AI Dark Output: The $1.5 Trillion Measurement Problem in Global GDP
Key Takeaways
- ▸AI is producing substantial unmeasurable economic value ('Dark Output') that GDP and labor statistics cannot capture, leaving economists unable to quantify corresponding gains against visible costs
- ▸SemiAnalysis estimates $1.5 trillion in immediate substitution dark output (tasks AI could automate), with new dark output expected to eventually exceed this significantly
- ▸Service sector dominance of AI output amplifies the measurement problem—these sectors already show distorted productivity metrics, making AI-driven gains even harder to detect
Summary
SemiAnalysis has published a major analysis identifying 'AI Dark Output'—economic value created by artificial intelligence that national accounting systems cannot currently measure. Drawing parallels to the 1980s-90s 'computer paradox' where Robert Solow noted "You can see the computer age everywhere, but in the productivity statistics," the report documents how AI's rapid expansion is generating invisible economic gains while highly visible costs (job displacement, energy consumption, capital spending) are captured in real-time.
The analysis identifies two categories of dark output: substitution (AI replacing human labor, ~$1.5T identified) and new output (entirely new economically feasible tasks previously too expensive). Most critically, service-sector AI work—where measurement already fails—will likely dominate future dark output. Incoming Federal Reserve Chairman Kevin Warsh acknowledged the measurement risk in December 2025, warning that reliance on backward-looking data could trigger policy errors and recession misdiagnosis. The scale potentially rivals the computer revolution's statistical blind spot, with profound implications for inflation assessment, productivity tracking, and investment valuations.
- Policy and market risk: If AI productivity gains remain statistically invisible, the Fed could tighten prematurely, valuations could face 'bubble' scrutiny despite real gains, and unemployment statistics will appear worse than economic reality
Editorial Opinion
This analysis captures one of the defining economic measurement crises of our era. If AI-driven dark output reaches 30%+ of unmeasured activity—as the 1990s computer revision suggested—policymakers face a years-long blind spot in inflation, employment, and growth signals. The risk is acute: tightening monetary policy during genuine non-inflationary productivity expansion, or dismissing AI capex as bubble dynamics precisely when real value creation is accelerating. Unlike past measurement revisions discovered post-hoc, this one is happening in real-time, visible mainly through its effects: job displacement and soaring corporate spending with no apparent output.
