The literature on AI-driven job displacement spans a wide range of projections — from apocalyptic (47 percent of U.S. jobs at high risk, per Frey and Osborne 2013) to sanguine (less than 10 percent, per OECD revisions accounting for task heterogeneity within occupations) — and this variance is not merely a measurement problem but reflects genuine conceptual disagreements about what is being projected and how. A sober assessment requires distinguishing between several related but different claims: automation feasibility (can AI technically perform a task?), automation adoption (will firms deploy AI to perform that task, and at what pace?), net employment effect (what new work will be created as a result?), and distributional consequences (which workers and communities will bear the costs?). Conflating these four questions produces both the panic and the complacency that characterize much popular discussion.
The Frey-Osborne framework, which assessed the susceptibility of 702 occupations to computerization and found 47 percent at high risk, has been the most cited and most criticized projection. Its primary methodological limitation is that it treats occupations as uniform wholes: if any significant fraction of a role is automatable, the occupation is classified as at risk. OECD researchers Arntz, Gregory, and Zierahn responded with a task-level analysis that found only 9 percent of U.S. jobs highly automatable when within-occupation task heterogeneity is accounted for. The McKinsey Global Institute estimated that 60 percent of occupations have at least 30 percent of their tasks automatable with existing technologies — a finding that, importantly, suggests extensive task automation without necessarily implying equivalent job elimination.
The more recent empirical evidence from large language model deployment suggests that the task composition finding is particularly important. MIT-IBM Watson AI Lab research found that, contrary to prior automation waves that concentrated on routine manual and cognitive tasks, generative AI disproportionately affects high-wage, high-education cognitive work: writing, analysis, coding, legal research, customer interaction. This inverts the pattern of previous automation waves and has direct implications for projections: the workers most at risk from AI-driven displacement are not the least-educated and lowest-paid (who dominated previous displacement) but mid-to-high-wage white-collar knowledge workers whose skills have historically been considered safe from automation. The political economy implications are significant: these workers have more political voice, more mobility, and more resources to adapt — but also more expectations of permanence in their labor market position, making displacement more socially and psychologically disruptive relative to their expectations.
The adoption curve is as important as the feasibility question. Even for tasks that AI can technically perform, adoption by firms takes time, requires capital investment, involves organizational change, and is constrained by regulation (in healthcare, law, and finance) and by the practical difficulty of deploying AI in workflows that are not fully standardized. Goldman Sachs research from 2023 estimated that generative AI could automate 25 percent of current work tasks globally, but projected this displacement to unfold over a decade or more — a much slower adoption trajectory than the technological capability alone would imply. The economic history of automation confirms that adoption lags capability by years to decades: the technology for fully automated checkout has existed for decades, but cashier jobs have declined gradually rather than catastrophically.
Sober projections also require attention to the "so-so automation" problem identified by Acemoglu and Restrepo: automation that replaces human labor without creating significant productivity gains or new tasks is economically harmful, reducing employment without increasing output sufficiently to generate new demand for labor. Industrial robots in manufacturing and some forms of AI automation in service industries may fall into this category — they reduce labor costs and increase capital returns without proportionally increasing economic output or expanding markets in ways that generate new jobs. The so-so automation hypothesis is empirically contested but important because it challenges the standard economic assumption that automation is always productivity-enhancing and therefore ultimately employment-neutral.
Law 5's framework of transparent archive is directly applicable here: the displacement projections discussed above are all based on samples, surveys, and modeled assumptions rather than real-time tracking of actual displacement and job creation. Building the institutional infrastructure for continuous, granular monitoring of AI-driven labor market changes — at the level of specific firms, occupations, and geographic communities — is the precondition for both accurate projection and effective policy response. The current state of labor market data in the United States is insufficient for this purpose: the Bureau of Labor Statistics' occupational employment surveys have multi-year lags, do not capture the task content of work with sufficient granularity, and do not track the reasons for job loss with the specificity needed to identify AI displacement distinctly from cyclical layoffs or other causes.
The distributional dimension is ultimately more important than the aggregate projection. Even scenarios where AI creates as many jobs as it destroys will be experienced as catastrophic by workers whose specific skills become obsolete, in communities whose employment base contracts, and in the time period between displacement and reabsorption. The aggregate statistics average over genuine human devastation, and policy designed to manage average outcomes will fail workers in the distributional tails. Sober projections therefore require not just aggregate employment forecasts but granular distributional analyses: which occupations, which skill levels, which geographic concentrations, which demographic groups, and over what time horizons will displacement be concentrated?