人口红利已经转成了人口负担

不是变老。甚至也不是过时。沉到水下的，是文凭。

一张装裱好的文凭正面朝上，沉在浅浅的静水池底；俯视视角下，水面折射的光纹掠过文凭表面。

不是变老。甚至也不是过时。#

不是变老。甚至也不是过时。沉到水下的，是文凭。

这篇文章讲的不是抚养比，也不是退休时间表。那是一种会计式框架：老人太多，年轻人太少，前者压在后者身上。那个问题确实存在；但本文说的不是它。本文说的也不是劳动者逐渐被淘汰。那是一种更容易让人安心的叙事：技能一点点失效，工资慢慢被压低，受影响的人还有十年时间看见变化，然后转向。那种故事里的退化是可计量的，它沿着一条缓慢下行的曲线移动。本文要看的这条曲线，不是缓慢下行。它是断崖式下坠。

人口红利原本就是一笔加了杠杆的赌注。赌法很简单：劳动年龄人口增长快于被抚养人口，同时充裕的农业劳动力把制造业工资压在边际产出之下。二者合在一起，形成了一套格局。它不是年龄结构本身的性质，而是一种特定的生产格局。一代制度围绕这个假设修起了文凭基础设施：这套格局会延续得足够久，足以让那笔投资完成摊销。文凭是贷款。格局的延续是资产。贷款是真实的，还款还在继续；而当 AI 提前叫停这笔押注时，底层资产的市值已经跌到未偿余额以下。

贷款是真实的。还款还在继续。资产价值低于欠款余额。没有一个机制能让它按面值清算，也没有办法像这个类比暗示的那样，干干净净地退出文凭投资。当然，这个类比有边界，后文会把边界说清楚。

这里讨论的是资产负债表，不是损益表。损益表会显示：劳动者仍在就业，工资仍然到账，文凭还装裱在墙上。资产负债表显示的是另一件事。同一个劳动力市场里，同时存在两个受损位置。看得见的那个在职业阶梯底部：22 岁的人，原本该进入的初级岗位已经不存在了。看不见的那个在中段：35 到 55 岁的人仍有工作，但压在这份工作之上的索取权，正在以超过资产承载能力的速度累积。

先从看得见的地方开始。

看得见的损伤在 22 岁#

斯坦福数字经济实验室的 Brynjolfsson、Chandar 和 Chen 追踪了覆盖数百万美国劳动者的 ADP 薪资微观数据，这是美国规模最大的薪资数据集。他们发现了一个被总体就业统计遮住的、按群体分层的结果。1 具体到 22 至 25 岁的软件开发者，就业人数从 2022 年末的峰值到 2025 年 7 月大约下降了 20%。在所有高 AI 暴露职业中，更宽泛的 22 至 25 岁群体相对趋势下降了 13% 到 16%。机制是一种招聘冻结效应：AI 接过了那些新人原本被雇来完成的任务负荷，于是公司停止把新人招进来。文凭仍然带来身份地位。可那个文凭本来要通向的岗位，在这一代人抵达之前就已经被 AI 自动化了。

行业层面的岗位发布数据把这个图景继续展开。根据 REZI 在 2025 年汇总的行业报告，并引述 IntuitionLabs 与 HR Katha 对雇主岗位发布数据的分析：美国入门级技术岗位发布量在 2023 到 2024 年间下降了 67%；英国技术类毕业生岗位在 2024 年下降了 46%；同一时期，入门级金融岗位下降了 24 个百分点。2 这些数字是说明性的，因为数据链经过了中间分析，而不是直接来自 NACE、LinkedIn 或 Indeed；但它们和斯坦福 ADP 的发现方向一致，也捕捉到同一个时刻：ChatGPT 的采用时间线与入门岗位崩塌的时间线，重合得令人不安。

大型律所从一个高度依赖文凭的职业内部，给出了最刺眼的诊断。AI 工具正在自动化文件审阅、尽职调查和基础法律研究，而这些正是初级律师的传统训练工作。律所开始减少初级律师招聘；与此同时，正如斯坦福法学院教授 David Freeman Engstrom 在 Axios 2026 年 5 月的一篇报道中所说，它们正在赶在律师退休之前，“抢着提取律师的知识”。3 “抢着提取”这个说法，是专业群体在意识到生产下一代专家判断的基底已经被拆掉之后才会说出来的话。文件审阅不只是就业。它曾经是助理律师成长为合伙人的机制。

22 岁的人毕业时拿到了文凭。文凭证明他有资格得到入门工作。入门工作本身就是训练。现在，训练消失了。

可是，斯坦福 ADP 数据里还有第二个发现。它得到的关注远少于入门岗位崩塌：在同样的高 AI 暴露职业中，35 到 49 岁劳动者的就业没有出现有意义的变化。《财富》2026 年 4 月对斯坦福后续数据的报道还发现，在最高 AI 暴露类别中，30 岁及以上劳动者的就业从 2022 年末到 2025 年 5 月实际增长了 6% 到 12%。1 职业阶梯的底部被抽掉了。中段却还在。人口红利叙事原本预测，增长中的劳动年龄人口会成为增长的结构性引擎；但现在，劳动年龄人口的中段仍有工作，正在试图进入中段的人却没有。某个东西脱钩了。

要理解脱钩的是什么，得回到人口红利本来是什么。

红利是一套格局，不是一种结构#

人口红利从来都不是人口年龄本身的属性。

Bloom 和 Williamson 在 1998 年确认，1965 到 1990 年间，东亚和东南亚人均收入增长的三分之一到二分之一，可以归因于人口红利。4 机制是：劳动年龄群体比被抚养人口增长更快，由此产生的比例变化，暂时扩大了相对于消费的生产能力。这是发展经济学报道中流通最广的人口红利版本。就它说到的范围而言，它是真的。

它漏掉的是第二个条件。红利之所以可兑现，并不只是因为劳动年龄人口占比有利。它可兑现，是因为农业部门的剩余劳动力让制造业工资长期低于边际产出。蔡昉 2010 年的论文明确建立了这一点：只有当剩余劳动力仍然存在、乡村通向工厂的迁移管道仍然运转、工业工资成本仍被锚定在生产率以下时，红利才存在。5 关键不是劳动年龄群体有多大，而是它能以什么价格被调用。

当剩余劳动力耗尽，也就是当农村到城市的迁移因为剩余已经被吸收而放慢时，制造业工资会上升到边际生产率附近，成本优势消失，红利机制随之结束。IMF 的 Das 和 N'Diaye 在 2013 年把这个过程形式化，并预测中国的刘易斯拐点会在 2020 到 2025 年间到来。6 蔡昉在 2010 年写作时已经认为它正在发生。红利的终点从来都是结构性的：它不是日历上的某一天，而是一个条件，即廉价劳动力格局耗尽廉价劳动力的那一刻。

AI 比人口转型本身更快终结了这套格局。红利时代修建起来的文凭基础设施，包括十二年的专业训练轨道、从初级到高级的机构金字塔、被写进这些文凭里的工资预期，都建立在一个假设之上：劳动力稀缺会持续得足够久，让投资可以被摊销。AI 没有改变人口。它改变的是人口生产性产出在边际上值多少钱。曾经给人口红利提供正当性的格局结束了。为了服务那套格局而建的基础设施，还没有调整。

文凭先从顶端溶解，而不是从底部#

过去的自动化，是慢慢沿着技能阶梯往上走，替代中段和底部的例行任务，也就是那些人类相对可编程机器并没有多少比较优势的工作。

这是大多数读者随身携带的背景模型。用在 AI 上，它是错的，而且错在一个特定方向上：AI 首先暴露的是顶端。

Eloundou、Manning、Mishkin 和 Rock 2024 年在《Science》发表研究，发现美国约 80% 的劳动力至少有 10% 的工作任务会受到大语言模型影响，约 19% 的劳动者至少一半任务可能受到影响。7 真正打破旧模型的发现是：收入越高的职业，暴露程度越高。专业服务、法律、金融和知识工作这些类别，正是其市场价值曾经由文凭投资来证明的类别，也恰恰处于最暴露的行列。Felten、Raj 和 Seamans 用一套独立构建、并与 O*NET 工作能力相连的 AI 职业暴露指数，从另一种方法得出同样结果：高学历、高收入、白领职业得分最高。8

过去的自动化冲向流水线。AI 冲向的是专业人士的内室。

译者就是那只预警鸟。2025 年经由 CEPR VoxEU 发表的研究发现，在美国本地劳动力市场中，机器翻译采用率每提高 1 个百分点，译者就业增长大约降低 0.7 个百分点；累计看，它消除了约 28000 个本来会在 2010 到 2023 年间存在的译者岗位。9 英国作家协会 2024 年调查发现，36% 的英国译者表示因生成式 AI 失去工作，43% 表示收入下降。翻译是一条有文凭支撑的专业轨道。它被机器翻译替代，不是数据录入的故事，而是判断的故事：AI 匹配了文凭原本被设计来生产的输出。

Acemoglu 和 Restrepo 的任务框架，也就是评估自动化浪潮中替代与重置的正式装置，识别出了具体的失效模式。10 只有当重置效应，也就是创造出劳动仍有比较优势的新任务，超过替代效应时，自动化对劳动者才是净正面的。他们 2022 年发表在《Econometrica》的论文发现，过去四十年美国工资结构变化的 50% 到 70% 可归因于替代，而且近几十年重置效应已经变弱。历史上让自动化对劳动者整体为正的机制，依赖新任务仍然需要人类判断。AI 正在瞄准的，正是人类过去比较优势最强的任务层级。它在文凭投资最大的地方，反转了历史模式。

文凭的交换价值下降了。针对它的索取权没有下降。

数据看不见的那一群人#

然而，也正是在这里，可见记录开始不够用了：职业生涯建立在那套刚刚结束的格局之上的人，仍然在就业。

斯坦福数据显示 35 到 49 岁就业没有有意义变化，《财富》2026 年 4 月的数据又显示，高 AI 暴露类别中年长劳动者就业增长 6% 到 12%。这些并不是需要消解的矛盾。它们就是问题本身。可见记录展示的是损益表。损伤发生在资产负债表上。

现金流为正，累积索取权却已经超过资产价值。工资还在到账，养老金公式却已经锁定。文凭还挂在墙上，交换价值却已经跌破成本基础。

IMF 2024 年 1 月关于生成式 AI 与工作未来的 Staff Discussion Note，记录了定义 35 至 55 岁群体处境的两个结构性条件。11 第一，曾在高暴露、高互补职业中就业的年长劳动者，一旦被替代，比壮年劳动者更不容易在同一职业类别中重新就业。第二，IMF 这份讨论说明的分析意味着，对于职业剩余期限较短的劳动者，企业可能并不认为投入再培训是划算的。于是形成了一种结构性市场条件：暴露最高的人，反而最不可能获得适应性投资。这不是个人动机失败，而是市场失败。在一个格局已经改变的领域里，再培训一名 50 岁劳动者的 ROI 算不过来，公司会理性地按这个计算行事。

Pizzinelli 和 Tavares 在 2025 年沃顿养老金研究委员会论文中，精确记录了定义这一群体的结构性悖论。12 与年轻同侪相比，更大比例的年长劳动者已经处在高 AI 互补职业中，也就是 AI 增强而非替代、累积判断确实有价值的岗位。他们的位置很好。与此同时，一旦被替代，他们又更不容易完成成功的职业转换，原因包括岗位流动性更低、转换成本更高，以及 IMF 记录的那种理性低投资模式。Pizzinelli 和 Tavares 报告称，年长劳动者的培训参与率显著低于年轻同侪；论文引用的全文数据表明，2023 年 60 至 65 岁成年人和 25 至 44 岁成年人之间存在明显差距。位置优势和适应能力脱钩了。这个群体被最优地安置在一套刚刚改变的格局里。

22 至 25 岁的损害，和 35 至 55 岁的损害，在类型上不同，而不是沿同一条轴线分布的位置不同。前者是损益表崩塌：入门岗位被抹掉，立刻出现在 ADP 数据和岗位发布数据库里。后者是资产负债表崩塌：索取权不断累积，压向一个价值已经下跌的资产，在索取权到期前不可见。二者同时存在。第二种更危险，因为到目前为止，公共讨论里唯一读得出来的文件一直是损益表。

要说明这个位置为什么是负的，而不只是变差了，必须精确命名。

成本基础、固定索取权、恶化的价值#

在资产负债表上，索取权超过资产价值的那一行有一个名字。这就是 “underwater” 的意思：资不抵债，沉在水下。

有四种机制可能把人力资本推入负区间。其中三种不会。一种会。

沉没成本锁定。 十二年的文凭化是一项投资，投入的时间无法追回。但沉没成本一旦沉没，就归零，不会低于零。时间已经过去；这个损失真实存在，也被时间本身限定。这个机制制造的是后悔和失去的替代选择，而不是负区间。它是背景，不是本文的主张。

反模式训练。 Erik Dane 2010 年在《Academy of Management Review》发表的论文建立了这个机制：随着领域专长加深，认知图式会变得更大、关联更密、更稳定。13 这种稳定性让专家在一个范式内高效。范式转换时，它会变成负债：根深蒂固的图式会把新信息导回旧框架。Brynjolfsson、Li 和 Raymond 2025 年发表在《Quarterly Journal of Economics》的随机现场实验，研究了一家《财富》500 强公司的 5172 名客服人员，发现，在那个呼叫中心场景中，新手和低技能劳动者在 AI 辅助下提升了 34%，而经验最丰富的劳动者速度小幅提升、质量小幅下降。14 机制是：AI 编码并分发专家最佳实践，让资深劳动者累积起来的优势不再那么具有区分度；而在某些狭窄的组织场景里，如果劳动者有权覆盖 AI 输出，这种经验还可能成为质量拖累。这是真实的。但它只在特定组织情境中真实，不是普遍机制。它是次要的。

组织拖累。 公司保留一名产出可以被 AI 匹配的资深劳动者，就是在为一种它不再需要以这个成本购买的东西支付溢价。这是昂贵的零价值位置，不是低于零的位置。公司承担成本；成本真实存在；但单独看，它不会让劳动者的人力资本变成负值，只会让这名劳动者相对于替代方案变得昂贵。这也是背景。

无法匹配的财政索取权。 这是唯一能承载字面负值主张的机制。

养老金是在旧格局的生产率假设下承诺的。工资是按旧工资溢价谈成的。医疗福利由合同固定。这些索取权以美元计价，有明确的合同结构，面对的却是一个已经下降的生产价值。缺口真实存在，原则上可以计算，并且会随时间复利。

养老金是合同。工资是合同。文凭投资已经摊销。只有生产价值，也就是方程里唯一从未被合同固定的那一项，可以自由下跌。

Korinek 和 Lockwood 2026 年 1 月的布鲁金斯学会论文建立了宏观经济框架：AI 可能逐步侵蚀支撑现代社会安全网的两大税基，即劳动收入和人类消费，并在最需要这些资金的时候制造财政压力。15 35 至 55 岁群体正好位于交叉点上。2025 至 2045 年间，他们的财政索取权会不断累积，而为这些索取权提供资金的生产性基础正在压缩。养老金基金和 AI 生产率曲线运行在两本不同账簿上。它们最终会在基金没有预提准备的某一刻对账。

这就是会计意义上的“负值人力资本”：不是说人有负的内在价值，而是说某个特定群体的会计位置已经变号，也就是养老金、医疗和工资预期，对上 AI 价格点下的生产性产出，已经从正翻成了负。标准人力资本理论在形式上不允许负值；本文的主张说的是一个会计恒等式，不是一个理论属性。这个区别重要。正因为主张更窄，它才更可辩护。

这个隐喻能承载什么，不能承载什么#

房贷类比能承载其中一部分，但不是全部。

它能承载的是：成本基础与当前市场价值之间的结构关系；合同索取权会独立于资产贬值继续存在这一点；以及信号价值和交换价值可以脱钩。墙上的文凭仍有信号价值，可在一个已经重新定价产出的市场里，那张面值完整的文凭并不值它表面显示的价值。

它不能承载的是：

不是数值精确。房贷有一个固定的美元缺口，贷款余额比估价高出一个具体数字。文凭的“贷款金额”是一种模糊的终身收入预期，它曾经按一套生产格局定价，而那套格局后来改变了。会计恒等式是结构性的，不是算术。缺口真实存在，但不能压缩成一个单一数字。

不是字面锁死。劳动者会再培训，会转向。职业转换会发生，再培训会发生，中年重塑也会发生。“无法离场”这个说法，是对结构性困难的修辞表达，不是对法律约束的描述。劳动者确实会离开。他们会以清仓价离开，而这正是这个类比要命名的东西。映射的是清仓出售，不是政府救助。

也不是暗示救助。2008 年房贷危机之后有 HAMP、HARP 和本金削减。本文没有提出任何等价的政策主张。这里命名的是会计恒等式，不是在为立法开处方。文凭持有人认识到自己的位置已经沉到水下之后该怎么办，是另一个问题，而且更难。

隐喻是框架。分析是上一节那个不等式。框架帮助读者抓住不等式；不等式本身不依赖框架也成立。

这个分析主张必须经受最强反驳，而最强反驳是真实的、经过校准的、有经验证据支撑的。

最强反驳，以及它止步的地方#

最强反驳不是一种修辞姿态。它是一个真实的、经过校准的、经过经验估计的模型，并且有记录在案的预测成功。本文要守住自己的论线，就必须先正面面对它。

让步一：Bessen 论需求弹性。 Bessen 2018 年的 NBER 框架确立了一点：自动化会扩大还是收缩就业，取决于被自动化服务的需求弹性。16 ATM 案例很干净：ATM 部署降低了单笔交易成本，让银行能够运营更多网点，于是柜员总需求反而上升。在 ATM 扩散期间，柜员就业增加了。这个模型已经被延伸到 AI：法律服务存在结构性需求配给，因为司法可及性缺口意味着当前需求被价格压住；AI 驱动的降本应该释放潜在需求，甚至可能扩大整体法律就业。这是真的。承认它。

让步二：Acemoglu 和 Restrepo 的重置效应。 任务框架是双向的。近期扩展专门估计了 AI 的 50% 到 80% 重置区间。PwC 2025 年《Global AI Jobs Barometer》发现，AI 暴露行业的生产率增长翻了两番（2018 到 2024 年，从 7% 到 27%）；AI 暴露行业的工资增长约为低暴露行业的两倍；即便整体岗位发布下降 11.3%，要求 AI 技能的岗位仍增长 7.5%；具备 AI 技能的劳动者获得 56% 的工资溢价。10 这就是任务框架预测的重置效应。它真实，而且可测。承认它。

让步三：Bessen 等人的再吸收。 Bessen 及其同事发表在《Review of Economics and Statistics》的经验后续研究显示，在该研究的五年测量窗口内，被自动化替代的劳动者累计工资损失约为一年工资收入的 8% 到 9%。17 总量调整叙事有经验证据支撑。这个也承认。

现在是反驳。它不是作为经验发现提出，而是作为会计恒等式提出：

需求弹性在新的均衡价格上扩张就业。35 至 55 岁劳动者的工资是在旧格局价格上设定的。工资是合同性的，弹性是总量性的；二者落在不同账簿上。

Bessen 的模型描述的是新劳动投入如何按新均衡价格出清。它并不说明附着在现有劳动者身上的遗留成本结构。ATM/柜员类比之所以成立，是因为柜员可以按新的单网点经济性被雇用；既有柜员的薪水不是障碍。2026 年的文凭层劳动者，不能突然以 AI 赋能后的价格点提供服务。内部成本结构，也就是旧格局中谈成的累积工资、已经按旧公式归属的养老金，阻止了这件事。障碍不是劳动者的能力。障碍是劳动者的成本基础。

56% 的 AI 技能溢价，流向的是获得 AI 互补性的劳动者。锁定论证讨论的是另一类劳动者：他们面对结构性障碍，难以获得这种互补性。这些障碍包括 Dane 记录的认知固化机制，以及 IMF SDN 记录的企业低投资机制。溢价和锁定，描述的是同一个劳动力市场中的不同劳动者子群体。用溢价来反驳锁定，是把两者混为一谈。

时间论证是最锋利的部分。Bessen 等人的五年测量窗口，也就是累计工资收入损失发生的时期，对于一名 2026 年被替代的 50 岁劳动者来说，恰好与多数养老金公式下财政索取权累积的峰值年份重合。2026 到 2031 年之间损失的工资收入，不会在测量窗口结束后凭空消失：那些缺失年份本来会锁进养老金公式，最终会在劳动者余生的退休期里生成更低的待遇。工资损失是真的。养老金缺口也是真的。它们在不同账簿上对账。

相关的历史先例不是“这一次不同”。而是：“模式相同。” Krueger 1993 年的研究发现，在 1980 年代计算机化浪潮中，计算机使用者有 10% 到 15% 的工资溢价18。重置效应真实存在，它惠及能够部署新技术的劳动者，同时也以工资分布的永久位移为代价；被替代的文职劳动者在自己的一生中并没有从中受益。模式是：总量劳动力市场用二十年调整；在转型中被替代的那一代人，无法在自己的世代里恢复。这不是说 AI 与所有历史自动化都发生了结构性断裂。它是 1980 年代已经产生过的同一种世代不对称，只是发生在不同层级，具有不同的可见性。

耶鲁预算实验室基于 CPS 持续追踪 AI 对美国劳动力市场的影响，它给出的图景比斯坦福按职业划分的发现更谨慎。19 它的分析认为，总体 AI 就业市场影响表现为稳定，也就是在经济整体就业统计中没有重大扰动。这不是矛盾。总量稳定完全可以和特定“群体 + 职业”交叉点上的集中结构性损害同时存在。本文的主张说的是那种集中，而不是宏观标题数字是否令人警觉。

反驳仍然站得住。第二幕的故事还在等待：它会把眼下可见的东西，与学徒培养基底枯竭之后才会变得可见的东西连接起来。

生产专家判断的工厂关门了#

回到 22 岁的人。

斯坦福 ADP 数据记录、REZI 行业汇总相互印证的入门岗位清空，不只是一代人失去第一份工作。它是学徒培养基底正在被经济性地擦除。专业金字塔底部的初级岗位，包括审阅文件的助理律师、阅读常规影像的初级放射科医生、制作演示文稿的初级咨询顾问、处理样板文本的初级译者，并不只是就业机会。它们是下一代资深专家判断被制造出来的机制。文件审阅是训练。样板文本是训练。常规影像是训练。每一级台阶都有经济价值；正是这个价值，证明了雇用那个最终会爬过这一级台阶的人是合理的。

Axios 在 2026 年 5 月报道了大型律所面对这一局面时正在做什么：赶在退休之前，赶在那些通过这条管道建立起专业能力的资深合伙人离开之前，抢着提取律师的知识。3 当一个职业意识到生产专家判断的基底已经被拆除时，它就会这样做：它试图在持有知识的人消失之前捕捉现有知识，因为已经没有流程能生产下一代持有人。

这个结构性推断是诚实的，不是经验性的。如果初级岗位在 AI 条件下已经不合经济理性，因为 AI 做训练层级工作的成本低于初级律师，那么这条管道就不会在市场激励下重建。过去，雇用初级律师的市场理由是她的经济产出；现在这个理由不再成立，而她原本会从中获得的训练价值，从来不是支付她薪水的主要理由。2035 到 2040 年的知识缺口，也就是当 AI 在边界案例上犯错、而人类训练出来的判断已经不存在时出现的缺口，并不是有经验证据背书的预测。它是从当下经济性推出的结构性推断，因此应该按这个身份持有它：诚实，但不确定。

第一代人的职业路径被关上。下一代人没有专家判断可以继承。等文明试图调用那种已经不在的东西时，它才会发现缺口。

22 岁处的可见损伤，不是一个会在经济调整后自行纠正的入门级扰动。它拆掉的是专家判断跨世代复制自身的机制。两个位置的结构现在已经完全显现：职业阶梯底部的损伤，可测量、已记录、正在加速；中段的损伤，在资产负债表上累积，到索取权到期前不可见；未来时态中的损伤，是推断性的、结构性的，尚未出现在任何数据集里，因为正在被摧毁的是一条管道，而管道总是先在远端干涸。

资产负债表会显示什么#

资产负债表不会撒谎。它只是延迟。

财政索取权会到期。 养老金基金会在 2030 到 2050 年间遇到缺口。届时，35 至 55 岁群体开始领取福利，而这些福利抵押在一个已经被压缩的生产性基础之上。这不是关于劳动力市场的预测，而是关于算术的观察：养老金公式、摊销周期、在一套已经不存在的格局中谈成的福利合同结构。缺口累积时，现金流是正的。缺口显现时，现金流会变成负的。从累积到显现之间的延迟，正是损益表看起来还不错的那段时间。

魏玛手推车。 当固定名义索取权遇上坍塌的生产价值，历史模式就是印钞：延续名义索取权，同时让底层生产基础进一步下滑。对应到政策层面，就是文凭通胀：更多再培训项目，更多证书，更多 AI 技能徽章，更多文凭层级，被堆在一个已经重新定价底层产出的市场之上。手推车这个意象，指的是一个机构试图通过发行更多索取权，来兑现面对恶化基础的固定索取权时会发生什么。交换系统已经坏掉时，更多文凭不会恢复交换价值；它只会加速贬值，因为曾经赋予文凭市场价值的稀缺信号已经消失。

把会计恒等式完整摆出来时，它显示的是： 红利是一套格局，不是一种结构。文凭是相框，不是资产。资产负债表会显示损益表正在遮住的东西。

那个缓慢、不可见、由合同累积起来的负担，不在抚养比里。它在一个生产层级曾被承诺的东西，和 AI 时代经济性愿意支付的东西之间。这个缺口不在损益表上。它从来就不在那里。

资产负债表不会出现在季度财报里。它会在索取权到期时出现，出现在一本从未被设计来读懂这场已经完成的格局变化的账簿上。

参考文献#

  1. Brynjolfsson, E., Chandar, B., and Chen, R. (2025). "Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence." Stanford Digital Economy Lab Working Paper. ↩ ↩

  2. REZI AI (2025). "The Crisis of Entry-Level Labor in the Age of AI (2024-2026)." rezi.ai, citing IntuitionLabs and HR Katha analyses of employer job-posting data. ↩

  3. Axios (2026-05-02). "AI Threatens Big Law's Talent Pipeline." Also: Above the Law (2026-03). "AI Won't Replace Lawyers But Can Create Critical Shortage Of Good Ones." ↩ ↩

  4. Bloom, D., and Williamson, J. (1998). "Demographic Transitions and Economic Miracles in Emerging Asia." World Bank Economic Review, 12(3), 419-455. DOI: 10.1093/wber/12.3.419. ↩

  5. Cai, F. (2010). "Demographic Transition, Demographic Dividend, and Lewis Turning Point in China." China Economic Journal, 3(2), 107-119. DOI: 10.1080/17538963.2010.511899. ↩

  6. Das, M., and N'Diaye, P. (2013). "Chronicle of a Decline Foretold: Has China Reached the Lewis Turning Point?" IMF Working Paper WP/13/26. SSRN: 2222490. ↩

  7. Eloundou, T., Manning, S., Mishkin, P., and Rock, D. (2024). "GPTs are GPTs: Labor Market Impact Potential of Large Language Models." Science, 384, 1306-1308. arXiv preprint 2303.10130. ↩

  8. Felten, E., Raj, M., and Seamans, R. (2023). "Occupational Heterogeneity in Exposure to Generative AI." SSRN Working Paper 4414065. ↩

  9. Frey, C.B., and Llanos-Paredes, A. (2025). "How Machine Translation Has Displaced Translators." CEPR VoxEU, March 2025. Supporting data: Society of Authors (UK) Survey 2024. ↩

  10. Acemoglu, D., and Restrepo, P. (2019). "Automation and New Tasks: How Technology Displaces and Reinstates Labor." Journal of Economic Perspectives, 33(2), 3-30. DOI: 10.1257/jep.33.2.3. And (2022). "Tasks, Automation, and the Rise in U.S. Wage Inequality." Econometrica, 90(5), 1973-2016. DOI: 10.3982/ECTA19815. PwC (2025). "The Fearless Future: 2025 Global AI Jobs Barometer." ↩ ↩

  11. Cazzaniga, M., Jaumotte, F., Li, L., Melina, G., Panton, A., Pizzinelli, C., Rockall, E., and Tavares, M. (2024). "Gen-AI: Artificial Intelligence and the Future of Work." IMF Staff Discussion Note SDN/2024/001. ↩

  12. Pizzinelli, C., and Tavares, M. (2025). "AI and the Future of Work in an Aging Economy." Wharton Pension Research Council Working Paper WP2025-14. SSRN: 5345347. ↩

  13. Dane, E. (2010). "Reconsidering the Trade-off Between Expertise and Flexibility: A Cognitive Entrenchment Perspective." Academy of Management Review, 35(4), 579-603. DOI: 10.5465/amr.35.4.zok579. ↩

  14. Brynjolfsson, E., Li, D., and Raymond, L. (2025). "Generative AI at Work." Quarterly Journal of Economics, 140(2), 889-942. NBER Working Paper 31161. ↩

  15. Korinek, A., and Lockwood, B. (2026). "Public Finance in the Age of AI: A Primer." Brookings Institution Working Paper. January 2026. Also: NBER Working Paper 34873. ↩

  16. Bessen, J. (2018). "AI and Jobs: The Role of Demand." NBER Working Paper 24235. ↩

  17. Bessen, J., Goos, M., Salomons, A., and van den Berge, W. (2025). "What Happens to Workers at Firms that Automate?" Review of Economics and Statistics, 107(1), 125-141. ↩

  18. Krueger, A. B. (1993). "How Computers Have Changed the Wage Structure: Evidence from Microdata, 1984-1989." Quarterly Journal of Economics, 108(1), 33-60. ↩

  19. Yale Budget Lab. (2025-2026). "Tracking the Impact of AI on the Labor Market." The Budget Lab at Yale. Multiple updates 2024-2026, including February 2026 update. ↩

延伸阅读#

  • Bessen, J., Goos, M., Salomons, A., and van den Berge, W. (2025). "What Happens to Workers at Firms that Automate?" Review of Economics and Statistics, 107(1), 125-141. 这是 Bessen 2018 的经验后续研究，记录了自动化企业中的劳动者在五年测量窗口内累计工资损失约为一年工资收入的 8% 到 9%。要理解第八节的时间论证，以及为什么总量工资恢复不能拯救被锁定的 35 至 55 岁群体的养老金公式，这篇是关键背景。
  • Autor, D. (2024). "Applying AI to Rebuild Middle Class Jobs." NBER Working Paper 32140. Autor 认为，AI 可以把专家判断民主化，让中等技能劳动者受益。这是本文分析必须与之共存的乐观互补论证。同一个民主化机制，在本文中被描述为管道破坏；区别在于观察它的规范角度不同。
  • Acemoglu, D. (2024). "The Simple Macroeconomics of AI." NBER Working Paper 32484. Acemoglu 自己评估了为什么 AI 的宏观生产率收益可能小于宣传口径：如果自动化集中在并不与广泛人类劳动互补的任务上，收益就会受限。它直接关系到理解第四节和第八节引用的重置效应变弱问题。
  • Cowen, T. (2013). Average Is Over. Dutton. 这是文凭层论证的预示性框架：在一个世界里，能高度熟练地与复杂工具协作的人，其回报会和不借助这些工具工作的人急剧分化。它提供了 35 至 55 岁群体资产负债表论证所在的分配背景。
  • Goldin, C. (2014). "A Grand Gender Convergence: Its Last Chapter." American Economic Review, 104(4), 1091-1119. 讨论工资谈判的时间结构、职业路径锁定，以及为连续职业轨迹设计的薪酬结构，如何在轨迹中断时制造系统性惩罚。它是理解第六节养老金公式机制的重要背景。

简而言之#

不是变老。甚至也不是过时。沉到水下的，是文凭。

人口红利从来都不是人口年龄的属性。它是一套格局：劳动年龄群体增长快于被抚养人口，同时农村剩余劳动力把制造业工资压在边际产出以下。Bloom 和 Williamson 曾把东亚 1965 至 1990 年人均增长最高一半归因于这个配置；蔡昉和 IMF 后来确认，它会随人口时钟到期。AI 让它更早到期。一整代文凭基础设施，原本被设计为在这套格局下完成摊销，而这套格局已经提前结束。

同一个劳动力市场里，现在有两个受损位置。看得见的那个位于职业阶梯底部。斯坦福数字经济实验室的 Brynjolfsson、Chandar 和 Chen 追踪 ADP 薪资微观数据，发现 22 至 25 岁软件开发者的就业从 2022 年末到 2025 年中大约下降 20%；在高 AI 暴露职业中，同龄群体相对趋势下降 13% 至 16%。行业汇总岗位数据也朝同一方向移动：美国入门级技术岗位下降 67%，英国技术类毕业生岗位下降 46%。大型律所正赶在资深合伙人退休前提取他们的知识，因为过去训练下一代的文件审阅，现在由 AI 完成。文凭仍然赋予身份。可它本来要通向的岗位，在持有人到达之前已经被自动化。

看不见的损伤在中段。同样高暴露类别里的 35 至 55 岁劳动者没有出现就业下降；有些群体甚至就业增长。这不是对入门岗位清空的反驳，而是第二个受损位置。损益表仍然为正：工资到账，文凭挂墙。资产负债表却是另一回事。养老金是在旧格局的生产率假设下承诺的。工资是按旧工资溢价谈成的。医疗福利由合同固定。这些索取权会复利累积；它们所抵押的生产价值已经下降。Eloundou 等人在《Science》中记录，AI 暴露最高的是高收入白领职业，也就是那些文凭投资原本最需要证明其市场价值的类别。过去的自动化冲向流水线。AI 冲向专业人士的内室。

最强反驳，包括 Bessen 的需求弹性、Acemoglu 和 Restrepo 的重置效应、AI 技能工资溢价，描述的是新劳动投入如何在新均衡价格上出清。它没有回答现有劳动者背后的遗留成本结构。障碍不是能力，而是成本基础。生产专家判断的学徒培养基底也已经被经济性地擦除：在 AI 经济性下，初级岗位已经不合算，这意味着管道不会在市场激励下重建。

资产负债表不会撒谎。它只是延迟。财政索取权会在 2030 到 2050 年间到期，落在一本从未被设计来读懂这场已经完成的格局变化的账簿上。

The Demographic Dividend Has Become a Demographic Burden

A framed diploma lying face-up on the bottom of a shallow pool of still water, seen from above, with light refracting across its surface.

Not aging. Not even obsolescence.#

Not aging. Not even obsolescence. The credential is underwater.

This is not a story about dependency ratios or retirement timelines — the accounting frame where too many old people weigh against too few young ones. That story is real; it is not this story. This is not a story about workers becoming obsolete — the comfortable narrative where skills erode gradually, wages compress slowly, and the affected worker has a decade to notice and pivot. The degradation in that story is measured. It tracks a declining curve. The curve examined here does not decline gradually. It drops.

The demographic dividend was a leveraged bet. The bet was simple: working-age populations growing faster than dependent populations, paired with abundant agricultural labor keeping manufacturing wages below their marginal product. That combination produced a regime — not a property of age structure, but a specific production regime. A generation of institutions built their credential infrastructure against the assumption that this regime would persist long enough to amortize the investment. The credential was the loan. The regime's continuation was the asset. The loan was real, the payments continued — and when AI called the bet early, the market value of the underlying asset dropped below the outstanding balance.

The loan is real. The payments continue. The asset is worth less than the balance owed. There is no mechanism to liquidate at par, no way to walk away from the credential investment as cleanly as the analogy suggests — though the analogy's limits are worth naming, and they will be named.

The argument is about a balance sheet, not an income statement. The income statement shows workers employed, wages arriving, credentials still framed on walls. The balance sheet shows something different. Two loci of damage exist in the same labor market simultaneously. The visible one is at the bottom of the career ladder — the twenty-two-year-old for whom the entry-level position no longer exists. The invisible one is in the middle — the thirty-five to fifty-five-year-old whose current employment is intact while the claims against that employment accumulate faster than the asset can service them.

Start with what is visible.

The visible damage is at twenty-two#

Brynjolfsson, Chandar, and Chen at Stanford's Digital Economy Lab tracked ADP payroll microdata covering millions of U.S. workers — the largest payroll dataset in the country — and found a cohort-stratified result that the aggregate employment statistics hide.1 For software developers specifically aged twenty-two to twenty-five, employment fell approximately twenty percent from its late 2022 peak to July 2025. Across all high-AI-exposure occupations, the broader twenty-two to twenty-five cohort declined thirteen to sixteen percent relative to trend. The mechanism is a hiring-freeze effect: firms stop bringing in new entrants as AI takes over the task load those entrants were hired to perform. The credential still confers status. The position it was built to fill has been AI-automated before the cohort arrived to fill it.

Industry-level postings data extends the picture. Per industry reports aggregated by REZI (2025), citing IntuitionLabs and HR Katha analyses of employer job-posting data: U.S. entry-level technology postings fell sixty-seven percent between 2023 and 2024; UK technology graduate roles fell forty-six percent in 2024; entry-level finance positions fell twenty-four percentage points over the same period.2 These numbers are illustrative — the data chain runs through intermediate analyses rather than directly from NACE, LinkedIn, or Indeed — but they move in the same direction as the Stanford ADP finding and capture the same moment: the ChatGPT adoption timeline matches the entry-level collapse timeline with uncomfortable precision.

Big Law offers the starkest diagnostic from inside a credentialed profession. As AI tools automate document review, due diligence, and basic legal research — the traditional training work of junior associates — law firms have begun reducing junior associate hiring at the same moment they are, as Stanford Law professor David Freeman Engstrom noted in an Axios May 2026 report, "racing to extract the knowledge of their lawyers" before retirement.3 That phrase — racing to extract — is what professionals say when they have recognized that the substrate producing the next generation of expert judgment has been dismantled. The document review was not just employment. It was the mechanism by which an associate became a partner.

The twenty-two-year-old graduated with a credential. The credential justified entry-level work. The entry-level work was training. The training is gone.

And yet the Stanford ADP data contains a second finding that receives far less attention than the entry-level collapse: workers aged thirty-five to forty-nine in the same high-AI-exposure occupations show no meaningful change in employment. Fortune's April 2026 coverage of follow-up Stanford data found workers aged thirty and over in the highest-AI-exposure categories actually saw employment grow six to twelve percent from late 2022 to May 2025.1 The bottom of the career ladder has been knocked out. The middle is intact. The dividend story predicted that a growing working-age population would be the structural engine of growth — but the working-age middle cohort is employed and the cohort trying to enter it is not. Something has decoupled.

To understand what decoupled, go back to what the dividend actually was.

The dividend was a regime, not a structure#

The demographic dividend was never a property of population age.

Bloom and Williamson established in 1998 that between one-third and one-half of East and Southeast Asia's per capita income growth during 1965 to 1990 was attributable to the demographic dividend.4 The mechanism: the working-age cohort grew faster than the dependent population, producing a ratio change that temporarily expanded productive capacity relative to consumption. This is the version of the dividend that circulates in development economics journalism. It is true as far as it goes.

What it omits is the second condition. The dividend was not available merely because the working-age ratio was favorable. It was available because surplus labor in the agricultural sector kept manufacturing wages below their marginal product. Cai Fang's 2010 paper established this explicitly: the dividend was available only as long as that surplus labor remained, keeping the migration pipeline from countryside to factory alive and keeping industrial wage costs anchored below productivity.5 Not the size of the working-age cohort — but the price at which it could be deployed.

When the surplus is exhausted — when rural-to-urban migration slows because the surplus has been absorbed — manufacturing wages rise to meet marginal productivity, the cost-advantage evaporates, and the dividend mechanism ends. Das and N'Diaye at the IMF formalized this in 2013, projecting that China's Lewis turning point would arrive between 2020 and 2025.6 Cai had already argued it was underway at the time of writing (2010). The dividend's endpoint was always structural — not a calendar event but a condition: the moment the cheap-labor regime ran out of cheap labor.

AI has ended the regime faster than the demographic transition would have. The credential infrastructure built during the dividend era — the twelve-year professional training tracks, the institutional pyramid from junior to senior, the wage expectations priced into those credentials — rested on the assumption that labor scarcity would persist long enough to amortize the investment. AI has not changed the population. It has changed what the population's productive output is worth at the margin. The regime that justified the dividend is over. The infrastructure built to service that regime has not adjusted.

The credential dissolves at the top, not the bottom#

Past automation moved up the skill ladder slowly, replacing routine tasks at the middle and bottom — the kind of work where humans had little comparative advantage over programmable machines.

This is the background model most readers carry. It is wrong for the AI case — wrong in a specific direction. AI is exposing the top first.

Eloundou, Manning, Mishkin, and Rock published in Science in 2024 the finding that approximately eighty percent of the U.S. workforce has at least ten percent of their work tasks affected by large language models, and that approximately nineteen percent of workers may see at least half their tasks impacted.7 The finding that breaks the prior model: higher-income occupations show the highest exposure. Professional services, legal, financial, and knowledge work — the categories whose market value was justified by the credential investment — are among the most exposed. Felten, Raj, and Seamans, using an independently constructed AI Occupational Exposure index tied to O*NET workplace abilities, reached the same result from a different methodology: highly-educated, highly-paid, white-collar occupations score highest.8

Past automation came for the assembly line. AI comes for the chambers.

The translator is the canary. Research published via CEPR VoxEU in 2025 found that each one percentage point increase in machine translation adoption across U.S. local labor markets lowered translator employment growth by approximately 0.7 percentage points — cumulatively eliminating an estimated 28,000 translator positions that would otherwise have existed over 2010 to 2023.9 A Society of Authors 2024 survey found that thirty-six percent of UK translators report lost work due to generative AI, and forty-three percent report that their income has decreased. Translation is a credentialed professional track. Its displacement by machine translation is not a data entry story. It is a judgment story: the AI matched the output the credential was built to produce.

Acemoglu and Restrepo's task framework — the formal apparatus for evaluating displacement versus reinstatement across automation waves — identifies the specific failure mode.10 Automation is net-positive for workers when the reinstatement effect (new tasks created where labor has comparative advantage) outpaces the displacement effect. Their 2022 Econometrica paper finds that fifty to seventy percent of U.S. wage structure changes over four decades are attributable to displacement — and that the reinstatement effect has weakened in recent decades. The historical mechanism that made automation net-positive for workers depended on new tasks requiring human judgment. AI is targeting the tier of tasks where humans previously had their strongest comparative advantage. This inverts the historical pattern at precisely the point where the credential investment was largest.

The credential's exchange value has fallen. The claim against it has not.

The cohort the data does not see#

And yet — and this is where the visible record stops being adequate — the cohort whose career was built on the regime that has just ended is still employed.

The Stanford data showing no meaningful change in thirty-five to forty-nine employment, the Fortune April 2026 data showing six to twelve percent employment growth for older workers in high-AI-exposure categories — these are not contradictions to resolve. They are the problem. The visible record is showing the income statement. The damage is on the balance sheet.

Cash flow positive — accumulated claims exceeding asset value. Wage still arriving — pension formula already locked. Credential still framed on the wall — exchange value already fallen below the cost basis.

The IMF's January 2024 Staff Discussion Note on generative AI and the future of work documented two structural conditions that define the thirty-five to fifty-five cohort's position.11 First, older workers previously employed in high-exposure, high-complementarity occupations are less likely to find re-employment in the same occupation category than prime-age workers following displacement. Second, the IMF SDN's analysis implies that firms may not find it beneficial to invest in retraining workers with a shorter career horizon — a structural market condition in which the workers most exposed are least likely to receive adaptation investment. This is not a failure of individual motivation. It is a market failure: the ROI on retraining a fifty-year-old in a regime-changed field does not compute, and firms rationally act on that calculation.

Pizzinelli and Tavares, in their 2025 Wharton Pension Research Council paper, document the structural paradox that defines the cohort precisely.12 A larger share of older workers than younger peers is already employed in high-AI-complementarity occupations — jobs where AI augments rather than replaces, where the accumulated judgment has genuine value. They are positionally well-placed. And simultaneously, they are less likely to make successful occupational transitions if displaced, due to lower job fluidity, higher switching costs, and the rational underinvestment pattern the IMF documents. Pizzinelli and Tavares report that older workers' training participation rates are markedly lower than their younger counterparts — the specific figures in the paper, which cite the full-text data, indicate a substantial gap between adults aged sixty to sixty-five and those aged twenty-five to forty-four in 2023. Position and adaptation capacity have decoupled. The cohort is optimally positioned for a regime that has just changed.

The harm at twenty-two to twenty-five and the harm at thirty-five to fifty-five are different in type, not distributed differently along the same axis. One is income-statement collapse — entry-level wipe, visible immediately in ADP data and job-posting databases. The other is balance-sheet collapse — claims accumulating against an asset whose value has fallen, invisible until the claims come due. Both exist simultaneously. The second is the dangerous one, because the income statement has been the only readable document in the discourse so far.

What makes the position negative rather than merely deteriorated requires naming precisely.

Cost basis, fixed claim, deteriorated value#

On the balance sheet, the line where claims exceed asset value has a name. It is what "underwater" means.

Four mechanisms could put human capital in negative territory. Three of them do not. One does.

Sunk-cost lock-in. Twelve years of credentialing is an investment whose time cannot be recovered. But the sunk cost, once sunk, collapses to zero — not below it. The time is gone; that loss is real and bounded by the time itself. This mechanism produces regret and foregone alternatives, not negative territory. Supporting context, not the claim.

Anti-pattern training. Erik Dane's 2010 paper in the Academy of Management Review established the mechanism: as domain expertise deepens, cognitive schemas become larger, more interrelated, more stable.13 That stability is what makes experts efficient within a paradigm. Under paradigm shift, it becomes a liability — entrenched schemas route new information through old frameworks. Brynjolfsson, Li, and Raymond's 2025 Quarterly Journal of Economics randomized field experiment with 5,172 customer support agents at a Fortune 500 firm found, in that call-center setting, that novice and low-skilled workers improved thirty-four percent with AI assistance while the most experienced workers showed small gains in speed with small declines in quality.14 The mechanism: the AI encodes and distributes expert best practices, making the accumulated advantage of the senior worker less differentiating and, in narrow organizational contexts where the worker has authority to override AI outputs, a source of quality drag. This is real. It is real in specific organizational contexts, not universally. It is secondary.

Organizational drag. A firm retaining a senior worker whose output can be AI-matched is paying a premium for something it no longer needs at that cost. This is an expensive zero-value position, not a below-zero one. The firm bears the cost; the cost is real; it does not, in isolation, make the worker's human capital negative — it makes it costly relative to the alternative. Supporting context.

Unmatched fiscal claims. This is the mechanism that carries the literal negative claim, and it is the only one that does.

The pension was promised at the old regime's productivity assumption. The salary was negotiated against the old wage premium. The healthcare benefits are contractually fixed. These claims are denominated in dollars, with explicit contractual structure, against a productive value that has fallen. The gap is real, calculable in principle, and compounds with time.

The pension is contractual. The salary is contractual. The credential investment is amortized. Only the productive value — the only term in the equation that was never contractual — is free to fall.

Korinek and Lockwood's January 2026 Brookings Institution paper establishes the macroeconomic frame: AI may gradually erode the two main tax bases — labor income and human consumption — that underpin modern social safety nets, producing fiscal strain at precisely the moment when that funding is most needed.15 The thirty-five to fifty-five cohort sits at the intersection. Their fiscal claims accrue across the 2025 to 2045 window while the productive base that funds those claims compresses. The pension fund and the AI productivity curve are running on different ledgers. They will reconcile at a moment the fund has not provisioned for.

This is what "negative human capital" means in the accounting sense: not that humans have negative inherent worth, but that the accounting position of a specific cohort — pension plus healthcare plus salary expectations against productive output at AI-enabled price points — has flipped sign. Standard human capital theory does not allow negative values as a formal matter; the claim here is about an accounting identity, not a theoretical property. The distinction matters. The claim is narrower and more defensible for being narrower.

The metaphor and what it does not carry#

The mortgage analogy carries some of this and not all of it.

What it carries: the structural relationship between cost basis and current market value. The intuition that contractual claims survive asset depreciation independently. The understanding that signal value (the credential on the wall) and exchange value (what the credential commands when it goes looking for work) can decouple — a face-value diploma in a market that has repriced the output is not worth what the face value suggests.

What it does not carry:

Not numerical precision. A mortgage has a fixed dollar gap — the balance exceeds the appraisal by a specific number. A credential's "loan amount" is a fuzzy lifetime-earnings expectation priced against a production regime that has since changed. The accounting identity is structural; it is not arithmetic. The gap is real without being expressible as a single figure.

Not literal lock-in. Workers retrain and pivot. Career changes happen, retraining happens, midlife reinventions happen. The "cannot walk away" claim is a rhetorical rendering of a structural difficulty, not a description of juridical constraint. Workers do walk away. They walk away at fire-sale prices — which is exactly what the analogy is naming. The fire-sale is the mapping; the bailout is not.

Not an implied bailout. The 2008 mortgage crisis concluded with HAMP, HARP, and principal reductions. The article makes no equivalent policy claim. The accounting identity is being named here, not legislated against. What the credential-holder does with the recognition that the position is underwater is a different question, and a harder one.

The metaphor is a frame. The analysis is the inequality from the previous section. The frame gives the reader traction on the inequality; the inequality stands without the frame.

The analytical claim must survive the strongest counter — and the strongest counter is real, calibrated, and empirically grounded.

The strongest counter and where it stops#

The strongest counter is not a rhetorical move. It is a real, calibrated, empirically estimated model with documented predictive successes, and the article must confront it before it can hold the line.

Concession one — Bessen on demand elasticity. Bessen's 2018 NBER framework establishes that whether automation expands or contracts employment depends on the elasticity of demand for the automated service.16 The ATM case is clean: ATM deployment lowered per-transaction costs, which allowed banks to operate more branches, which raised total teller demand. Teller employment rose during a period of ATM proliferation. The model has been extended to AI: legal services face structural demand rationing, since the access-to-justice gap means current demand is suppressed by price — AI-driven cost reduction should unlock latent demand, potentially expanding legal employment overall. This is true. Concede it.

Concession two — Acemoglu and Restrepo reinstatement. The task framework is bidirectional. Recent extensions estimate fifty to eighty percent reinstatement ranges for AI specifically. PwC's 2025 Global AI Jobs Barometer found productivity quadrupled in AI-exposed industries (from seven to twenty-seven percent, 2018 to 2024); wages in AI-exposed industries grew at roughly twice the rate of less-exposed industries; jobs requiring AI skills grew 7.5 percent even as overall postings fell 11.3 percent; AI-skilled workers earn a fifty-six percent wage premium.10 This is the reinstatement effect the framework predicts. It is real and measurable. Concede it.

Concession three — Bessen et al. reabsorption. Empirical follow-up from Bessen and colleagues published in the Review of Economics and Statistics shows that automation-displaced workers face cumulative wage losses of approximately eight to nine percent of one year's wage income over the five-year measurement window of the study.17 The aggregate adjustment story is empirically supported. Concede this too.

Now the rebuttal, stated as accounting identity rather than empirical finding:

Demand elasticity expands employment at the new equilibrium price. The thirty-five to fifty-five worker's wage was set at the old regime's price. The wage is contractual; the elasticity is aggregate; they meet on different ledgers.

Bessen's model describes how new labor inputs at new equilibrium prices clear. It says nothing about the legacy cost structure attached to existing workers. The ATM/teller analogy holds because tellers could be hired at the new per-branch economics; the existing teller's salary was not the barrier. The credential-tier worker in 2026 cannot suddenly provide services at the AI-enabled price point. The internal cost structure — the accumulated salary negotiated in the old regime, the pension already vesting on the old formula — prevents it. The barrier is not the worker's capability. It is the worker's cost basis.

The fifty-six percent AI-skill premium accrues to workers who acquire AI complementarity. The lock-in argument concerns workers who face structural barriers to acquiring it — the cognitive-entrenchment mechanism Dane documented, the firm-underinvestment mechanism the IMF SDN documented. The premium and the lock-in describe different worker subpopulations within the same labor market. Citing the premium as evidence against the lock-in conflates the two.

The temporal argument is the sharpest point. Bessen et al.'s five-year measurement window — the period over which cumulative wage income losses accumulate — coincides exactly, for a fifty-year-old displaced in 2026, with the years of peak fiscal-claim accrual under most pension formulas. The lost wage income between 2026 and 2031 does not vanish on the other side of the measurement window: the pension formula that locked in during those missing years will produce a lower benefit for the rest of the worker's retirement. The wage loss is real. The pension gap is also real. They are reconciled on different ledgers.

The relevant historical precedent is not "this time is different." It is: "the pattern is the same." Krueger's 1993 research found a ten to fifteen percent wage premium for computer users during the 1980s computerization wave18 — the reinstatement effect was real, it benefited workers who could deploy the new technology, and it came at the cost of a permanent shift in the wage distribution that did not benefit displaced clerical workers in their lifetime. The pattern: aggregate labor markets adjust over twenty years; the cohort displaced during the transition does not recover in their generation. This is not a claim of structural rupture with all historical automation. It is the same generational asymmetry the 1980s produced, at a different tier and with different visibility.

The Yale Budget Lab's ongoing CPS-based tracking of AI's impact on U.S. labor markets presents a more cautious picture than the Stanford occupation-specific findings.19 Its analysis characterizes aggregate AI job-market impact as showing stability — not major disruption — in economy-wide employment statistics. This is not a contradiction. Aggregate stability is consistent with concentrated structural damage at specific cohort-plus-occupation intersections. The claim here is about that concentration, not about whether the macroeconomic headline number is alarming.

The rebuttal stands. The second-act story is still waiting — the one that connects what is visible now to what will become visible when the apprenticeship substrate runs dry.

The factory that builds expert judgment is closed#

Return to the twenty-two-year-old.

The entry-level wipe documented in the Stanford ADP data and corroborated in the REZI industry aggregations is not only a generation losing its first jobs. It is the apprenticeship substrate being economically erased. The junior roles at the base of professional pyramids — associate lawyers reviewing documents, junior radiologists reading routine images, junior consultants building decks, junior translators handling boilerplate — were not only employment opportunities. They were the mechanism by which the next generation of senior expert judgment was built. The document review was training. The boilerplate was training. The routine images were training. Each rung had economic value; that value justified hiring the person who would eventually climb past it.

Axios reported in May 2026 on what Big Law firms are doing in the face of this: racing to extract the knowledge of their lawyers before retirement, before the senior partners who built their expertise through that pipeline leave.3 This is what a profession does when it recognizes that the substrate producing expert judgment has been dismantled: it tries to capture the existing knowledge before the people who hold it are gone, because no process exists to produce the next generation of holders.

The structural inference is honest, not empirical. If junior roles are economically irrational under AI — because AI does the training-level work more cheaply than a junior associate can — the pipeline does not rebuild under market incentives. The market that justified hiring the junior associate for her economic output no longer justifies it, and the training value she would have extracted was never the primary justification for her salary. The 2035 to 2040 knowledge gap — when no human-trained judgment exists for the edge cases AI gets wrong — is not a forecast with empirical backing. It is the structural inference from current economics, and it should be held as such: honest, not certain.

The first generation has its career path foreclosed. The next generation will have no expert judgment to inherit. The civilization will discover the gap when it tries to use what is no longer there.

The visible damage at twenty-two is not an entry-level disruption that will correct when the economy adjusts. It is the dismantling of the mechanism by which expert judgment reproduces itself across generations. The two-locus structure is now fully visible: damage at the bottom of the career ladder — measurable, documented, accelerating; damage in the middle — accumulating on the balance sheet, invisible until the claims come due; damage in the future tense — inferential, structural, not yet appearing in any dataset because what is being destroyed is a pipeline, and pipelines run dry at the far end first.

What the balance sheet will show#

The balance sheet does not lie. It only delays.

The fiscal claims will come due. Pension funds will encounter the gap during 2030 to 2050, when the thirty-five to fifty-five cohort begins drawing benefits collateralized against a productive base that has been compressed. This is not a forecast about the labor market. It is an observation about arithmetic — pension formulas, amortization cycles, the contractual structure of benefits negotiated in a regime that no longer exists. The cash flow was positive while the gap accumulated. It will be negative when the gap materializes. The delay between accumulation and materialization is precisely the interval during which the income statement looked fine.

The Weimar wheelbarrow. When fixed nominal claims meet collapsed productive value, the historical pattern is currency printing — extending nominal claims while the underlying productive base falls further. The policy analogue is credential inflation: more retraining programs, more certifications, more AI-upskilling badges, more credential layers stacked above a market that has already repriced the underlying output. The wheelbarrow is the image of what happens when an institution attempts to honor fixed claims against a deteriorated base by issuing more of the claim. More credentials when the exchange system is broken does not restore exchange value; it accelerates devaluation, because the scarcity signal that gave the credential its market value is already gone.

What the accounting identity shows when held in full view: The dividend was a regime, not a structure. The credential is a frame, not an asset. The balance sheet will show what the income statement is hiding.

The burden that accumulates — slowly, invisibly, contractually — is not in the dependency ratio. It is in the gap between what a productive tier was promised and what AI-era economics will pay. That gap is not on the income statement. It never was.

The balance sheet does not show up in quarterly earnings. It shows up when the claims come due — on a ledger that was never designed to read the regime change that is already complete.

References#

  1. Brynjolfsson, E., Chandar, B., and Chen, R. (2025). "Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence." Stanford Digital Economy Lab Working Paper. ↩ ↩

  2. REZI AI (2025). "The Crisis of Entry-Level Labor in the Age of AI (2024–2026)." rezi.ai, citing IntuitionLabs and HR Katha analyses of employer job-posting data. ↩

  3. Axios (2026-05-02). "AI Threatens Big Law's Talent Pipeline." Also: Above the Law (2026-03). "AI Won't Replace Lawyers But Can Create Critical Shortage Of Good Ones." ↩ ↩

  4. Bloom, D., and Williamson, J. (1998). "Demographic Transitions and Economic Miracles in Emerging Asia." World Bank Economic Review, 12(3), 419–455. DOI: 10.1093/wber/12.3.419. ↩

  5. Cai, F. (2010). "Demographic Transition, Demographic Dividend, and Lewis Turning Point in China." China Economic Journal, 3(2), 107–119. DOI: 10.1080/17538963.2010.511899. ↩

  6. Das, M., and N'Diaye, P. (2013). "Chronicle of a Decline Foretold: Has China Reached the Lewis Turning Point?" IMF Working Paper WP/13/26. SSRN: 2222490. ↩

  7. Eloundou, T., Manning, S., Mishkin, P., and Rock, D. (2024). "GPTs are GPTs: Labor Market Impact Potential of Large Language Models." Science, 384, 1306–1308. arXiv preprint 2303.10130. ↩

  8. Felten, E., Raj, M., and Seamans, R. (2023). "Occupational Heterogeneity in Exposure to Generative AI." SSRN Working Paper 4414065. ↩

  9. Frey, C.B., and Llanos-Paredes, A. (2025). "How Machine Translation Has Displaced Translators." CEPR VoxEU, March 2025. Supporting data: Society of Authors (UK) Survey 2024. ↩

  10. Acemoglu, D., and Restrepo, P. (2019). "Automation and New Tasks: How Technology Displaces and Reinstates Labor." Journal of Economic Perspectives, 33(2), 3–30. DOI: 10.1257/jep.33.2.3. And (2022). "Tasks, Automation, and the Rise in U.S. Wage Inequality." Econometrica, 90(5), 1973–2016. DOI: 10.3982/ECTA19815. PwC (2025). "The Fearless Future: 2025 Global AI Jobs Barometer." ↩ ↩

  11. Cazzaniga, M., Jaumotte, F., Li, L., Melina, G., Panton, A., Pizzinelli, C., Rockall, E., and Tavares, M. (2024). "Gen-AI: Artificial Intelligence and the Future of Work." IMF Staff Discussion Note SDN/2024/001. ↩

  12. Pizzinelli, C., and Tavares, M. (2025). "AI and the Future of Work in an Aging Economy." Wharton Pension Research Council Working Paper WP2025-14. SSRN: 5345347. ↩

  13. Dane, E. (2010). "Reconsidering the Trade-off Between Expertise and Flexibility: A Cognitive Entrenchment Perspective." Academy of Management Review, 35(4), 579–603. DOI: 10.5465/amr.35.4.zok579. ↩

  14. Brynjolfsson, E., Li, D., and Raymond, L. (2025). "Generative AI at Work." Quarterly Journal of Economics, 140(2), 889–942. NBER Working Paper 31161. ↩

  15. Korinek, A., and Lockwood, B. (2026). "Public Finance in the Age of AI: A Primer." Brookings Institution Working Paper. January 2026. Also: NBER Working Paper 34873. ↩

  16. Bessen, J. (2018). "AI and Jobs: The Role of Demand." NBER Working Paper 24235. ↩

  17. Bessen, J., Goos, M., Salomons, A., and van den Berge, W. (2025). "What Happens to Workers at Firms that Automate?" Review of Economics and Statistics, 107(1), 125–141. ↩

  18. Krueger, A. B. (1993). "How Computers Have Changed the Wage Structure: Evidence from Microdata, 1984–1989." Quarterly Journal of Economics, 108(1), 33–60. ↩

  19. Yale Budget Lab. (2025–2026). "Tracking the Impact of AI on the Labor Market." The Budget Lab at Yale. Multiple updates 2024–2026, including February 2026 update. ↩

Further Reading#

  • Bessen, J., Goos, M., Salomons, A., and van den Berge, W. (2025). "What Happens to Workers at Firms that Automate?" Review of Economics and Statistics, 107(1), 125–141. — The empirical follow-up to Bessen 2018; documents cumulative wage losses of approximately eight to nine percent of one year's wage income over a five-year measurement window for workers at automating firms. Essential for understanding the temporal argument in §8 and why aggregate wage recovery does not rescue the locked-in thirty-five to fifty-five cohort's pension formula.
  • Autor, D. (2024). "Applying AI to Rebuild Middle Class Jobs." NBER Working Paper 32140. — Autor's argument that AI can democratize expert judgment to middle-skill workers; the optimistic-complement case that the article's analysis must coexist with. The democratization mechanism is the same one the article describes as producing the pipeline destruction — seen from a different normative angle.
  • Acemoglu, D. (2024). "The Simple Macroeconomics of AI." NBER Working Paper 32484. — Acemoglu's own assessment of why AI's macroeconomic productivity gains may be smaller than advertised if the automation is concentrated in tasks that are not complementary to broad human labor. Directly relevant to understanding the reinstatement-weakening finding cited in §4 and §8.
  • Cowen, T. (2013). Average Is Over. Dutton. — The anticipatory frame for the credential-tier argument: a world where the returns to being highly capable of working with sophisticated tools diverge sharply from the returns to working without them. Provides the distributional context within which the thirty-five to fifty-five balance-sheet argument sits.
  • Goldin, C. (2014). "A Grand Gender Convergence: Its Last Chapter." American Economic Review, 104(4), 1091–1119. — On the temporal structure of wage bargaining, career-path lock-in, and how compensation structures designed for continuous career trajectories produce systematic penalties when those trajectories are interrupted. Relevant background for understanding why the pension-formula mechanism operates as described in §6.

In brief#

Not aging. Not even obsolescence. The credential is underwater.

The demographic dividend was never a property of population age. It was a regime — a working-age cohort growing faster than dependents, paired with surplus rural labor that kept manufacturing wages below their marginal product. Bloom and Williamson attributed up to half of East Asia's 1965-to-1990 per-capita growth to this configuration; Cai Fang and the IMF later established its expiration on the demographic clock. AI has expired it sooner. A generation of credential infrastructure was built to amortize against a regime that has just ended early.

Two loci of damage now exist in the same labor market. The visible one sits at the bottom of the career ladder. Brynjolfsson, Chandar, and Chen at Stanford's Digital Economy Lab tracked ADP payroll microdata and found employment for software developers aged twenty-two to twenty-five fell roughly twenty percent from late 2022 to mid-2025; across high-AI-exposure occupations, the same cohort declined thirteen to sixteen percent against trend. Industry-aggregated postings data — U.S. entry-level technology down sixty-seven percent, UK technology graduate roles down forty-six percent — moves in the same direction. Big Law is racing to extract the knowledge of senior partners before retirement, because the document review that trained the next generation is now done by AI. The credential still confers status. The position it was built to fill has been automated before its holder arrived.

The invisible damage sits in the middle. Workers thirty-five to fifty-five in the same high-exposure categories show no employment decline; some cohorts show employment gains. That is not the rebuttal to the entry-level wipe. It is the second locus. The income statement is still positive — wages arriving, credentials framed on walls. The balance sheet is something else. The pension was promised at the old regime's productivity assumption. The salary was negotiated against the old wage premium. Healthcare benefits are contractually fixed. These claims compound; the productive value they were collateralized against has fallen. Eloundou and colleagues in Science documented that AI exposure is highest in higher-income white-collar occupations — the categories whose market value the credential investment was meant to justify. Past automation came for the assembly line. AI comes for the chambers.

The strongest counter — Bessen on demand elasticity, Acemoglu and Restrepo on reinstatement, the AI-skill wage premium — describes how new labor inputs clear at new equilibrium prices. It says nothing about the legacy cost structure attached to existing workers. The barrier is not capability. It is cost basis. And the apprenticeship substrate that produced expert judgment has been economically erased: junior roles are now irrational under AI economics, which means no pipeline rebuilds under market incentives.

The balance sheet does not lie. It only delays. The fiscal claims will come due between 2030 and 2050, on a ledger that was never designed to read the regime change that is already complete.