AI Dark Output: Why AI Changes Everything—Except the Productivity Stats

Almost every tech headline says AI is changing the world. Big companies pour money into it; chip valuations and Magnificent Seven market caps dwarf whole regions. Look at official macro data, though, and something odd appears: the productivity lift from AI is hard to see. That contradiction is the starting point for a useful idea—AI dark output.

We’ve heard this song before

In the 1980s and 1990s, computers were everywhere—offices, factories, the press. Nobel economist Robert Solow quipped in 1987 that you can see the computer age everywhere but in the productivity statistics. That line became the Solow paradox. Only much later did accounting catch up: a major U.S. GDP revision in 2013 capitalized R&D and intellectual-property investment and lifted estimated 1990s output by on the order of trillions of dollars. We had been missing a Germany-sized block of activity because we did not know how to count it.

Compared with today’s AI measurement problem, that computer episode looks modest. Much of the value AI creates may never appear cleanly in traditional GDP, price indexes, or industry accounts.

What “AI dark output” means

Call the real economic value AI creates—but that national accounts miss or badly distort—“AI dark output,” by analogy with dark matter: you may not observe it directly, but you can sense it through how other variables behave. The problem is sharper because AI’s gains concentrate in services, historically the weakest link in national statistics. That weakness is decades old; AI amplifies it.

• Substitution dark output: work once done by people, now done by AI—often at a fraction of the old price, so the old market transaction shrinks or vanishes from the books.
• New dark output: useful work that simply would not have been done before, because human cost was too high—now done for pennies in tokens, with almost no invoice trail.

Why the mismatch matters

Incoming Fed chair Kevin Warsh has argued that if you only trust existing data, your view is backward-looking: you may miss that the economy can grow faster without the inflation the old gauges imply. Soft-spoken as that sounds, the sharper reading is that the statistical system is lagging an AI-shaped production function. Meanwhile markets see every cost—data centers, GPUs, power, water, displaced jobs—while most of the output stays invisible. Critics can say AI only burns money and cuts headcount. Dark output is not an excuse to ignore those costs; it is a warning that the other side of the ledger is under-measured. If AI is industrial-revolution scale, we need data that can see more than the destruction.

Cheap screws became countable output. Cheap AI work often does not. Until the measuring stick catches up, a real boom can look, in the numbers, like a quiet stall—or a bubble.