The Most Valuable Skill in the Economy Is the One Nobody's Being Trained For

Goldman Sachs's economists have spent the past year documenting how AI is hollowing out the bottom of the labor market: young tech workers aged 20 to 30 are being displaced at four times the overall rate of job losses, and U.S. entry-level postings are down 35% since January 2023 ^[1]. Elsewhere in the same firm, a partner named Chris Churchman has been warning that AI integration risks what he calls cognitive atrophy — the erosion of first-principles thinking and tacit knowledge as financiers outsource their reasoning to models ^[2]. The two warnings have never been connected, not even inside Goldman. But they are two halves of one cycle, and the firm is watching both from the same building. The first half is the elimination. SignalFire's research found a 50% decline in new hires with less than a year of experience between 2019 and 2024, and the share of employees aged 21 to 25 at large tech firms fell from 15% to 6.8% ^[3]. IBM's chief HR officer, Nickle LaMoreaux, put the question plainly: what are organizations doing about training, and how are they redefining jobs, rather than simply counting how many positions can be cut ^[4]. The second half is what that eliminated work was actually for. The tedious entry-level tasks were never just production; they were the apprenticeship. GoTo's CEO, Rich Veldran, called foundational tasks a rite of passage that builds intuition and critical thinking ^[5]. And the institutions that would have to replace that apprenticeship are not ready: the CIPD found only 19% of organizations believe their leaders are equipped for AI integration, and only a third provide any employee training at all ^[6]. So the judgment deficit is not a prediction. John Nosta describes an AI rebound effect, in which polished outputs mask a regression of workers' baseline skills — the skill set falls below where it started ^[7]. A software consultant built an app without writing a line of code, then found he could no longer edit the 100,000-line codebase it produced.

The skill set actually falls below baseline. — John Nosta

Balyasny's Gappy Paleologo makes the same point from the top of the finance industry: stock-picking is a demanding cognitive function that cannot be replicated by a model whose experience of a company is mediated by layers of processing ^[8]. The costs are already concrete. Oregon's Court of Appeals is tracking the time its staff spends on AI-fabricated legal filings — $10,000 and $8,000 sanctions for fake citations, and a $110,000 federal penalty ^[9]. A UK business lost three days of website uptime and paid $485 to repair AI-generated code, while real-world productivity gains from AI chatbots average 2.8%, a fraction of what controlled experiments promised ^[10]. And 55% of companies that replaced workers with AI now say they moved too fast ^[11]. Here is the part no single story says. The people who would catch these errors — the ones who can tell a hallucinated citation from a real one, who can read the code the model wrote — are the same people whose training ground is being dismantled. The errors compound at the exact moment the capacity to verify them shrinks. That is not a coincidence of timing; it is one mechanism. Consulting firms are shifting to outcome-based pricing because the production work itself has been commoditized to near zero ^[12]. The capabilities that still command a premium are the ones AI cannot supply — the BS Detector that spots an AI error, the judgment that overrides an efficient-but-wrong recommendation ^[13]. The most valuable skill in the economy is the one the pipeline is no longer producing.