In 1987 the economist Robert Solow delivered the most quoted line in the study of technology and growth: you could see the computer age everywhere except in the productivity statistics. Nearly four years after ChatGPT arrived, a new report from Google's own economics team suggests we have a fresh version of the same puzzle. AI appears to be everywhere, and yet its mark on employment, output and growth remains stubbornly hard to find. The interesting thing is that Google's data does not just restate the paradox. It quietly explains it.
The report, called ATLAS v1.0, maps roughly 15 million anonymised interactions with Gemini across more than 150 countries. On the surface the numbers read like a story of mass adoption: AI now shows up in 68 percent of detailed occupations, covering nearly 90 percent of US employment. Look closer and the picture thins out. In a typical occupation that uses AI at all, the technology touches only about a fifth of tasks. Just 3 percent of occupations use it for more than three-quarters of what they do, and those are narrow roles like software quality testers and document specialists. Adoption is broad, in other words, but shallow.
The shape of that usage matters. Attempts to hand a task over end to end account for under 10 percent of AI conversations in the higher-value "non-routine cognitive" work, the hypothesis-testing and creative design that makes up the bulk of professional output. The dominant pattern is collaborative: give me a rough draft, summarise this report, help me brainstorm. Google is blunt about what this does and does not show. It finds no evidence, it says, that AI is about to trigger mass automation of white-collar work, that it is irrelevant to blue-collar jobs, or that its purpose is simply to replace people. The tool is being used as an assistant, not a substitute.
The optimistic reading, favoured by the American Enterprise Institute's analysis of the report, is that we are still on the downstroke of a J-curve. When a firm adopts a genuinely important technology, it has to rewire its processes before the payoff arrives, spending on retrained staff and rebuilt workflows that nobody officially counts as output. For a while the company looks like it is spending more to produce the same, until the reorientation finishes and productivity finally jumps. On this view the missing numbers are a matter of patience, not disappointment.
Boards are not feeling patient. Steve Lucas, chief executive of the data firm Boomi, says the mood in the meetings he sits in has flipped from "build an AI strategy" to "show me the return," a shift he sums up as "ROI supersedes AI." He points to the four big US hyperscalers, whose combined capital spending on AI jumped from around $400 billion last year to over $700 billion this year, and warns that the grace period for unproven projects is real but closing fast. Even Greg Brockman of OpenAI sounded relieved that customers have started scrutinising cost, saying the world felt "rational again." That is the tension the ATLAS report crystallises. If AI is mostly polishing drafts rather than transforming work, the question is whether the productivity is merely late, or whether the tool is being asked to do something smaller than the spending assumes.