This is a brief philosophical prompt from the Import AI newsletter, not a substantive news article. It asks if readers feel they are living through a revolution, but provides no specific information or events.
SynthePulse Insight · AI deep reading
Version 1 · 1 source
The U.S. AI economy is growing at over 2,000% per year, but traditional GDP statistics almost entirely miss this phenomenon. Economists warn that this measurement blind spot could leave policymakers unprepared for labor market shocks.
Economists from the University of Virginia, Anthropic, and the Bank of Canada jointly published a paper stating that the U.S. AI economy is growing at an unprecedented rate, yet this growth is nearly invisible in traditional GDP statistics. They estimate that U.S. nominal AI GDP in 2025 is about $250 billion, with a quality-adjusted real annual growth rate of roughly 2,600%.
The difficulty in measuring growth stems from two factors: while data center construction is massive, it is not yet large enough to significantly boost overall GDP; and AI inference (i.e., the use of AI systems) contributes most of the economic impact but suffers from confounding factors—'the unit price for a given level of AI capability is falling almost as fast as quality-adjusted output is growing,' resulting in modest nominal AI income growth.
The paper proposes three methods to measure the AI economy: nominal compute spending, raw compute capacity, and quality-adjusted AI output. In terms of nominal compute spending, the U.S. went from $37 billion in 2023 to $90 billion in 2024, and then to $219 billion in 2025. For raw compute capacity, actual capacity grows even faster due to efficiency gains from new chips—'U.S. AI compute capacity is growing more than 200% per year.'
Quality-adjusted AI output growth is the most striking: accounting for algorithmic progress (measured by inference price at a fixed benchmark performance) and assumptions about declining training costs, 'these efficiency gains imply that quality-adjusted AI output grew by about 2,290% in 2024 and about 2,271% in 2025.'
The paper warns that if the AI economy cannot be accurately measured, policymakers may be caught off guard. 'AI is the latest in a series of rapidly improving technologies following semiconductors and the internet that have raised measurement concerns,' but a key difference is that AI may have a larger impact on labor: 'In previous cases, rapidly improving technologies were, on net, complements to human labor. AI is the first rapidly improving domain that may massively substitute for human labor.'
The authors note: 'A Treasury Department running ten-year revenue projections based on conventional data will substantially underestimate the probability of a shock to the labor tax base—and thus fail to design responses such as tax reform, sovereign wealth funds, or other benefit-sharing plans. An invisible windfall cannot be shared.'
To address measurement challenges, the paper offers three recommendations: First, statistical agencies should develop 'AI satellite accounts' that produce metrics like nominal compute spending to aid overall GDP calculation; second, statistical agencies, businesses, and academia should collaborate to generate better raw data, such as the allocation of training versus inference compute resources; third, policymakers should incorporate AI production capacity measurements into medium-term economic forecasts.
Researchers from the UK AI Safety Institute argue that using AI systems to oversee AI training processes (automated alignment research) is not a panacea. 'Errors in automated alignment research may be harder to identify than human baselines' for reasons including: AI research is optimized for human approval, AI errors are unintuitive to humans, research content is highly correlated, research volume is enormous, and alignment solutions may rely on arguments that humans cannot evaluate.
They suggest improving the situation through measurement (e.g., testing AI's ability to continue research by logging successful projects), generalization experiments (e.g., testing different training agents on post-knowledge-cutoff data), and scalable oversight (e.g., compressing research paper corpora).
This article is primarily based on Import AI's coverage of two papers: one by economists from the University of Virginia, Anthropic, and the Bank of Canada, and another by the UK AI Safety Institute. Import AI is a secondary source, but the papers themselves are primary sources. All numbers and conclusions come from direct quotes in the coverage, without introducing external knowledge.
The AI economy is growing at an extraordinary rate, but traditional statistical methods render it nearly invisible. Policymakers must improve measurement tools, or they will be unable to respond to potentially severe labor market shocks. Meanwhile, relying on AI for safety oversight is not straightforward; its inherent difficulties require systematic solutions.
Primary report
Primary source