The inability to enumerate what is occurring to work renders it impossible to cope with what is occurring to workers due to automation. At the end of the 19th century, Massachusetts reacted to the grinding logic of mechanized industry with a pretentious irrelevance: it attempted to count. Fairness was not created by the state labor bureau, the predecessor of the modern Bureau of Labor Statistics, but it made visibility an institution. The time, remuneration, mutilations, and the process of grinding livelihoods became arguably readable. Gradually, a common scoreboard helped a sprawling economy to take in the painful change without breaking down into out-right war.
Measurement is now confronted by a stranger, as civic technology. Artificial intelligence does not come to the world in the form of a loom at the factory. It comes as software that slips into the current processes, quietly or executive swagger. It is capable of summarizing, drafting, categorizing, translating and coding at a pace which crumples the previous correlation between effort and performance. The promise of AI is not just increased productivity in the workplaces constructed on credentialed attention, which is legal review, customer support, and finance operation, marketing collateral. It is the redefined formulation of the number of people that are expected to work to produce the same output, as it “should” be.
The problem is that the instruments of national policy that were created to observe the labor-market change were constructed in the times when the changes were going to be slower. BLS continues to sample about 60,000 households in a month and the response rates have declined. The former can explain what has already occurred with the commendable granularity. It can hardly collect anything of what is going on within firms until layoffs manifest themselves in the form of unemployment, or job descriptions transform into something unfamiliar. Even the keenest economists are reduced to looking through a fog, debating about whether an early indicator is actual or it is just the normal noise of the economy.
Other more recent studies indicate that the fog is disappearing. A long-run index of “occupational churn” constructed using over 100 years of U.S. Census data was unexpectedly peaceful between 1990 and 2017 before showing a higher-order change since 2019 and what appears to be the trend of reallocation by the AI. The identical work observes that the proportion of STEM employment has increased between 6.5 in 2010 and close to 10 in 2024, a structural skew of the labor market toward more technical labor which is paired with intensive investment in frontier tools. Meanwhile, the percentage of jobs in retail sales declined significantly throughout the 2010s, and not all the low-paying service jobs have recovered as much as older job polarization models would predict.
But the most dramatic unit of account is only the “jobs”. Wages can be set in motion, and more silently. The global bodies have cautioned that AI will claim a big portion of work; IMF has estimated that some 40 percent of the jobs in the world will be exposed to it, more in the developed world. In the case of a machine, having the capacity to complete a significant slice of what a job used to dictate, bargaining power may be reduced without a pink slip ever being extended. The labor market is able to appear sound in that world as the conditions of work shift against workers.
The second complication is adaptation. A survey of companies that have implemented AI has identified more retraining than layoffs, and pilot workforce programs in the U.S. have found training to be generally positively associated with earnings growth among the workers who have lost their jobs. Nevertheless, the benefits are not evenly spread: the returns to training in more heavily exposed jobs were 25 percent lower on average than the returns in less exposed jobs, and the returns were further reduced on top, to more exposed jobs. The factual item does not lay blame on training; it explains how malevolent the new task environment can be on employees when they are requested to run faster than software when trying to master the new product.
In the meantime, in the non-academic world, projections vary. The productivity increased, according to one research note, by an average of about 15 percent when generative AI takes the place of humans in the future, and advocates the idea that the rate of baseline replacement is between 6-7 percent, but with a broad range of variation. The diffusion is significant, as rapid diffusion is amenable to be counteracted by retirements, career mobility, and new company formation; rapid diffusion places an institution (training systems, safety nets, employers, and local economies) in a schedule that it was not created to fulfill.
The debate on whether AI is producing more jobs than it is eliminating will not be solved through counting. But it is what causes the argument to be pegged to reality. Provided that AI is restructuring work by freezing employment, enforcing silent productivity requirements, and administering wage pressure before the loss of jobs turns visible, then a democracy that does not want to spook the unemployment rate at all is opting to learn too late.
