AI's speed gains mask a fundamental productivity problem

Matt Scherer, a fellow at the Open Markets Institute focused on policy responses to the AI bubble, argues that generative AI's apparent productivity benefits mask a fundamental economic problem: faster individual work does not equal real productivity gains.

The intuition behind AI's promise is straightforward. ChatGPT, Claude, and similar systems write faster than humans; people report 10× productivity improvements; empirical studies show significant speed gains in tasks like coding. It seems obvious that economy-wide productivity should follow. Yet four years after ChatGPT's release, neither individual companies nor the broader economy have seen measurable productivity improvements from generative AI. Some call this a "productivity paradox," echoing economist Erik Brynjolfsson's term for the 1970s–80s IT era, when technology advanced but productivity statistics barely moved. Brynjolfsson and others predict that as companies restructure for AI, real gains will follow.

But Scherer identifies a simpler explanation: the confusion between colloquial and economic productivity. Colloquially, productivity means producing more stuff faster. Economically—the definition that matters for business and the economy—it measures the ratio of economic value of outputs to the economic cost of inputs. The numerator depends on quality, not just quantity. The denominator includes all input costs: labor, capital, infrastructure. A sewing machine was revolutionary not because it was fast but because it was fast *and precise*—if it had produced misshapen or torn clothes, speed alone would not have mattered.

Generative AI fails this test. The AI infrastructure buildout has already cost $1.5 trillion（約240兆円）, with spending expected to accelerate further next year, plus perpetual costs for powering and cooling chips. AI's errors are severe and routine: it fabricates case law, deletes code databases, and inserts incorrect patient diagnoses. The result is that professionals spend more effort correcting AI mistakes than they would have spent completing tasks manually from the start. This drives up input costs (more labor for quality control) while degrading output value (errors reduce real-world utility). Both sides of the productivity equation suffer.

The author acknowledges that some argue these errors will soon become rare enough that speed gains will dominate. But technological progress typically shows diminishing returns; today's AI may be about as good as it will ever get. Even if reliability improved dramatically, the economic value of AI-generated output remains questionable. The app development market offers a telling example: AI nearly doubled the number of app releases per month, yet the number of actively-used apps has started to decline. More output did not create economic value.

Additionally, some AI use cases actively destroy productivity. Generative AI excels at cyberattacks but fails at secure code, harming cybersecurity specialists' productivity. Low-quality AI "workslop" wastes time in many settings and makes identifying fit-for-purpose work harder. Unlike the Industrial Revolution—which met desperate demand for clothing, food, and transportation by automating production—generative AI operates in markets where low-hanging fruit has been plucked and demand for additional output is unclear. Scherer concludes: there may be no productivity paradox at all, just a productivity illusion.

The article challenges a widespread assumption: that because generative AI makes individuals work faster, it must improve overall productivity. This assumption rests on conflating two different definitions of productivity. In everyday speech, productivity means doing more in less time; economically, it measures the ratio of economic value generated to the cost of inputs required. The disconnect between these meanings is where the productivity illusion lives.

The author traces this tension through history. The "productivity paradox" of the 1970s–80s occurred when IT advanced rapidly but economy-wide productivity barely budged; productivity eventually improved once companies restructured around computers. AI optimists predict the same trajectory. But the article argues the comparison fails: generative AI's error-proneness—fabricating case law, deleting databases, inserting wrong diagnoses—forces professionals to spend more time correcting mistakes than they would have spent doing the original task, which raises input costs and destroys the value of speed gains. The $1.5 trillion（約240兆円） already spent on AI infrastructure must somehow generate enough value to justify the cost; instead, current evidence (the near-doubling of app releases paired with declining use of active apps) suggests more output is not translating to economic value. The pattern may repeat across other knowledge-work fields, suggesting the AI productivity boom some predict may never arrive.

No comments yet. Be the first to share your thoughts!