Every few weeks, a new study lands claiming to have counted the jobs artificial intelligence will eliminate. The numbers are always alarming. The methodology is always the same. And the methodology is always wrong.

Not wrong because the researchers are incompetent. Wrong because they are measuring the wrong thing. Jobs are not the unit of analysis. Tasks are. And until the public debate, the policy community, and the governance frameworks catch up to that distinction, we will keep building the wrong policy for the wrong problem, with roughly predictable results.

We have been here before. Thirty years ago, the smartest economist in the most consequential job in American government sat with access to more economic data than any person in history and could not figure out where productivity had gone. The resolution of that puzzle tells us exactly where we are today, and more importantly, where this is heading. The governance implications are not subtle.

What a Job Actually Is

A job is a bundle of tasks. This seems obvious once stated, but the implications are radical and the public discourse almost entirely ignores them.

Consider a radiologist. The bundle of tasks that constitutes the job of radiologist includes reviewing imaging studies for anomalies, integrating findings with patient history, communicating diagnoses to referring physicians, consulting with other specialists, managing uncertainty in diagnostically ambiguous cases, and a significant amount of administrative and documentation work. AI systems, specifically deep learning models trained on large labeled datasets, are now performing some of those tasks at a level of accuracy that competes with or exceeds human performance. Reviewing imaging studies for certain anomalies is one of them.

The question the public debate asks is: will AI take the radiologist’s job? That is the wrong question. The right question is: which tasks within the radiologist’s bundle can AI perform, and how does removing those tasks from the human’s workload reconfigure the remaining bundle? The answer might be that the job transforms but persists, that fewer radiologists are needed to handle the same volume of cases, that radiologists shift toward higher-order diagnostic work and consultation, or some combination of all three. What it almost certainly does not mean is that radiologists simply disappear.

This framework, now foundational in labor economics, was formalized by David Autor, Frank Levy, and Richard Murnane in their 2003 paper “The Skill Content of Recent Technological Change: An Empirical Exploration” in the Quarterly Journal of Economics. Their central insight was that technology does not substitute for workers. It substitutes for specific tasks. The displacement effects flow from which tasks get automated, and the labor market adjusts around what remains.

Autor, Levy, and Murnane were describing what had happened with computers and information technology through the 1980s and 1990s. The historical pattern was clear: IT was particularly good at routine tasks, both cognitive and manual, that could be codified into explicit rules. Bookkeeping, data entry, repetitive assembly, form processing. These tasks, mostly performed by middle-skill workers, were systematically automated. The result was not mass unemployment but labor market polarization: growth in high-skill, high-wage work and growth in low-skill, low-wage service work, with the middle hollowing out. The jobs did not disappear wholesale. They restructured.

The current debate about generative AI complicates this picture considerably, because large language models are not especially good at routine codifiable tasks in the old sense. They are, it turns out, quite capable at non-routine cognitive work. This has led some observers to conclude that the old task-based framework breaks down for AI. It does not. It means the task-based framework applies to a different and broader set of tasks than the previous technological wave. The unit of analysis is still the task. The question is still which tasks are being displaced, and how the remaining bundle reconfigures.

Policy built at the job level cannot see any of this. It can only count and recount, which is exactly what it is doing.

Greenspan’s Mirror

In October 1995, Alan Greenspan addressed the Economic Club of Chicago and said publicly what he had been working through privately for months: he could not find the productivity gains that the information and communications technology revolution should have been producing. Computers were everywhere. The semiconductor, the microprocessor, and the satellite had transformed American business. And yet the productivity statistics were essentially flat. Greenspan worried out loud that all this apparent technological progress might be mere “wheel spinning”, changing production inputs without increasing output, rather than real advances in productivity.

He was not alone in the confusion. In 1987, MIT economist Robert Solow had put the same puzzle more memorably: “You can see the computer age everywhere but in the productivity statistics.” The observation became known as the Solow Paradox, and it haunted economic policy debates through the early 1990s.

The stakes were not abstract. Greenspan’s uncertainty about productivity had direct monetary policy consequences. If productivity was genuinely rising but not showing up in the data, then the economy could sustain lower unemployment and faster growth without triggering inflation, and premature rate hikes would choke off a real expansion. If productivity was not actually rising, then the apparent boom was inflationary and the Fed needed to act. Getting the unit of analysis wrong, or more precisely, using measurement frameworks calibrated to a wheat-and-steel economy to evaluate an information economy, had direct and concrete policy costs.

The resolution came slowly. By his February 1998 Humphrey-Hawkins testimony before the House Subcommittee on Domestic and International Monetary Policy, Greenspan was beginning to see it. The acceleration in capital investment in advanced technologies beginning in 1993 was producing a “noticeable pickup in productivity” that conventional measures were having difficulty capturing. The issue was partly measurement and partly lag: the productivity gains from information technology required substantial complementary investments in new processes, business models, and worker skills before they showed up in the aggregate numbers. The technology was real. The gains were real. The measurement frameworks were wrong, and the lag was long.

By October 1999, addressing the Business Council in Boca Raton, Greenspan had the picture fully in view. In a speech on information, productivity, and capital investment, he described precisely what had happened: information technology had reduced the hours of work required per unit of output by eliminating the redundancies that businesses had historically maintained to cope with uncertainty. The productivity gain was not showing up as more output per worker in the traditional sense. It was showing up as fewer workers needed to guard against the unpredictable, fewer inventory buffers, fewer backup teams, less friction throughout production and distribution systems. The measurement frameworks were looking for output per hour. The gains were accruing as reductions in wasted hours. The frameworks were measuring the wrong thing.

Solow himself completed the loop in 2000 when he told the New York Times that you could now, finally, see computers in the productivity statistics. The paradox had resolved. It had taken the better part of a decade.

The governance failure in the 1990s was that policy, regulation, and institutional frameworks were calibrated to the wrong unit of analysis for most of that decade. Not because the people involved were unintelligent. Because the measurement frameworks had not caught up to what the technology was actually doing.

We are in exactly the same position today. The AI discourse is measuring jobs, counting them, projecting their disappearance, building policy around the projections. The technology is operating at the task level. The frameworks cannot see what the technology is doing with any precision, and the lag between what is actually happening and what the data will eventually show is measured in years.

What the Resolution Tells Us

Greenspan’s retrospective account in The Age of Turbulence, published in 2007, is instructive less for what it says about the 1990s than for what it implies about now. The gains from information technology were real, ultimately enormous, and distributed in ways that the initial discourse had not predicted. The policy debate of the early 1990s was obsessed with displacement and disruption. The actual outcome was a productivity-driven expansion of historic length.

But “the outcome was fine” is not the governance lesson. The governance lesson is that the distribution of gains was neither automatic nor equitable, and that the policy frameworks that were in place during the transition were poorly suited to shaping that distribution. The hollowing of middle-skill employment that Autor, Levy, and Murnane documented was not a law of nature. It was in part a consequence of governance frameworks that could not see the task-level restructuring clearly enough to intervene at the right level.

The same dynamics are already visible with generative AI. The task displacement is real and proceeding faster than the previous wave. The gains will accrue somewhere. The question, as it was in the 1990s, is whether governance frameworks can see clearly enough, soon enough, to shape how those gains are distributed.

They cannot, at present, because they are measuring jobs.

A framework calibrated to count displaced jobs will always be fighting the last war. By the time a “job” is visibly gone, the task-level restructuring that produced that outcome has been underway for years, in ways that earlier intervention might have redirected. Retraining programs built around job categories miss the point: the category may survive while the tasks within it shift completely. Social insurance systems keyed to employment status miss the point: the relevant economic disruption may be occurring among workers who remain formally employed but whose real wages and bargaining power are eroding as their highest-value tasks are displaced.

The governance problem is a measurement problem. The measurement problem is a unit-of-analysis problem. And we have the historical record of the Solow Paradox to tell us exactly how long it takes to resolve a unit-of-analysis problem when the underlying technology is moving fast: longer than anyone wants to wait.

The Policy Implication

The practical implication is not complicated, though it is inconvenient for anyone who has already built a framework around job counts.

Governance frameworks for AI labor market effects need to be built at the task level. That means occupational analysis systems like O*NET, which already catalogs hundreds of occupational tasks at fine granularity, need to become primary inputs into policy rather than background reference material. It means labor market monitoring needs to track task displacement in real time, not occupational employment totals with a two-year lag. It means retraining and adjustment programs need to be designed around task bundles and how they reconfigure, not around job titles.

None of this is technically difficult. The data infrastructure exists. The analytical frameworks exist. The labor economics literature has been working at the task level for two decades. The governance frameworks have simply not caught up.

Greenspan eventually found the productivity. It was there the whole time, doing things the existing measures could not see. The same is true of what AI is doing to the labor market. The disruption is real, the gains are real, and the existing measures are looking in the wrong place.

Build the frameworks around tasks, and you can see the problem clearly enough to govern it. Keep counting jobs, and you will spend the next decade discovering the answer too late to do much about it.

The foundational task-based framework in labor economics is Autor, Levy, and Murnane (2003), available through JSTOR or the NBER working paper archive. Greenspan’s October 1995 Economic Club of Chicago remarks are archived at FRASER, the St. Louis Fed’s historical archive. His February 1998 Humphrey-Hawkins testimony and his October 1999 Business Council speech are available directly from the Federal Reserve Board.