The productivity numbers for 2025 and 2026 are genuinely impressive. US nonfarm business productivity growth, which averaged roughly 1.5% annually through the 2010s and crept up to 1.8-2.1% annually in the early AI adoption period, jumped to 3.4% in 2025 and is tracking at 4.1% in 2026. Some economists are describing this as the beginning of the AI productivity boom that the technology’s proponents have been promising for five years. They’re right that the productivity jump is real. The context they’re less enthusiastic about discussing is why workers don’t feel it.
The relationship between aggregate productivity growth and individual worker wages is not mechanical or automatic. It depends on the bargaining structure of the labor market — specifically, on workers’ ability to claim a share of the productivity gains their labor helped generate. When productivity rose 3-4% annually during the postwar period, wages rose at roughly the same rate, because organized labor and tight labor market conditions gave workers leverage. When productivity started diverging from wages in the 1970s and 1980s, the divergence happened because the conditions that had kept productivity and wages linked — union density, low unemployment, limited capital mobility — eroded.
AI productivity gains are arriving in a labor market where those conditions are largely absent or actively weakening. Union density in the US private sector is at a historic low (6.1% as of 2025). The sectors experiencing the biggest AI productivity gains — professional services, financial services, knowledge work broadly — are among the least organized. And uniquely with AI, the productivity gains are generated by a capital asset (the AI system, the data, the infrastructure) rather than by an increase in human effort or skill that would create a natural justification for wage sharing.
This last point is underappreciated. In previous productivity cycles, workers could argue (and often successfully) that productivity gains reflected their improved skills, harder work, or better organization — and those contributions deserved compensation. With AI-driven productivity, the gain is explicitly attributable to a technology investment made by the employer. The worker’s contribution is being the human who runs the process alongside the AI. That’s a weaker bargaining position, and the wages reflect it.
The pattern in 2026 wage data confirms this directly. Sectors showing the highest AI-driven productivity gains — legal services, financial analysis, content production, software development — show the widest divergence between productivity growth and median worker wage growth. Legal services productivity (measured by outputs per attorney-hour) is up roughly 28% since 2023, driven by AI-assisted document review, research, and drafting. Associate attorney wages at mid-size firms are up 4% over the same period, which is below inflation. Partner compensation at those same firms is up substantially more. The productivity gains are flowing upward within the organizational hierarchy, not outward to the median worker.
This is not nefarious in the individual-firm sense. Each firm is responding rationally to its local labor market conditions. If associates at competing firms are also experiencing flat real wages, no individual firm has an incentive to raise wages above that equilibrium. The outcome is collectively suboptimal for workers and individually rational for each firm — a classic coordination problem that market mechanisms don’t solve.
The counterexample exists and is worth understanding. Trades and physical services workers — electricians, plumbers, construction managers, home health aides, and similar occupations with genuine supply constraints — are seeing wage gains in 2025-2026 that meaningfully exceed productivity growth. This is because their labor markets have the structural conditions that legal services and financial analysis lack: tight supply, no obvious AI substitution in the relevant task set, and in some cases meaningful union or guild organization (licensed trades in particular benefit from credentialing requirements that function as supply constraints).
The lesson is not that wages track productivity. It’s that wages track bargaining power, and productivity growth changes bargaining power in complicated ways that depend on who captures the productivity gains. When productivity growth reflects workers becoming more skilled or scarce, it typically increases worker bargaining power. When productivity growth reflects employers acquiring technology that reduces their dependence on workers, it typically decreases worker bargaining power, regardless of what happens to aggregate output.
There’s an additional dynamic at work in 2026 that compounds the basic productivity-wages divergence: the experience of abundance in the work product combined with scarcity in the paycheck. A knowledge worker using AI tools in 2026 can genuinely produce more — more documents, more analyses, more communications, more code — than she could without those tools. The quality of her output, measured against her pre-AI baseline, has improved substantially. She is, in a real sense, more capable. But if the value she produces is priced in a market where AI tooling means everyone in her field can produce more, the increased supply of output drives down prices, and her income doesn’t rise with her capability. She works alongside increasingly powerful tools, produces more, and earns roughly the same in real terms.
This is psychologically confusing in a way that simple wage stagnation isn’t. Workers who stagnate because their industry is declining can understand their situation, even if they resent it. Workers who stagnate while observably becoming more productive experience a particular kind of cognitive dissonance — they can see that they’re doing more, producing better work, and the results are not reflected in their compensation. The surveys of knowledge worker satisfaction in 2025-2026 consistently show high satisfaction with AI tools and low satisfaction with compensation, which is an unusual combination and an unstable one.
Historical parallels are imperfect here, but the industrial revolution’s “Engels’ pause” is the most relevant precedent. Between roughly 1780 and 1840, British industrial productivity rose substantially while real wages for the working class stagnated or declined. The gains went to capital owners. Workers eventually captured a share of the gains, through a combination of organizing, political pressure, and the tight labor market conditions created by industrial expansion itself. The pause lasted 60 years. It is not a comfortable precedent.
What would it take for the current productivity-wages divergence to close? Three things, historically. First, tighter labor markets than exist currently, which would require either slowed AI deployment (reducing supply of AI-assisted labor) or expanded demand for human workers in non-automatable sectors large enough to absorb the workers displaced from automatable ones. Second, increased worker bargaining power through organizing, credentialing, or legal structures — none of which is building quickly. Third, policy intervention that changes the distribution of AI productivity gains, either through taxation (taxing AI productivity and redistributing it to workers), regulation (requiring that productivity gains be shared with affected workers), or direct government investment in public employment.
None of these mechanisms is currently in place at the scale required. The productivity boom is real. The question of who it belongs to has not been answered by the market, and there are no serious institutional mechanisms for answering it another way.
The workers who feel poorer while producing more are not misunderstanding their situation. They’re reading the data correctly.
One email a week with practical Mac workflows, useful comparisons and honest tech recommendations.

