Is AI Suppressing Wages? Two Studies Reach Very Different Answers

There is no persuasive evidence yet that AI is suppressing wages across the U.S. labor market. An Apollo analysis finds markedly slower real wage growth in a small group of occupations with high observed Claude use, while a broader Dallas Fed analysis finds no average relationship between occupational AI exposure and wage growth.

The findings are less contradictory than the headlines suggest. They cover different occupation samples, define exposure differently and measure different wage outcomes over different periods; neither analysis can isolate a nationwide causal effect of AI on pay.

What Apollo’s 6.7-point estimate covers

The Apollo white paper estimates that real wages grew 6.7 percentage points more slowly after 2023 in occupations above its high-exposure threshold than in the comparison group. Its difference-in-differences model covers 321 occupations from 2015 through 2025, but only 11 occupations meet the baseline threshold; the employment estimate is not statistically significant. The wage coefficient falls to 1.89 percentage points at a lower exposure cutoff and to a statistically insignificant 1.74 points at a higher cutoff.

Those threshold results matter because the headline estimate is driven by a small treated group. Its members include computer programmers, customer service representatives, data-entry keyers, medical-records specialists, financial and investment analysts, database architects and several other occupations. Their wage paths cannot automatically represent all workers who use generative AI.

Apollo measures exposure using the share of an occupation’s tasks observed in interactions with Anthropic’s AI tools, weighted by task importance. This captures activity that capability-based indexes may miss, but it is not a census of employer adoption: it reflects one provider’s logs, excludes occupations without a usable match and does not show whether every interaction was part of paid work.

Why the Dallas Fed finds no average penalty

The Dallas Fed occupation analysis examines annualized wage trends since fall 2022 across 205 occupations and finds an almost horizontal overall relationship with AI exposure. After accounting for an occupation’s experience premium, a one-standard-deviation increase in exposure is associated with a 0.05-percentage-point decline in wage growth at the median premium, but the 90 percent confidence interval includes zero. The estimate is negative 0.28 points for an occupation with no experience premium and positive 0.2 points at the 90th percentile of that premium.

The experience premium is the modeled percentage difference between experienced and entry-level occupational wages. It is not a record of the same worker’s pay before and after AI adoption. The pattern is therefore consistent with AI substituting for codified tasks while complementing tacit knowledge, but it does not prove that AI produced either effect.

The methodology matrix behind the disagreement

• Exposure: Apollo uses observed activity involving Claude and converts it into a binary high-exposure group. Dallas uses the Felten–Raj–Seamans capability-based index as a continuous measure.
• Sample: Apollo’s headline comparison depends on a small treated group within a larger matched panel. Dallas estimates a relationship across a broader occupational distribution.
• Period: Apollo sets the post period at 2023–2025 and compares it with an earlier panel beginning in 2015. Dallas compares post-2022 annualized wage trends with prepandemic trends.
• Wage outcome: Apollo models annual mean occupational wages after adjusting them with CPI-U. Dallas studies occupational wage-growth trends and interacts exposure with modeled returns to experience.
• Question: Apollo asks whether occupations above a particular observed-use threshold changed differently after 2023. Dallas asks whether greater capability-based exposure predicts wage growth across occupations and whether that relationship varies with the experience premium.

These choices can produce different answers without either result being arithmetically inconsistent. A concentrated change may disappear in a broad average, while a large coefficient from a small threshold-defined group may change substantially when the cutoff or group composition changes.

Exposure is not the same as an economic effect

The distinction between the two exposure measures is substantive, not semantic. The ILO review of AI-exposure indicators cautions that such measures differ conceptually, rely on static task descriptions or subjective judgments and often omit economic feasibility and adoption constraints. Exposure can identify where work might change; by itself, it does not predict wages, employment or displacement.

Observed product activity narrows one uncertainty but does not remove it. Claude logs reveal that certain tasks were performed with AI assistance, yet they do not identify the employer’s technology stack, the intensity of use across all workers, changes in output quality or the share of productivity gains passed into pay.

Neither analysis settles causation

Apollo’s fixed effects remove persistent differences between occupations and shocks shared across years, making its design more informative than a simple cross-sectional correlation. A causal interpretation still requires high- and low-exposure occupations to have followed comparable wage trajectories without AI and requires no other post-2023 change to have affected the groups differently. The paper itself acknowledges possible confounding from post-pandemic labor-market dynamics and incomplete occupation matching.

The Dallas result has the opposite aggregation problem. A near-zero average can conceal wage pressure in routine occupations and gains in jobs where experience becomes more valuable. Its interaction estimates show heterogeneity, but the experience premium remains an occupation-level proxy rather than direct evidence about individual workers, AI adoption or bargaining power.

The defensible conclusion is therefore qualified: AI may be restraining wage growth in some highly exposed occupations, but the available evidence does not establish a general U.S. wage-suppression effect. Apollo supplies a narrow warning signal from observed use; the Dallas Fed shows that the signal weakens on average and changes with occupational returns to experience. Broader adoption data and worker-level wage histories will be needed to determine whether either pattern is truly caused by AI.

Also read:

• When AI Search Tools Confuse Two Different People With the Same Name
• YouTube Directors Conquer the Big Screen: Two Horror Hits Show the Platform’s Power
• 409A Valuation vs. Fundraising Valuation: Why Your Startup Has Two Different Prices