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.
GPT-Live lets ChatGPT listen and speak at once, and OpenAI says it can bring in GPT-5.5 for harder work without stopping the conversation. The launch’s real test is whether that experience holds up across long sessions, regional capacity and uneven networks—and whether its related WebRTC proposals become broadly deployed infrastructure rather than a company-specific advantage.
Five states enacted AI mental-health restrictions in 2026, but the laws regulate different conduct: Tennessee targets representations that a system is a qualified professional, while Colorado and Maine restrict how licensed clinicians can use AI in therapy.
Xero's survey of 1,081 New Zealand SME customers finds 61% are proactively using AI, but most users are learning by experimentation. Its new 12-week bootcamp with ASB is a response to that confidence gap; it is not yet evidence of a productivity gain.
Apple alleges that OpenAI and former Apple employees sought confidential hardware information through recruiting, supplier contacts and a former engineer’s post-employment system access. The complaint has not been tested in court, and OpenAI says it is unaware of evidence that it has merit.
Amazon’s market value passed $3 trillion after AWS revenue accelerated and the company raised its 2026 capital-spending plan to $220 billion. The rally reflects confidence in Amazon’s cloud demand, but the company’s own capacity forecast, the broad uses of its budget and strong rival cloud growth leave the return on that spending unproven.
The Trump administration says it has completed a voluntary framework for testing the cyber capabilities of the most advanced U.S. AI models and will discuss it with major developers Tuesday. The key scope, testing and disclosure rules have not been made public.
MediaTek now says its second custom cloud AI accelerator is on track for high-volume production in 2028, later than its April goal of entering production by the end of 2027. The company is betting that supply-chain coordination and advanced packaging can support a much larger cloud-chip business, but it has not disclosed a customer, package supplier, cost comparison or 2028 revenue guide.
A Nvidia stockholder’s derivative complaint seeks to make directors and senior officers answer for alleged copyright and biometric-data violations in the company’s AI work. The immediate question is not whether those underlying allegations are true, but whether the suit can clear the procedural and factual hurdles for a board-level claim.
Dongfang Suanxin says its planned 14nm DF2000 will use stacked near-memory computing to reach 15 TB/s of bandwidth. The company’s system-level comparison with NVIDIA highlights an architectural bet, but it does not establish a performance win or answer the questions of yield, cost and software adoption.
Mexico’s exports of computer equipment have surged as U.S. data-center construction and tariff differences reshape trade. The figures point to a bigger role in assembling and shipping equipment, but imported components, near-full factories and limited investment leave the higher-value parts of the chain elsewhere.
Bayreuth Festival says its anniversary Ring uses AI to recombine 150 years of Wagner imagery into a different visual experience at each performance. Early reviews and boos after Götterdämmerung do not settle the experiment, but they focus attention on the people choosing and deploying the system when its projections compete with singers and drama.
A Block technical analysis says a firmware path used to create some COLDCARD secrets could be reproducible; reported address sweeps make the risk urgent, but the analysis does not establish that every affected wallet can be recovered remotely.
Honolulu’s state-run airport has added 17 AI-generated songs to its hourly sound system. The available reporting does not show that they replaced the licensed music already played there, but it has prompted local curator Roger Bong to argue that a public gateway should be more transparent about who decides what Hawaiʻi sounds like.
Trump Media's Truth API offers licensed, machine-readable delivery of selected Truth Social posts to institutional customers. Its commercial appeal rests on the speed with which traders receive communications from President Donald Trump and other officials, but the available accounts leave unresolved exactly what advantage buyers receive.
OpenAI and Anthropic both support disclosure, incident reporting and outside scrutiny for the largest frontier-AI developers. Their proposals—and Illinois’ new law—show that the more consequential unresolved question is whether an evaluator or government agency can decide when a model’s risk is too high for deployment.
AMD’s Kria AI system-on-module and robotics developer platform combine an X100 processor, FPGA-equipped carrier board and open software stack. But the headline 3.4x real-time result comes from an AMD-commissioned simulation run on a Strix Halo mini PC configured as an X100 proxy—not on the forthcoming Kria hardware—and its public descriptions contain methodological differences that make independent reproduction the next test.
A Manhattan federal judge allowed Reddit’s core DMCA and conspiracy claims against Perplexity and SerpApi to proceed over alleged scraping through Google results. The ruling accepts a plausible theory tied to Reddit’s Google license, but leaves unresolved whether Reddit can prove authorization, protected works and actual circumvention.
MediaTek says its AI ASIC business could contribute about $2 billion in the fourth quarter of 2026. That is a company target for one unnamed US hyperscaler project, distinct from its longer-term market-share goal and a separate report on revenue mix.
OpenAI says it closed a $122 billion funding round at an $852 billion post-money valuation. Amazon’s filing sets out three linked elements: $15 billion already invested in OpenAI, a $35 billion share-purchase commitment, and an AWS commercial commitment expanded by $100 billion over eight years.
Two House committee chairs have asked DoorDash to identify the Chinese AI models it uses and the security testing behind them. DoorDash’s public benchmark shows Kimi K2.6 in one experimental code-review configuration, while its stated production reviewer used Claude models—leaving deployment scope and data controls unresolved.