means 6.7% should not be treated as a single stable effect for all AI-exposed work.
The underlying occupation table also shows why a group-level result is not a verdict on each job. Between 2022 and 2024, customer-service representatives, data-entry keyers and medical-records specialists—all above the 0.5 threshold—each recorded positive changes in real wages, while computer programmers, medical transcriptionists and statistical assistants recorded declines. Those descriptive two-year changes are not the regression estimate, but they underscore the variation that the average conceals. A contemporaneous account similarly notes large wage gains in some moderately exposed occupations and large declines in some less-exposed ones.
The paper also finds a -10.7% relative wage effect for its lowest pre-period wage quartile and no statistically significant effect for the highest. Its -24.3% service-occupation estimate is based on 239 occupation-year observations; the authors explicitly warn that it may be sensitive to the exposed occupations included. Those subgroup figures are useful clues about distribution, not separate proof that AI has reduced the pay of every lower-wage or service worker.
Apollo’s company-reported appendix table lists occupation-level Anthropic scores and reported 2022–24 employment and real-wage changes. Source: Apollo Global Management white paper.
Apollo’s main advance over theoretical exposure studies is that the index is based on actual Claude interactions rather than a model’s prediction of tasks AI could perform. Yet it remains a provider-specific proxy for occupational use. The paper says Anthropic was then the only provider to release such usage data for public research, and only 321 of roughly 800 BLS occupations could be matched.
Its literature review gives a further reason for restraint: a cited 2026 survey-based study found that chat-log measures can overclassify generic activities such as editing and programming, whereas surveys capture the purpose workers associate with their jobs. That does not make the Claude measure unusable. It does mean the paper has not directly observed which employers deployed AI, which workers used it at work, or whether their pay-setters changed wages because of it.
The regression controls for differences fixed within an occupation and for annual economy-wide shifts. It includes no time-varying occupation controls. That leaves room for forces that moved differently across occupations after 2023. One reported complication is demand for data-center construction, which can raise pay in comparatively low chat-AI-exposure trades such as electrical and construction work and widen a relative gap without a pay cut in the high-exposure group.
Apollo itself frames the work as an early attempt to measure labor-market effects from observed usage. Its accompanying summary says high-exposure employment levels were unchanged, but the underlying regression does not support a conclusive unchanged-employment finding. The document also says its views may not represent Apollo and that its information is not investment research or a recommendation.
The best retained comparison uses a different country and a different design. Anders Humlum and Emilie Vestergaard’s NBER working paper links adoption surveys to Danish administrative records. It finds precise null effects on earnings and recorded hours at worker and workplace levels, ruling out effects larger than 2% two years after ChatGPT’s launch, even as employers adopted chatbot initiatives and workers reported productivity benefits.
That study cannot disprove Apollo’s U.S. occupation-level estimate: its workforce, observation period, adoption measure and outcome design differ. But it weakens any claim that early generative-AI use already has a uniform, detectable pay effect. It also offers a more direct way to test the mechanism—linking actual adoption to workers and workplaces rather than assigning exposure at the occupation level.
The next test is replication with measures that span multiple providers and distinguish AI use at a workplace from an occupation’s average task score. Researchers would need to follow comparable employers or workers that adopt at different times, report wages and hours over a longer period, and test whether the result remains when high-demand infrastructure trades and other occupation-specific shifts are accounted for.
Until then, Apollo’s result identifies a question, not a settled causal answer: are firms using AI to redistribute productivity gains away from workers in certain occupations, or are an unusually small treated group and a turbulent labor market producing a temporary relative gap? The evidence needed is connected adoption, employer and worker data—not a broader reading of one provider’s log-based index.
Palantir’s co-founder and CEO says he could back a Democrat in 2028, while excluding a Democratic Socialists of America member. His Democratic history complicates talk of a political conversion; the more immediate question is how his government software company will answer congressional scrutiny.
ByteDance’s video model is reaching creators through multiple platforms with 30-second 4K output. Posted prices and draft workflows show why the choice of distributor still matters.
Rox’s $1.3 billion valuation backs AI agents for large enterprise sales teams. Its growth and cost claims face a crowded market, while integration with existing systems leaves the case for replacing Salesforce unresolved.
Google’s Gemini Decision 1 scores predefined choices across text, images, audio, video and PDFs for $0.10 per million input tokens. Its shared-input billing targets recurring workflow costs, but cheaper decision models, rival caching and low-cost generative alternatives complicate the savings claim.
Treasury and other agencies reportedly want Singapore cloud provider Megaspeed on a U.S. trade blacklist. Commerce denies a reported cabinet confrontation over the delay, while Nvidia says it has blocked shipments since June 2025.
United is restricting unauthorized AI agents from searching, booking and managing travel through its platforms and intermediaries. The airline cites security and safety risks as it reserves control over future integrations.
A controlled study found that pharmaceutical ads shifted AI recommendations toward promoted drugs without reducing overall clinical accuracy. The results expose a gap in accuracy tests, but do not establish effects on prescriptions or patient outcomes.
Former AWS computing chief Dave Brown will succeed Santosh Janardhan in early 2027 as Meta weighs its next capacity budget after a quarter of negative free cash flow.
Google is rolling out its Gemini app on Apple CarPlay, bringing voice conversations linked to Gmail and Google Calendar. It joins existing in-car AI options, while some personalized features remain restricted by age and region.
Google Labs’ Tabletop lets groups generate and use shared itineraries, polls and other interactive tools. The US adult-only experiment builds on an idea already offered by shareable AI app builders, leaving its practical advantage to be tested.
Anthropic’s IPO has reportedly attracted orders multiple times the available shares. Its planned $3 billion–$3.5 billion Japanese allocation would exceed Japan’s reported domestic IPO proceeds so far this year, but indications of interest are not completed sales.
Google warned that charging for news links would force it to stop distributing Indonesian news, while Meta and TikTok challenged other parts of the copyright bill. Existing Indonesian payment mechanisms and overseas deals complicate the platforms’ warnings about the cost to publishers.
Utrecht is investigating restrictions on camera glasses in municipal buildings, with findings expected in early 2027. The city must establish whether a ban is lawful, enforceable and compatible with assistive uses.
Research commissioned by cloud-storage supplier Wasabi estimates £3.8 billion in annual UK infrastructure overspend attributable to hard-to-forecast provider charges, including £1.2 billion for AI. Businesses report hiring and investment trade-offs, but recovering that money and turning it into jobs remain conditional.
The USDT issuer’s 800-million-parameter BotPsy beats a larger Qwen model on five of six company-reported agent benchmarks. Its local-AI ambitions still face missing handset measurements and explicit limits on sensitive uses.
Google DeepMind’s AlphaGenome 2 improves cell-specific DNA predictions and accepts new single-cell data without fine-tuning. Company benchmarks show gains at high confidence, but limited recall, complementary tools and greater computational demands leave experimental validation essential.
Coinbase is rolling out server-enforced trading limits and portfolio-loss triggers for AI agents connected to customer accounts. Product-category permissions remain forthcoming, and pausing new trades does not establish a ceiling on losses.
JetBrains IDEs 2026.3 bundle Air for running coding agents and reviewing their changes, while the standalone Air app is in maintenance mode. Existing provider subscriptions remain usable, but billing, security controls and execution options vary by integration.
Amazon’s shares and convertible notes could be worth $271.8 billion at the top of Anthropic’s proposed IPO range. Existing accounting gains, a 12% ownership cap and Anthropic’s computing commitments shape what that paper value means.
Nvidia chief Jensen Huang argued in May 2025 that US chip restrictions were strengthening Chinese rivals. Filings show a substantial commercial stake, but lost sales and market share alone cannot establish whether the controls failed their security purpose.