Can the Recovery Mechanism Survive AI? Skill Formation, Labor, and What Current Measurement Misses
Aysa Xuemo Fan · 2026
Read Can the Recovery Mechanism Survive AI? Skill Formation, Labor, and What Current Measurement Misses by Aysa Xuemo Fan (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2605-16283.txt.
Introduction
Fan frames a stock-building measure gap in AI's effects on skill building. Field data show AI use focused in skilled work aiding instant task results, but observe current skill rather than whether users grow able to perform tasks alone. Lab studies measure solo ability more head-on, but only in small groups with design-bound results. Fan merges both evidence streams by ID strength and offers a research program linking agreed usage records to solo tests, varying whether AI supplies answers, hints, feedback, or grading.
Catalog record
| CREATOR | Aysa Xuemo Fan |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 54480 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2605.16283
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Can the Recovery Mechanism Survive AI? Skill Formation, Labor, and What Current Measurement Misses Aysa Xuemo Fan Abstract Throughout the modern era, when new technologies displaced workers, societies adapted through the same mechanism: education raised the cognitive ceiling, producing workers capable of tasks machines could not yet reach. Generative AI may be the first technology to break this cycle, because it now operates at the top of that ceiling. Drawing on labor economics, deployment data from millions of AI conversations across multiple platforms, original reanalysis of two public datasets, and skill-formation experiments, this paper develops three contributions. First, a stock-versus-flow framework showing that economic data and education data tell divergent stories about the same technology: augmentation dominates current workers, but the developmental pipeline producing the next generation is under strain. Second, a systematic gap analysis of the evidence base, revealing that the knowledge dimension of cognition is unmeasured across all major studies, that the three studies measuring learning outcomes (each n < 200 n<200 ) consistently find AI improves performance without improving learning ( d = 1.21 d=1.21 in our cross-platform reanalysis), and that no study bridges professional and student populations. Third, an extended cognitive taxonomy (judgment under uncertainty, epistemic identity, and epistemic agency) applied to three cases from the evidence to distinguish AI interaction patterns that preserve learning from structurally similar ones that erode it. The paper argues that AI’s societal risk lies not in replacing teachers but in eliminating the productive struggle through which the next generation’s capacity forms, and proposes a research and design agenda targeting what current measurement systems miss. Part I: The Historical Pattern Every story about artificial intelligence’s impact on society is, whether its authors realize it or not, a story about a pattern that has played out many times before. The pattern is this: a new technology arrives, it displaces workers, those workers suffer, and eventually, sometimes generations later, society adapts. The mechanism of adaptation, every single time, has been education. To understand what is different about AI, we first need to understand what has been the same across centuries of technological change. The Pattern of Displacement In The Technology Trap , 15 draws a fundamental distinction between two types of technological change. Enabling technologies create new tasks that humans could not do before: the telescope enabled astronomy, spreadsheet software enabled complex financial modeling, search engines enabled rapid information retrieval. Workers gain new capabilities, and resistance to the technology is typically low. Replacing technologies, by contrast, perform existing tasks cheaper or faster: the power loom replaced hand-weaving, the assembly robot replaced manual assembly, autonomous checkout replaced cashiers. Workers lose their livelihoods, and resistance is fierce. The historical examples are vivid. In 1907, six hundred lamplighters in New York went on strike when gas lights were replaced by electric. Within twenty years, a profession that had been a neighborhood institution since 1414 was extinct ( 15 ) . In Verviers, Belgium, lamplighters attacked replacement workers; the government called in the army. The mechanized factory told the same story at a larger scale: middle-income artisan craftsmen who could fashion a product from start to finish were replaced by less-skilled factory operatives performing a narrower set of tasks aided by machinery. An artisan’s lifetime of accumulated skill was rendered obsolete by a machine a child could tend. The scale of these transformations is difficult to overstate. Before 1750, per capita income doubled every six thousand years; since then, every fifty years ( 15 ) . But the process was brutal for those caught in the transition. Working Englishmen during the Industrial Revolution, as 15 documents, “were made worse off as technological creativity was allowed to thrive. And those who lost out did not live to see the day of the great enrichment.” The transition period can last generations, and those who bear its cost often do not survive to enjoy its benefits. The Recovery Mechanism: Education Every time technology displaced workers, society eventually adapted through the same mechanism: education. This is the central empirical finding of two centuries of economic history, most comprehensively documented by 16 . Jan Tinbergen ( 45 ) first formulated it as a race: the pattern of inequality in any society is determined by whether the supply of educated workers keeps pace with technological demand. 16 extended this framework into a full empirical account of the United States from 1890 to 2005. When supply leads, as it did from 1915 to 1960 when the supply of skilled workers grew approximately one percent per year ahead of demand, wages compress and the middle class expands. When demand leads, as it has since 1980, wages diverge and the middle class erodes. 15 reports that seventy-seven percent of the variation in workers’ earnings stems from individual characteristics, primarily skills and education. “The wealth of workers,” he writes, “is in their skills.” The historical evidence is consistent. Steam power initially required only child operatives, but as machines grew complex, demand shifted to skilled engineers and machinists. During what 18 calls the “special century” (1870–1970), education supply kept ahead of demand, the mechanism 16 identify as the engine of the Great Compression and the golden age of shared prosperity. After 1980, the computer revolution began favoring skills that require higher education ( 15 ) . Demand outpaced supply, and the middle class eroded as a direct consequence. But the prescription remained the same: more school, more training, more degrees. 16 also document a detail that matters for what follows. Educational attainment in the United States plateaued for cohorts born after roughly 1950. The race was already tilting against education before the computer revolution arrived, which suggests the recovery mechanism was losing momentum well before AI added new pressure. There is an implicit assumption embedded in this entire analysis, so deeply embedded that economists rarely state it explicitly: no matter how disruptive the technology, education remains intact as the mechanism through which the next generation adapts. The factory replaces artisans; schools produce engineers. Computers replace clerks; universities produce programmers. The mechanism takes time, but it has worked for two centuries. The question this essay asks is: what if, for the first time, it doesn’t work at all? To answer it, the paper traces this question through three bodies of evidence. Part II establishes that AI is structurally different from every previous technology (it operates across the full cognitive hierarchy, blocking education’s traditional escape route) and introduces a stock-versus-flow framework that reconciles divergent economic and educational data. Part III asks whether learning actually occurs when students use AI, where its ceiling lies, and how we would know. Part IV proposes a diagnostic agenda targeting the gaps in the evidence base, extends the cognitive taxonomy to name capacities above AI’s current reach, and translates both into design principles. The Modern Data: AI Follows the Pattern (So Far) 22 ’s ( 22 ) Anthropic Economic Index provides the most detailed look to date at how AI is actually being used in the economy. At first glance, the data is reassuring. Across more than four million conversations analyzed, roughly fifty-seven percent of AI usage is augmentation rather than automation ( 22 ) : workers iterating, learning, and validating with AI assistance rather