Why does AI productivity appear weak when investment is high?

Brynjolfsson, Rock and Syverson (NBER Working Paper 25148, October 2018, revised January 2020) argue that general purpose technologies (GPTs) — "engines for growth" that are pervasive, improve over time, and spawn complementary innovation — require large, mostly intangible complementary investments that standard productivity statistics cannot see: "new business processes," "managerial experience," worker training, and software patching. Because these intangible investments are "not readily tallied on a balance sheet or in the national accounts," they generate a measurement artifact the authors name the Productivity J-Curve: measured total factor productivity (TFP) growth is "initially underestimated" while firms are pouring resources into unmeasured intangible capital, then later "overestimated" once that hidden capital stock starts generating measurable output. Applying this logic empirically to R&D, computer hardware, and software, they report an adjusted TFP measure "11.3% higher than official measures at the end of 2004, and 15.9% higher than official measures at the end of 2017."

The mechanism is a growth-accounting model extended to include an intangible capital stock U alongside tangible capital K and labor L. The authors infer the size of U using Tobin's q: when a firm's market value exceeds the replacement cost of its observable tangible assets, they read the premium as the market pricing in intangible assets correlated with that tangible investment — for example, a firm buying computer equipment also investing, unmeasured, in AI-engineer training. The resulting mismeasurement term decomposes into two opposing effects: one that understates true productivity growth because intangible investment goods are not counted as output when produced, and a second that overstates it later because the output flowing from that now-built intangible capital gets credited to "productivity" rather than to the capital that produced it. They root the pattern in prior GPT transitions — the multi-decade "Engels' Pause" of the British industrial revolution, the generation it took to redesign factory layouts around electrification, and Solow's own 1987 productivity paradox roughly two decades into the IT era — treating AI as the next candidate case, though they caution their AI-specific estimate captures effects that are "small but growing" and is offered "more speculatively" than the R&D, hardware, and software results.

This paper is the theoretical ancestor of the claim Is AI productivity finally showing up in economic data? invokes by name: that 2026 op-ed reads a weak 2025 jobs report alongside rising GDP as evidence the economy is moving from the J-Curve's "investment phase" (where this paper's understatement term dominates) into its "harvest phase" (where the overstatement term should start to show), but does so with a residual inference from one revised jobs report rather than the market-value-based measurement this paper develops to size intangible capital directly. It also stands in a different relationship to Do AI productivity gains feel larger than they actually measure?: that survey's "productivity paradox" of perceived gains exceeding measured ones runs in the same direction as this paper's early-stage underestimation claim, but locates the lag in delayed revenue recognition rather than in an unmeasured intangible capital stock inferred from market valuations — a distinct mechanism for a structurally similar mismeasurement pattern.

The excerpt's quantitative results (the 11.3% and 15.9% TFP adjustments) are derived from R&D, computer hardware, and software data running through 2017, not from AI-specific data; the paper is explicit that its treatment of AI intangible capital is comparatively speculative and the measured effect so far small. The q-theory method also requires intangible investment to be correlated with measurable tangible investment to be detected at all — by the authors' own caveat, any intangible investment uncorrelated with observable capital will not show up in their estimates. The implication is that this paper supplies the measurement framework and historical pattern later cited to interpret AI-era data, not a direct empirical measurement of an AI productivity J-Curve itself.