两篇最近研究（哈佛与史丹佛）提供较早的实证，显示在 AI 应用度高的职类，如 software development、customer service，22–25 岁及较 junior 的就业增速确有放缓迹象。理论上，资方认为年长员工累积了 AI 短期难以复制的 expertise 与 judgment；而青年初期更偏 book learning、标准化执行任务。多伦多大学模型亦称 AI 擅长制作简报、汇整资料、摘要文件等 implementation 工作，但仍缺乏新机会判断。作者同时警告，证据仍受利率与关税等混杂因素影响，且 AI 能力持续变动，未必能稳定落在「资深/初级」二分轴上。
文章更偏向解读为暂时的「expertise shock」而非永久替代。企业可能因 AI 是否真正提升生产力与节省成本未明，而先暂停招聘；在这种低进、低退的市场里，受冲击最重的是入门层。Avi Goldfarb 指出企业不知道未来需要何种技能，因此倾向缩招。但随著流程成型，模式可能逆转：OpenAI 正启动 AI 工作媒合平台，Shopify 与 Duolingo 甚至转向增聘实习生。资料上，18–25 岁使用者送出 46% 的 ChatGPT 讯息；Stack Overflow 2025 调查中，认为 AI 威胁就业者为 18–24 岁 15%，55–64 岁 13%。David Marchick 也表示「50%消失」是高估，Kogod 目前约 90% 教师将 AI 纳入教学，并提高软实力必修课比重。
The article says the “career-ladder collapse” narrative for 22- to 25-year-olds has become widespread, but not an economic law. LinkedIn co-founder Reid Hoffman argues AI’s overall aggregate labor impact remains limited so far; Anthropic’s Peter McCrory similarly thinks effects are still unclear. Harvard’s David Deming notes that past technological revolutions usually favor younger, highly educated workers. At the same time, the U.S. unemployment level remains high, near the highest since 2021, while Yale Budget Lab director Martha Gimbel argues the timing of labor-market turbulence predates AI’s rapid expansion.
Two recent studies from Harvard and Stanford provide early evidence that in AI-exposed occupations, such as software development and customer service, job growth appears slower for younger and junior cohorts, including ages 22–25. One mechanism is that older workers have experience-based expertise and judgment that is hard to automate, while younger workers often start with routine tasks and book learning. A University of Toronto model similarly says AI is strong in implementation tasks—building slide decks, compiling data, and summarizing documents—but still weak at opportunity judgment. The article warns that findings remain incomplete, with confounders like interest rates and tariffs, and that AI capability is still evolving, so a fixed seniority split may not hold.
A more likely interpretation is temporary “expertise shock,” not permanent displacement. Firms may pause hiring while waiting to see whether AI raises productivity or lowers labor costs, and this uncertainty is costly for entrants in a low-hiring, low-firing labor cycle. Avi Goldfarb says firms reduce hiring when future skill needs are unclear. But the pattern may reverse as processes settle and roles evolve: OpenAI is piloting an AI job-matching platform, and firms such as Shopify and Duolingo are reportedly hiring more interns. Data are mixed: users aged 18–25 send 46% of ChatGPT messages, while Stack Overflow 2025 shows only 15% of age 18–24 developers and 13% of age 55–64 developers view AI as a job threat. David Marchick says “50% evaporation” is exaggerated; at Kogod, about 90% of faculty now include AI in teaching and soft-skill requirements have been expanded.