The curve broke twice.
No one fully agrees why.
U.S. labor productivity grew about 3% a year from 1947 to 1973. Then it slowed. The 1995–2004 IT decade was a brief revival; after 2004 the curve broke again. This page tells the story through seven public datasets and the seven competing explanations they refute, support, or only partially accommodate.
The mystery isn't that growth slowed once. It's that growth slowed, came back, and slowed again — and almost every advanced economy did the same thing at the same time. That synchronicity is the single fact that shapes which explanations survive contact with the data.
Annual labor productivity growth, U.S. nonfarm business sector, 1948 to 2025. Each bar is one year. The horizontal lines mark period averages: a postwar boom near 3%, the 1973–95 slowdown, the brief IT-decade revival, the post-2004 break, and the recent surge of unclear durability. Hover any year to read it; click a period to focus on it.
What this means
Every hypothesis in the next six charts has to explain four things at once: the sharp 1973 break (3% to 1.5%), the 1995–2004 IT-driven exception, the 2004 break back to 1.3%, and the 2024–2025 surge whose cause we don't yet know. The repeated visit to ~1.5% across decades is what makes the technology-cycle hypothesis (§ V) the hardest to falsify.
Labor productivity decomposes into three components: total factor productivity (the residual — technology, management, intangibles), capital deepening (more capital per worker), and labor composition (workforce skill mix). The chart below shows BLS Table B contributions for each period. The key fact: post-2004, capital deepening barely fell. The slowdown is in TFP — the residual.
What this rules out
Stories that work only through capital — low investment, savings glut, infrastructure neglect — can't be the full story for the post-2004 break, because capital deepening only fell about 0.3 percentage points. Stories that work only through workforce composition can't be it either: the labor-composition contribution sits near 0.3 ppt the whole period. The slowdown is in the residual, and the residual is where technology lives.
GDP per hour worked, annual % growth, by period, for the G7. After rising in parallel before 1995, the U.S. pulled ahead in the IT decade. Since the mid-2000s every major economy slowed. In 2024 only the U.S. showed meaningful re-acceleration — most of Europe was flat or negative. The synchronicity is the single most important fact for sorting hypotheses.
What this rules out
Almost-synchronous slowdown across countries with different regulators, different education systems, different demographics, and different energy mixes is hard for U.S.-specific stories to explain. It strengthens common-shock stories — technology maturity (§ V) and capital cycles (§ VII) — at the expense of regulation (§ V) and education-as-cause.
The 1970s slowdown coincided with the OPEC oil crisis. The hypothesis is that energy-price shocks reallocate capital away from productive uses and force industry into energy-saving rather than growth-producing investment. It works if and only if the productivity curve tracks the oil-price curve.
Verdict — weak
Energy correlates with the timing of the 1973 break but fails the falsification test: when real oil prices fell sharply 1982–1998, productivity stayed at 1.5%. The post-2004 slowdown happened during ultimately falling real energy costs, especially the 2014–2020 cheap shale years. Denison estimated higher energy prices reduced productivity by only 0.1–0.3 ppt — far short of the ~1.3 ppt to explain. Energy was the precipitating shock for 1973; it isn't the cause of the long-run trend.
The U.S. regulatory state grew rapidly in the 1970s — OSHA (1970), EPA (1970), Clean Air Act (1970), Clean Water Act (1972), ERISA (1974) — and has continued to grow. Mercatus' RegData counts CFR restrictions (instances of "shall," "must," "required," "prohibited") at roughly 400,000 in 1970 versus 1.08 million in 2019. The Federal Register set a record at 106,109 pages in 2024.
Verdict — partial, 1970s only
Regulation grew fastest in 1970–1981 — that timing fits the 1973 break. But regulation also grew through the 1995–2004 productivity boom and continued growing during the post-2004 slowdown, with no visible inflection. Reagan-era deregulation didn't produce a productivity boom in the 1980s. The 2024 record-high coincided with a strong productivity year. Dawson–Seater (2013) estimate −2 ppt; mainstream industry-level studies find under 0.5 ppt. And regulation differs sharply across the G7, yet the slowdown doesn't.
GDP statistics were built for a steel-and-cars economy. Maybe modern GDP fails to value free internet services, intangible capital, and quality-adjusted product replacement. If true, the productivity slowdown is an illusion of statistics rather than a real change in the economy. Below: every credible estimate of how much "missing output" each mechanism could explain, vs the gap that needs explaining.
Verdict — largely rejected
Even summing every credible mismeasurement source — internet consumer surplus, free digital goods, creative destruction imputation, intangible capital — yields about 0.7 ppt, well short of the ~1.3 ppt the post-2004 gap requires. Most haven't accelerated since the early 2000s. Syverson (2017) closed the debate: the slowdown was global with no correlation to ICT intensity, and the "missing output" implied by the slowdown is $2.7 trillion — orders of magnitude bigger than any digital-surplus estimate. Mismeasurement is real and matters for welfare; it doesn't explain away the slowdown.
The 1995–2004 boom is, quantitatively, a story about IT investment plus IT-producer TFP. Information-processing equipment alone contributed 0.5 ppt to labor productivity in the 1990s. After the dot-com bust and the saturation of corporate IT, that contribution fell to 0.2 ppt and stayed there. The chart below shows the breakdown.
Verdict — partial; sub-component of the IT-cycle story
Capital deepening did slow after 2007, but only by ~0.3 ppt — about one-fifth of the post-2004 labor productivity gap. Most of the gap is TFP. The 1995–2004 episode looks like a one-time diffusion event for general-purpose IT (PCs, internet, ERP). This is the "low-hanging fruit" view (Cowen 2011, Gordon 2016): the corporate world only digitizes once. AI may be a second wave — Penn Wharton projects AI's TFP contribution rising through the 2030s — but the 2025 contribution from AI is estimated at 0.01 ppt. Most of the 2024–25 surge is something else.
Each row is a candidate explanation. Each column is one of the four facts the data forces every theory to explain. A green tile means the hypothesis fits the fact, a yellow tile means it partially fits, and a grey tile means it fails or contradicts.
Productivity grew 2.7% in 2024 — the strongest non-recovery year since 2004. TFP grew 1.3%, the first sustained TFP recovery since the IT cycle. Generative-AI adoption hit 26.4% of workers by H2 2024 (Bick–Blandin–Deming 2025). But the Penn Wharton Budget Model estimates AI's actual contribution to 2025 TFP at 0.01 percentage points. Most of the surge is something else — composition, post-pandemic reallocation, cyclical recovery.
Labor productivity from BLS / FRED OPHNFB, quarterly
1947–latest. TFP from BLS Productivity & Costs prod3 release
(Tables A & B) and FRBSF Fernald quarterly utilization-adjusted
TFP series. International from OECD Productivity Database +
Conference Board TED + Penn World Tables 11.0. Oil prices from
FRED WTISPLC. Regulation counts from Mercatus RegData
3.2 (CFR restrictions) and Office of the Federal Register.
Investment share from BEA NIPA Table 1.1.10 / FRED
A008RE1A156NBEA.
The 1973 and 2004 breaks are both estimated as roughly 1.3 percentage-point drops in trend labor productivity growth. Period boundaries follow the BLS major-business-cycle convention and Fernald (2014). Some scholars see a continuous deceleration rather than two breaks; the BLS convention is what's used here.
Total factor productivity is what remains of output growth after you subtract measured contributions of capital and labor. It's called the measure of our ignorance — anything mismeasured, any unmeasured input, any markup variation shows up there. Utilization adjustment (Basu–Fernald–Kimball 2006) materially affects estimates; revisions are routinely large.
Every annual figure comes from a public dataset you can verify. The verdicts below each chart reflect the academic consensus as of early 2026 — Syverson 2017 on mismeasurement, Fernald–Inklaar–Ruzic 2025 on the global pattern, Penn Wharton 2025 on AI's near-term contribution. Where consensus disagrees (Dawson–Seater vs mainstream on regulation), both magnitudes appear in the chart and the readout names the disagreement.
- BLS · Productivity & Costs (prod3 release) ↗
- FRBSF · Fernald quarterly utilization-adjusted TFP ↗
- Syverson (2017) · "Challenges to Mismeasurement Explanations" · JEP ↗
- Byrne, Fernald & Reinsdorf (2016) · Brookings ↗
- Gordon (2016) · The Rise and Fall of American Growth ↗
- Aghion, Bergeaud, Boppart, Klenow, Li (2019) · Missing Growth from Creative Destruction · AER ↗
- Fernald, Inklaar & Ruzic (2025) · Review of Income and Wealth ↗
- Penn Wharton Budget Model · AI productivity projections ↗
- BLS productivity prints have a ±1 ppt 80% confidence interval per quarter. Decade averages are more reliable but still revised at comprehensive NIPA revisions.
- TFP is a residual. Anything mismeasured, including intangible capital not yet capitalized by BEA, shows up there.
- International comparisons depend on hours-worked methodology that differs sharply across countries. Period boundaries follow Fernald (2014) and the OECD 2025 Compendium nowcast.
- The 2024–2025 surge is real in the data but its durability and cause are unproven. AI optimists and tech-maturity baseliners are consistent with the same dataset for the next 24 months at least.