不再有平庸者：二阶独奏

一幅剪纸风格的齐国宫廷合奏图：一排穿袍乐师鼓着腮帮吹竽，只有中央一人静止不动，腮帮扁平，手持乐器却没有发声。

我们已经背熟的恐慌#

等你读到一篇新的关于 AI 与工作的文章时，你其实已经读过它了。数字会变，结构不会。先出现一个预测 —— 世界经济论坛《2025 未来就业报告》预计，到 2030 年将有 9,200 万个岗位被替代，同时创造 1.7 亿个新岗位，净增 7,800 万个；这个净增数字不知为何总是无法让人安心（WEF 2025）。随后是一波裁员，被解释为确认。2025 年，美国雇主宣布裁撤 120 万个岗位，比 2024 年增加 58%；报道沿着熟悉的弧线展开：AI 已经到来，筛选已经开始，选择你的阵营（Challenger Gray & Christmas 2026）。

本文讨论的是“合奏式不透明”的知识工作 —— 咨询、法律、产品、企业战略、研究写作。外科、航空、个人竞技运动，以及那些本来就会单独衡量你的职业，不在本文范围内；AI 不会改变那些原本就已经存在的可读性结构。

多数报道没有指出的是：在那 120 万个宣布裁撤的岗位中，只有 54,836 个 —— 占总数 4.5% —— 把 AI 列为原因。裁员潮是真的。AI 归因大多不是。多数裁员与 AI 实际改变产出无关：Davenport 和 Srinivasan 在 Harvard Business Review 2025 年 12 月对 1,000 多名高管所做的调查显示，60% 削减人手的组织，是出于对 AI 未来影响的预期，而不是因为 AI 已经交付了收益；只有 2% 报告了与实际 AI 实施相关的大规模裁员（Davenport & Srinivasan 2026）。Davenport 和 Srinivasan 把这读作借口：裁员本来就会发生；AI 提供了遮盖。

一幅剪纸风格的现代办公室插图：一排相同的知识工作者坐在相同的办公桌前，周围是一摞摞相同文件，个人贡献隐藏在集体产出之中。

这个差距 —— 被讲述的故事与真正驱动它的东西之间的差距 —— 正是本文的起点。如果 AI 并没有造成大部分替代，那造成它的是什么？为什么 AI 叙事对讲述它的人如此有用？

从未要求独奏的机构#

南郭先生 —— 下文会解释这个中国古代寓言中的模仿者 Mr. Nan Guo —— 通常并不是骗子。更多时候，我们现在想称为模仿者的人，是被训练其他人的同一套激励训练出来的 —— 被一个奖励集体产出、从不要求个人交账的系统塑造出来，然后在规则改变时让系统感到惊讶。如果我们想理解为什么合奏能运转那么久，就要从建造合奏的人开始，而不是从坐在合奏里的人开始。

几十年来，组织绩效管理的主导方向一直在远离个人问责。Cappelli 和 Tavis 在 2016 年一篇著名 HBR 文章中追踪了这条弧线：公司正在从年度评分系统转向持续发展对话，从个人打分转向团队指标，从问责转向培养（Cappelli & Tavis 2016）。这种转向并不荒唐。个人归因令人痛苦，政治成本高，组织上也有破坏性。麦肯锡 2024 年发现，超过三分之一的管理者被认为缺乏就绩效进行直接对话的能力 —— 这是感知调查，不是行为审计，而且很可能还是保守估计（McKinsey 2024）。

很长时间里，这种安排是稳定的。经济学家知道原因。信号理论 —— Michael Spence 在 1973 年一篇后来获得诺贝尔奖的论文中发展出来 —— 描述了当雇主无法区分高能力和低能力劳动者时，劳动力市场会发生什么：两组人取得同样的凭证，拿到同样的工资，并共同存在于所谓的混同均衡中（Spence 1973）。方程并不必要。很长时间里，凭证足够便宜，虚张声势者和专家都能取得同样信号并赚取同样工资。只要信号便宜，测试从未到来，这个均衡就成立。

中国有个古老故事。齐宣王喜欢听三百人合奏竽 —— 一种古代簧管乐器。一个叫南郭的人不会吹，却站在合奏队伍里，手指动着，领着和真正乐师一样的米。齐宣王死后，他的儿子齐湣王即位。齐湣王想听乐师一个一个吹。南郭当夜逃走。这个故事一直被记作关于模仿的寓言。“南郭先生”成了中文里对混在合奏中假装演奏者的简称。但这个故事一直是从错误的一端讲起的。雇用南郭的人 —— 那位组织三百人合奏而不是单独试奏任何人的宫廷主管 —— 并不天真。他在建造一种结构，在这种结构里，个人能力这个问题永远不必被提出。真正有意思的角色，不是那个不会吹的人，而是那个专门组织三百人合奏、从而让任何人都不必独奏的人。

这不是在反驳 Acemoglu 关于宏观生产率的担忧；那是问题的另一个层级。本文讨论的是哪些个体会变得可读，而不是蛋糕会长多大。

在那些 AI 大规模运作的职业中，AI 所做的，与其说是揭露欺诈，不如说是一场可读性事件 —— 一个长期依赖集体产出的系统，第一次开始产生能够读出个人的信号。那 60% 因预期 AI 而裁员、不是因 AI 已经交付收益而裁员的组织，响应的并不是一项新技术，而是在抓住一个他们以前无法正当化的重组理由。合奏让个人归因在政治上困难。AI 时刻让它在政治上容易。病理是旧的；催化剂是新的。

同时也在教学的合奏#

在继续之前：合奏值得辩护，本文也欠它一个辩护。

不透明不只是平庸的掩护 —— 它也是学徒制的基底。新人不能在第一天独奏，也不应该被要求独奏。医院有查房，律所有合伙制，咨询团队带着分析师，管弦乐队有声部。在所有这些地方，新人由集体托举，同时发展自己的判断。不透明是保护性的，不是逃避性的。合奏吸收错误，而学徒建立避免错误的能力。

ATM 是一面有用的历史镜子。从 1980 年代开始，ATM 在美国部署，最终达到约 40 万台。你可能会以为银行柜员就业会崩塌。它没有。柜员就业实际上从约 50 万增长到 60 万，因为更低的网点运营成本让银行开设了更多网点，并把柜员角色转向需要人类判断的关系型银行服务（Bessen 2015）。ATM 没有揭露柜员；它改变了柜员所做的事。合奏改变了形状，但仍然存在。

Goldin 和 Rouse（2000）的盲选研究显示了同一机制的反向运作：根据该研究，在乐手和评委之间放置实体屏风，使女性从初选晋级的概率增加约 50% —— 这个幅度后来受到统计批评的争议（Gelman 2019），但方向仍然成立。刻意移除不透明，改变了谁能通过。合奏此前承载的不只是学徒，也包括永久居民。

每个合奏中都有两种不透明。第一种是发展性的 —— 粮仓里的种子。第二种是永久性的 —— 永远不让种子出来的粮仓。本文批判的是第二种。

一幅剪纸风格插图：剖面显示的传统粮仓中长出一株幼芽 —— 粮仓里的种子，具有自身解决期限的发展性不透明。

发展性不透明会被解决。学徒制有时间线，而在时间线尽头，学徒被期待知道，也被期待被人知道。合奏教会独奏；最终它会要求独奏。本文关心的合奏，是那个从未提出要求的合奏 —— 一种结构，在那里十二年、十五年、二十年过去，个人判断的问题就是从未到来。Dunning 和 Kruger（1999）记录了它从内部看起来是什么样：测试表现处于底部四分位的人，会估计自己大约在第 62 百分位，不是因为他们不诚实，而是因为表现良好所需的知识，正是识别糟糕表现所需的知识 —— 元认知差距和表现差距是同一件事。这个发现的幅度有争议；方向没有。

这些都不是敌视学徒制。问题是学徒制本应发展成什么 —— 以及当合奏从不发问时，会发生什么。

AI 实际自动化了什么，又安装了什么#

经验证据图景确实复杂，假装不复杂是不诚实的。

把 AI 视作伟大均衡器的最强版本是真实的，也值得认真对待。Noy 和 Zhang（2023）在发表于 Science 的一项预注册随机对照试验中发现，对于中等难度的专业写作任务，ChatGPT 显著抬高了下限 —— 低能力写作者提升最多，分布被压缩，不平等下降（Noy & Zhang 2023）。如果 AI 给最弱表现者带来最大提升，那么合奏的不透明就不是通过揭露南郭先生而消解，而是通过让他们真正更有能力而消解 —— 这就意味着“二阶独奏”这个说法是错的。Peng 等人（2023）在编码中发现了类似现象：GitHub Copilot 在一个定义清楚的任务上让开发者平均提速 55.8%，经验较少的开发者受益更多（Peng et al. 2023）。两个领域的分布都被压缩了。

调节变量是任务类型。

趋同发现对我们可以称为“前沿以内的生产任务”很稳健：写一份草稿，为指定函数生成代码，完成一个质量标准明确、AI 错误能被任何人检查出来的定义清楚的交付物。这些任务趋同。发散发现出现在任务从按规格生产转向判断什么算正确时 —— 尤其是当任务需要对“这对吗？”这个问题给出领域特定答案时。

Dell'Acqua 及其同事（2023/2026）对 758 名 BCG 咨询顾问做了田野实验。在专业能力前沿以内的任务上 —— 标准咨询套件 —— 低于平均水平的顾问使用 AI 后质量提升 43%；高于平均水平的顾问提升 17%。趋同。但当任务转向一个落在 AI 能力前沿之外的商业问题解决任务时，使用 AI 的顾问给出正确解决方案的概率比不使用 AI 的人低 19 个百分点（Dell'Acqua et al. 2023）。AI 给了错误的信心。论文说，工作者“在方向盘前睡着了” —— 停止质疑那些看起来权威的输出。论文命名的“锯齿状技术前沿”，正是 AI 可靠帮忙的任务与 AI 可靠误导的任务之间的边界 —— 而关键技能，是知道自己站在哪一边。

Vendraminelli 及其同事 2025 年 HBS 工作论文中所谓的 GenAI Wall Effect，让这个机制更清楚。在一家大型英国公司的受控实验中，AI 在执行任务上为相邻外部人抹平了场地 —— 但缺乏领域知识来评估 AI 输出的远距离外部人，即便有 AI 辅助，也无法缩小与内部人的差距（Vendraminelli et al. 2025）。墙不是努力问题。它是评估问题。远距离外部人无法判断 AI 输出是否合适；内部人可以。AI 放大了这个差异。

METR（2025）在编码中发现了相邻现象：经验丰富的开源开发者在复杂代码库上使用 AI 工具时反而慢了 19%，尽管他们预期会提速 24% —— 而且事后他们仍然相信 AI 让自己更快了（METR 2025）。这项研究规模小（N=16；发现具有提示性而非决定性），但感知差距很重要：即便专家也会误判评估问题。

三个职业显示出同样结构：

在法律工作中，生产任务是用 AI 辅助起草一份诉状；任何合格助理都能用旧时间的一小部分完成。评估任务是在签上自己名字之前识别哪些引文是幻觉。这两件事需要的能力不同。第一件需要熟悉格式。第二件需要足够懂法理，知道 Fabulozo Airlines v. Carillo 并不存在。

在管理咨询中，生产任务是生成一套关于行业动态的 40 页幻灯片；AI 可以在几分钟内生成一个貌似合理的版本。评估任务是识别 AI 的哪些框架假设不适用于这个客户的真实问题 —— 那些嵌入数据选择、竞争框架、隐含市场定义中的假设。抓住这些，需要在足够多语境中做过足够多次这项工作，才能识别看似合理的答案何时是错的。

在产品管理中，生产任务是撰写需求文档和用户故事；AI 做得不错。评估任务是识别 AI 生成的综合是否漏掉了一个会在生产环境中出问题的边界情况 —— 这需要把产品实际推过生产环境足够多次，脑中有一套边界情况如何呈现的模型。

这三个职业的共同点是：评估步骤需要无法从 AI 借来的领域特定判断，因为 AI 正是被评估的对象。

评估型 AI 正在到来。它不会消解这个机制，因为构建评估型 AI 需要领域量规，而只有已经拥有领域判断的人才能制定这些量规 —— 这正是 GenAI Wall 显示的东西。护城河也许会缩小，但仍然需要有人告诉评估型 AI，目标领域里的“正确”长什么样。即便护城河缩小得比这个论证预测得更快，“建立评估能力”在 24 个月视野内仍然是可执行的 —— 而读者眼下要做的大多数决定，恰恰都在这个视野内。

劳动力市场数据支持这个总体方向：Brynjolfsson、Chandar 和 Chen（2025）使用 ADP 工资记录和 AI 暴露指数发现，在最受 AI 暴露影响的职业中，22-25 岁劳动者的相对就业下降 16%，而同一职业中的有经验劳动者保持稳定或增长（Brynjolfsson et al. 2025）。如果 AI 正在自动化入门级生产任务，而资深评估任务大体仍然完整，这个模式正是你会预期看到的。

评估测试已经开始#

关于评估失败代价，最具体的证据来自法律 —— 因为那里有书面记录，也因为法院在遇到无能时有结构性激励去点名它。

2023 年，律师 Steven Schwartz 使用 ChatGPT 为一宗航空人身伤害案件做研究。他提交的诉状包含多项案例引用：Varghese v. China Southern Airlines、Shaboon v. EgyptAir、Petersen v. Iran Air 等等。这些案例并不存在。当 P. Kevin Castel 法官要求提供副本时，Schwartz 提交了宣誓书证明它们真实存在，却没有核查。法院最终分别对 Schwartz、他的合伙人 LoDuca 以及律所处以 5,000 美元制裁 —— 合计 15,000 美元 —— 并认定存在“有意识回避以及向法院作出虚假和误导性陈述”（Mata v. Avianca 2023）。生成很容易。失败的是评估测试。

模式并不是 Schwartz 是个异类。斯坦福法学院 2025 年一项预注册研究，测量了专门为减少幻觉而设计的 AI 法律研究工具中的幻觉率：Lexis+ AI 在 17% 的查询中生成伪造引文；Westlaw AI-Assisted Research 为 33%；GPT-4 为 43%（Magesh et al. 2025）。这些不是通用聊天机器人。它们是专门构建、带检索增强生成、瞄准引文准确性这个具体问题的工具。它们仍然在大约六分之一到三分之一的查询中产生幻觉。评估测试不是思想实验。它是正在运行的基线专业要求。

加州上诉法院在 Noland v. Land of the Free（2025）中把这一点具体化。律师 Amir Mostafavi 提交的开篇诉状中，23 处引文有 21 处是伪造的；法院对他处以 10,000 美元制裁，并把他移交州律师协会。到了是否判给对方律师费的问题，法院拒绝了 —— 因为对方律师也没有抓住伪造引文，显然是在法院自己提出问题后才意识到。这是一个新兴信号：识别对方诉状中的 AI 幻觉，正在成为基线专业能力的一部分 —— 不是新的法律义务，而是一种正在悄悄形成的职业期待（Noland v. Land of the Free 2025）。

一幅剪纸风格插图：深色桌面上一份单独文件，被上方聚焦灯光照亮，其中一个具体主张用琥珀色标出，其余文字仍未检查 —— 一个评估瞬间真正的样子。

把文章放下一分钟。想想你最近依赖的一份 AI 生成工作 —— 一份备忘录、一份草稿、一份市场摘要、一段代码、一份研究综合、一套幻灯片。具体说出你在依赖之前核查过的最后一个 AI 生成主张。不要说“我通常都会审阅一切”。不要说“我大体有感觉”。说出过去两周里，你从某个具体 AI 输出中拿出并对照一手来源验证过的一个具体主张。如果你说不出来，你是在把 AI 用于生产，而不是用于评估。这不是判决。它是诊断。

如果你能说出一个，注意一下做这件事是什么感觉 —— 花了多久，你做了什么，下一次你是否还会这样做。那种感觉，就是评估能力被动用时的感觉。那是测试已经开始的感觉。

测试不会公平#

对以上所有论证，有一个真正的反对意见，它值得的不只是一条脚注。

最终出现的具体测试会是政治性的。哪些幻觉算数？哪个“正确”才是正确？每一种评估制度都会偏向某些知识而不是另一些知识；谁设置测试，谁就塑造哪些工作者能够通过。法律引文核查、战略框架核查、产品边界情况核查，都是同一要求的不同版本 —— 但它们不是中性指标。它们是被建构出来的类别，而且会由有利益的人来建构。

回想 Goldin 和 Rouse（2000）的盲选研究：屏风提高女性晋级概率，是因为它移除了一种偏见来源 —— 但试奏标准本身（曲目、评判量规、“正确”音准的标准）是由处在权力位置的人设定的，而这些标准带着自己的假设。盲选比此前的做法更公平。它们并不在任何绝对意义上公平。

但是：以测试可争议为由拒绝建立评估能力，就等于在 1970 年因为试奏标准可争议而拒绝盲选。标准确实可争议。盲选仍然重要。

机制方向不是捏造的。无论咨询、法律、产品、企业战略中出现的具体测试是什么，它们都会要求评估能力，而不仅仅是生产能力。在自己的领域中建立评估判断 —— “正确”长什么样，AI 最可能在哪里漏掉它，当输出看起来合理时应该追问什么 —— 这是在各种可能测试设计中都成立的投资。具体实施会是政治性的。底层需求不会。

投资动作#

如果你在上一节说不出一个核查过的主张，羞愧不是正确反应。有用的事，是识别出你所在领域里真正的评估能力长什么样 —— 不是作为一个类别，而是作为一种习惯 —— 然后开始建立它。

从下一份你依赖的 AI 输出开始。具体说出你所在领域里的“正确”长什么样。然后测试输出是否满足这个定义。不是它看起来是否合理 —— 而是它是否满足定义。持续做。评估能力通过重复发展；最初几次会缓慢、不确定，而且很可能会暴露出你过去依赖的东西比你检查过的更多。

读到这里的职业中段人士，站在一个特定决策点。工作十二到十五年 —— 发展性不透明本应已经解决。学徒窗口已经关闭，或正在关闭。真正的判断是否在其中形成，是接下来二十四个月会回答的问题，一次 AI 输出回答一次。这不是敌意观察。处在这个阶段的人 —— 职业中段、有证书、有经验，身处本文一直讨论的合奏式不透明职业 —— 拥有早期职业者没有的东西：多年的真实领域暴露，可用来扎根评估。

问题是，这些年建立的是核查习惯，还是生产习惯。二者并不相同；AI 让这种差异以前所未有的方式变得可见，因为过去输出总是集体的。

发展性与永久性的区别在这里很重要：如果合奏曾作为学徒托举你，并逐渐给你更多个人责任，那么 AI 正在测试的是你已经建立的判断。你会在使用它时认出它。如果合奏从未发问 —— 如果十二年过去，你一直是那个为幻灯片做贡献、却从未必须为幻灯片是否正确辩护的人 —— 那么 AI 正在安装的，是一场本来早该到来的测试，只是它来晚了。

南郭当夜逃走，因为他知道自己无法通过独奏。二阶独奏不同。不是齐湣王叫你演奏。它是晚上 11 点那个安静瞬间：你把又一份 AI 草稿粘进又一份你将签上自己名字的文件里，然后决定自己要如何阅读它 —— 像一个能抓住错误的人那样读，还是像一个希望一切都没问题的人那样读。

参考文献#

Bessen, J. "Toil and Technology." Finance & Development (IMF), March 2015. https://www.imf.org/external/pubs/ft/fandd/2015/03/bessen.htm

Brynjolfsson, E., Chandar, B., & Chen, R. "Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence." Stanford Digital Economy Lab Working Paper, August 2025. https://digitaleconomy.stanford.edu/

Cappelli, P., & Tavis, A. "The Performance Management Revolution." Harvard Business Review, October 2016, pp. 58–67. https://hbr.org/2016/10/the-performance-management-revolution

Challenger, Gray & Christmas, Inc. 2025 Year-End Job Cut Report. January 8, 2026. https://www.challengergray.com/

Davenport, T.H., & Srinivasan, L. "Companies Are Laying Off Workers Because of AI's Potential — Not Its Performance." Harvard Business Review, January 2026.

Dell'Acqua, F. et al. "Navigating the Jagged Technological Frontier: Field Experimental Evidence of the Effects of AI on Knowledge Worker Productivity and Quality." HBS Working Paper 24-013 (September 2023); published Organization Science, 2026. https://ssrn.com/abstract=4573321

Goldin, C., & Rouse, C. "Orchestrating Impartiality: The Impact of 'Blind' Auditions on Female Musicians." American Economic Review, 90(4):715–741, September 2000.

Magesh, V. et al. "Hallucination-Free? Assessing the Reliability of Leading AI Legal Research Tools." Journal of Empirical Legal Studies, 2025. DOI: 10.1111/jels.12413. https://arxiv.org/abs/2405.20362

McKinsey & Company (Hancock, B. & Weddle, B.). "What Works — and Doesn't — in Performance Management." November 2024.

METR. "Measuring the Impact of Early-2025 AI on Experienced Open-Source Developer Productivity." July 10, 2025. ArXiv: 2507.09089. https://arxiv.org/abs/2507.09089

Mata v. Avianca, Inc., No. 1:22-cv-01461, 678 F.Supp.3d 443 (S.D.N.Y. June 22, 2023).

Noland v. Land of the Free, L.P., California Court of Appeal, Case No. B331918, September 12, 2025.

Noy, S., & Zhang, W. "Experimental Evidence on the Productivity Effects of Generative Artificial Intelligence." Science, 381(6654):187–192, July 14, 2023. DOI: 10.1126/science.adh2586.

Peng, S., Kalliamvakou, E., Cihon, P., & Demirer, M. "The Impact of AI on Developer Productivity: Evidence from GitHub Copilot." ArXiv: 2302.06590, February 2023. https://arxiv.org/abs/2302.06590

Spence, M. "Job Market Signaling." Quarterly Journal of Economics, 87(3):355–374, August 1973.

Vendraminelli, L. et al. "The GenAI Wall Effect: Examining the Limits to Horizontal Expertise Transfer Between Occupational Insiders and Outsiders." HBS Working Paper 26-011, September 2025. https://ssrn.com/abstract=5462694

World Economic Forum. "The Future of Jobs Report 2025." January 2025. https://www.weforum.org/reports/the-future-of-jobs-report-2025/

Dunning, D., & Kruger, J. "Unskilled and Unaware of It: How Difficulties in Recognizing One's Own Incompetence Lead to Inflated Self-Assessments." Journal of Personality and Social Psychology, 77(6):1121–1134, December 1999. PubMed PMID: 10626367.

Acemoglu, D., & Restrepo, P. "The Wrong Kind of AI? Artificial Intelligence and the Future of Labor Demand." Cambridge Journal of Regions, Economy and Society, 13(1):25–35, 2020. NBER Working Paper 25682. https://www.nber.org/papers/w25682

No More Mediocres — The Second-Order Solo

A paper-cut illustration of the Qi court ensemble — a row of robed musicians playing the yu with cheeks puffed, except one still figure at center whose cheeks are flat and whose instrument is held but unsounded

The Panic We Know By Heart#

By the time you read a new article about AI and jobs, you have already read it. The numbers change; the structure does not. A projection appears — the World Economic Forum's 2025 Future of Jobs Report projects 92 million roles displaced by 2030, with 170 million new roles created, a net gain of 78 million that somehow never seems to land as reassuring (WEF 2025). A wave of layoffs follows, interpreted as confirmation. In 2025, US employers announced 1.2 million job cuts — a 58% increase over 2024 — and the coverage ran its familiar arc: AI has arrived, the sorting has begun, choose your side (Challenger Gray & Christmas 2026).

This piece is about ensemble-opaque knowledge work — consulting, legal, product, corporate strategy, research writing. Surgery, aviation, individual sports, and the other professions that already measure you solo are not in scope; AI does not change legibility structures that were already there.

Here is what most of that coverage did not note: only 54,836 of those 1.2 million announced cuts — 4.5% of the total — cited AI as a reason. The layoff wave was real. The AI attributions was mostly not. Most of those cuts had nothing to do with AI actually changing output: according to a December 2025 survey of over 1,000 executives by Davenport and Srinivasan in Harvard Business Review, 60% of the organizations that reduced headcount did so in anticipation of AI's future impact, not because AI had delivered gains — and only 2% reported large layoffs tied to actual AI implementation (Davenport & Srinivasan 2026). Davenport and Srinivasan read this as pretext: the layoffs were already coming; AI provided the cover.

A paper-cut illustration of a modern office — a row of identical knowledge workers at identical desks, surrounded by identical stacks of documents, individual contribution hidden inside the collective output

That gap — between the story being told and what is actually driving it — is where this article begins. If AI is not causing most of the displacement, what is? And why is the AI story so useful to the people telling it?

The Institution That Never Called for Solos#

The 南郭先生 — Mr. Nan Guo, the mimic of an old Chinese parable explained below — was not usually a fraud. Most often, the person we now want to call a mimic was trained by the same incentives that trained everyone else — was shaped by a system that rewarded collective output, never once asked for an individual accounting, and was then surprised when the rules changed. If we want to understand why the ensemble worked for so long, we start with the people who built it, not the people who sat inside it.

For decades, the dominant direction of organizational performance management was away from individual accountability. Cappelli and Tavis traced this arc in a well-known 2016 HBR piece: companies were moving from annual-rating systems toward continuous development conversations, from individual scoring toward team-based metrics, from accountability to cultivation (Cappelli & Tavis 2016). The shift was not irrational. Individual attribution is painful, politically costly, and organizationally disruptive. Over a third of managers, McKinsey found in 2024, are perceived as lacking the skills to have direct conversations about performance — a perception survey, not a behavioral audit, and likely conservative at that (McKinsey 2024).

For a long time, this arrangement was stable. Economists have a name for why. Signaling theory — developed by Michael Spence in a 1973 paper that eventually won a Nobel Prize — describes what happens in labor markets when employers cannot distinguish between high- and low-ability workers: both groups acquire the same credential, receive the same wage, and coexist in what is called a pooling equilibrium (Spence 1973). The equations are not necessary. For a long time, the credential was cheap enough that the bluffer and the expert could acquire the same signal and earn the same wage. That equilibrium held as long as the signal was cheap and the test never came.

There is an old Chinese story. King Xuan of Qi liked to hear the yu — an ancient reed instrument — played by an ensemble of 300. A man named Nan Guo could not play, but he stood in the ensemble, fingers moving, receiving the same rice as the real musicians. When King Xuan died and his son King Min took the throne, King Min wanted to hear the players one by one. Nan Guo fled that night. The story has been remembered as a parable about mimicry. 南郭先生 — "Mr. Nan Guo" — became Chinese shorthand for the miming ensemblist. But the story has always been told from the wrong end. The man who hired Nan Guo — the court master who organized a 300-player ensemble rather than auditioning any player alone — was not naive. He was building something in which the question of individual competence would never have to be asked. The interesting character is not the man who could not play. It is the court master who assembled an ensemble of 300 precisely so that no one would ever have to.

This is not an argument against Acemoglu's concern about macro productivity; that is a different level of the problem. This is an argument about which individuals become legible, not about how big the pie grows.

What AI is doing, in the professions where it operates at scale, is less an exposure of fraud than a legibility event — the moment when a system that long ran on collective output begins, for the first time, to produce individually readable signals. The 60% of organizations cutting staff in anticipation of AI, not because of AI-delivered gains, are not responding to a new technology so much as seizing a new justification for a restructuring they could not previously justify. The ensemble made individual attribution politically difficult. The AI moment makes it politically easy. The pathology is old; the catalyst is new.

The Ensemble That Also Taught#

Before going further: the ensemble earned a defense, and this article owes it one.

Opacity is not only cover for mediocrity — it is the substrate of apprenticeship. Juniors cannot solo on Day 1 and should not be asked to. Hospitals have rounds, law firms have partnerships, consulting teams carry analysts, orchestras have sections. In all of these, the new person is carried by the collective while their judgment develops. The opacity is protective, not evasive. The ensemble absorbs error while the apprentice builds the capacity to avoid it.

The ATM is a useful historical mirror here. When ATMs were deployed across the United States from the 1980s onward — eventually numbering around 400,000 machines — you might expect bank teller employment to have collapsed. It did not. Teller employment actually grew, from roughly 500,000 to 600,000, because lower branch operating costs allowed banks to open more branches, shifting the teller role toward relationship banking that required human judgment (Bessen 2015). The ATM didn't expose tellers; it transformed what tellers did. The ensemble changed shape, but it held.

The Goldin and Rouse (2000) blind audition study shows the same mechanism working in reverse: according to the study, placing a physical screen between musician and evaluator increased by about 50% the probability a woman would advance from a preliminary round — the magnitude is contested by a later statistical critique (Gelman 2019), but the direction survives. Deliberately removing opacity changed who got through. The ensemble had been carrying not just apprentices but permanent inhabitants.

Two kinds of opacity are at work in every ensemble. The first is developmental — the seed in the silo. The second is permanent — the silo that never lets the seed out. The critique is of the second.

A paper-cut illustration of a single sprout growing inside a traditional grain silo shown in cross-section — the seed in the silo, developmental opacity with its own resolution

Developmental opacity resolves. The apprenticeship has a timeline, and at the end of that timeline the apprentice is expected to know, and to be known. The ensemble teaches the solo; eventually it calls for one. The ensemble this argument is concerned with is the one that never called — the structure in which twelve, fifteen, twenty years passed and the question of individual judgment simply never arrived. Dunning and Kruger (1999) documented what this looks like from the inside: people in the bottom quartile of tested performance estimated themselves at around the 62nd percentile, not because they were dishonest but because the same knowledge needed to perform well is needed to recognize poor performance — the metacognitive gap and the performance gap are one and the same thing. The magnitude of that finding is debated; the direction is not.

None of this is hostile to apprenticeship. The question is what apprenticeship was supposed to resolve into — and what happens when the ensemble never asks.

What AI Actually Automates, and What It Installs#

The empirical picture is genuinely complicated, and it would be dishonest to pretend otherwise.

The strongest version of the case for AI as a great equalizer is real, and it deserves an honest hearing. Noy and Zhang (2023), in a preregistered randomized controlled trial published in Science, found that for mid-level professional writing tasks, ChatGPT substantially raised the floor — low-ability writers improved most, the distribution compressed, and inequality fell (Noy & Zhang 2023). If AI gives the weakest performers the biggest lift, then the ensemble's opacity dissolves not by exposing the 南郭先生 but by making them genuinely more capable — which would mean the second-order solo is simply wrong. Peng et al. (2023) found something similar in coding: GitHub Copilot gave developers an average 55.8% speedup on a well-defined task, with less experienced developers benefiting more (Peng et al. 2023). The distribution compressed in both domains.

The moderator is task type.

The convergence findings are robust for what we might call inside-frontier production tasks: writing a draft, generating code for a specified function, completing a defined deliverable where the quality criteria are clear and the AI's errors are checkable by anyone. These tasks converged. The divergence findings appear when the task shifts from producing to spec to judging what counts as correct — and especially when the task requires a domain-specific answer to the question: is this right?

Dell'Acqua and colleagues (2023/2026) ran a field experiment with 758 BCG consultants. On tasks inside the expertise frontier — the standard consulting suite — below-average consultants using AI improved quality by 43%; above-average consultants improved by 17%. Convergence. But when the task shifted to one business problem-solving task that fell outside the AI's competence frontier, consultants using AI were 19 percentage points less likely to produce correct solutions than those without (Dell'Acqua et al. 2023). The AI gave false confidence. The workers "fell asleep at the wheel," as the paper puts it — ceasing to question output that appeared authoritative. What the study named the "jagged technological frontier" is the boundary between the tasks where AI reliably helps and the tasks where AI reliably misleads — and the critical skill is knowing which side you are on.

The mechanism becomes even clearer in Vendraminelli and colleagues' 2025 HBS working paper on what they call the GenAI Wall Effect. In a controlled experiment at a large UK firm, AI leveled the playing field for adjacent outsiders on execution tasks — but distant outsiders, who lacked the domain knowledge to evaluate AI output, could not close the gap with insiders even with AI assistance (Vendraminelli et al. 2025). The wall is not an effort problem. It is an evaluation problem. Distant outsiders could not judge whether the AI output was appropriate; insiders could. AI amplified that difference.

METR (2025) found something adjacent in coding: experienced open-source developers working on complex codebases were 19% slower when using AI tools, despite expecting a 24% speedup — and even after the fact, they still believed AI had made them faster (METR 2025). The study was small (N=16; the findings are suggestive, not definitive) but the perception gap matters: even experts misjudged the evaluation problem.

Three professions show the same structure:

In legal work, the production task is drafting a brief with AI assistance; any competent associate can do it in a fraction of the old time. The evaluation task is identifying which citations are hallucinated before signing your name to them. These require different things. The first requires fluency with the format. The second requires knowing the doctrine well enough to know that Fabulozo Airlines v. Carillo does not exist.

In management consulting, the production task is generating a 40-slide deck on industry dynamics; AI can produce a plausible version in minutes. The evaluation task is identifying which of the AI's framing assumptions are wrong for this client's actual problem — the assumptions baked into the data selection, the competitive frame, the implicit market definition. Catching those requires having done this work enough times in enough contexts to recognize when the plausible-looking answer is wrong.

In product management, the production task is writing requirement documents and user stories; AI does this well. The evaluation task is recognizing when the AI-generated synthesis has missed an edge case that will break in production — which requires having run products through production enough times to carry a mental model of how edge cases present.

What these three professions share is that the evaluation step requires domain-specific judgment that cannot be borrowed from AI, because AI is what is being evaluated.

Evaluator-AI is arriving. It will not dissolve the mechanism, because building an evaluator-AI requires domain rubrics that only someone who already has domain judgment can formulate — which is precisely what the GenAI Wall shows. The moat may shrink, but a human still has to tell the evaluator-AI what "correct" looks like in the target domain. And even if the moat shrinks faster than that argument predicts, "build evaluation competence" remains actionable for a 24-month horizon — and most of what the reader is deciding right now is on that horizon.

The labor market data backs the general direction: Brynjolfsson, Chandar, and Chen (2025), using ADP payroll records and AI exposure indices, found a 16% relative employment decline for workers aged 22–25 in the most AI-exposed occupations, while experienced workers in the same occupations remained stable or grew (Brynjolfsson et al. 2025). The pattern fits what you would expect if AI were automating the entry-level production tasks while the senior evaluation tasks remained largely intact.

The Evaluation Test, Already in Session#

The most concrete evidence for what evaluation failure costs comes from law — where written records exist, and where courts have a structural incentive to name incompetence when they encounter it.

In 2023, attorney Steven Schwartz used ChatGPT to research an aviation personal-injury case. The brief he filed included multiple case citations: Varghese v. China Southern Airlines, Shaboon v. EgyptAir, Petersen v. Iran Air, among others. These cases do not exist. When Judge P. Kevin Castel asked for copies, Schwartz submitted affidavits attesting to their reality without checking. The court eventually issued sanctions of $5,000 each to Schwartz, his partner LoDuca, and the firm — $15,000 total — and found "conscious avoidance and false and misleading statements to the Court" (Mata v. Avianca 2023). Generation was easy. Evaluation was the test that failed.

The pattern is not that Schwartz was an outlier. A 2025 preregistered study from Stanford Law measured hallucination rates in dedicated AI legal research tools specifically designed to reduce hallucination: Lexis+ AI produced fabricated citations in 17% of queries; Westlaw AI-Assisted Research in 33%; GPT-4 in 43% (Magesh et al. 2025). These are not general-purpose chatbots. They are purpose-built tools with retrieval-augmented generation aimed at the specific problem of citation accuracy. They still hallucinate in roughly one in six to one in three queries. The evaluation test is not a thought experiment. It is the baseline professional requirement, running now.

A California appeals court made this concrete in Noland v. Land of the Free (2025). Attorney Amir Mostafavi filed an opening brief in which 21 of 23 quotations were fabricated; the court sanctioned him $10,000 and referred him to the state bar. When it came to the question of opposing counsel's fees, the court declined to award them — because opposing counsel had also failed to catch the fabricated citations, apparently becoming aware of the problem only when the court itself raised it. This is an emerging signal that spotting AI hallucinations in opposing briefs is becoming part of baseline professional competence — not a new legal duty, but a professional expectation that is quietly taking shape (Noland v. Land of the Free 2025).

A paper-cut illustration of a single document on a dark desk, lit by a focused lamp from above, with one specific claim marked in amber while the rest of the text remains unchecked — what one moment of evaluation actually looks like

Put the article down for a minute. Think about the last AI-generated piece of work you relied on — a memo, a draft, a market summary, a piece of code, a research synthesis, a deck. Name, specifically, the last AI-generated claim you fact-checked before relying on it. Not "I usually review everything." Not "I have a general sense." Name one specific claim, from one specific piece of AI output, that you verified against a primary source in the last two weeks. If you cannot name one, you are using AI for production, not for evaluation. That is not a verdict. It is a diagnostic.

If you can name one, notice what doing that felt like — how long it took, what you did, whether you would do it again for the next piece. That feeling is the feeling of evaluation competence being exercised. It is the feeling of the test, in session.

The Tests Will Not Be Fair#

There is a genuine objection to everything above, and it deserves more than a footnote.

The specific tests that emerge will be political. Which hallucinations count? Which "correct" is correct? Each evaluative regime privileges some knowledge over others, and whoever sets the test shapes which workers pass. A legal citation check and a strategic-framing check and a product edge-case check are all versions of the same demand — but they are not neutral metrics. They are constructed categories, and they will be constructed by people with interests.

Recall Goldin and Rouse's (2000) blind audition study: the screen increased the probability of female advancement because it removed one source of bias — but the audition criteria themselves (the repertoire, the judging rubric, the standards for "correct" intonation) were set by people in positions of power, and those criteria carried their own assumptions. The auditions were fairer than what preceded them. They were not fair in any absolute sense.

But: refusing to build evaluation competence on the grounds that the tests are contestable is the position that would have rejected blind auditions in 1970 because the audition criteria were contestable. The criteria were contestable. The auditions mattered anyway.

The direction of the mechanism is not fabricated. Whatever specific tests emerge across consulting, law, product, and corporate strategy, they will demand evaluation competence rather than mere production competence. Building evaluation judgment in your own domain — what "correct" looks like, where the AI is most likely to miss it, what questions to press when the output seems plausible — is the investment that holds across the range of possible test designs. The specific implementation will be political. The underlying demand will not.

The Investment Move#

If you could not name a fact-checked claim in the preceding section, shame is the wrong response. The useful thing is to identify what genuine evaluation competence looks like in your own domain — not as a category, but as a habit — and start building it.

Start with the next piece of AI output you rely on. Name, specifically, what "correct" would look like in your domain. Then test whether the output meets that definition. Not whether it looks plausible — whether it meets the definition. Keep doing it. Evaluation competence develops through repetition, and the first few repetitions will be slow and uncertain and will probably reveal that you have been relying on more than you checked.

The mid-career professional reading this piece sits at a particular decision point. Twelve to fifteen years in — the developmental opacity should have resolved. The apprenticeship window has closed, or is closing. Whether genuine judgment developed inside it is a question that the next twenty-four months will answer, one AI output at a time. That is not a hostile observation. Someone at this stage — mid-career, credentialed, experienced, in the ensemble-opaque professions this article has been about — has something that earlier-career workers do not: years of actual domain exposure in which to ground evaluation.

The question is whether those years built a habit of checking or a habit of producing. The two are not the same, and AI makes the difference visible in a way that was not previously possible when output was always collective.

The developmental-vs-permanent distinction matters here: if the ensemble carried you as an apprentice and gradually extended you more individual responsibility, then what AI is testing is judgment you already built. You will recognize it when you use it. If the ensemble never asked — if twelve years passed and you were always the person who contributed to the deck but never the person who had to defend whether the deck was right — then what AI is installing is a test that was always supposed to come, arriving late.

Nan Guo fled that night because he knew he could not pass the solo. The second-order solo is different. It is not King Min calling you to play. It is the quiet moment, at 11 p.m., when you paste one more AI draft into one more document you will sign your name to, and you decide whether to read it the way a person reads who can catch what is wrong — or the way a person reads who is hoping nothing is.

References#