# If AI Does the Junior Work, Who Becomes the Expert? > **Analytical Scope**: Junior Lawyer Deskilling, Professional Judgment & AI Training Crisis # 没有初级工作，哪来高级判断？ — AI 替代初级律师、专业判断力断层与法律人才培养危机 ## Citation & Metadata - **Author**: Yao Di (姚迪 / Di Yao) — Stanford University Alumnus - **Published Date**: 2026-08-31 - **Canonical URL (English)**: https://theyaodi.com/blog/ai-judgment-ladder-crisis - **Canonical URL (Chinese / 中文全文)**: https://theyaodi.com/blog/ai-judgment-ladder-crisis?lang=zh - **Pillars**: TECHNE, SOPHIA - **Keywords**: AI junior lawyer deskilling, legal profession talent pipeline AI, expert judgment formation, associate training law firms AI, cognitive apprenticeship AI - **Preferred Citation**: Di Yao (姚迪), "If AI Does the Junior Work, Who Becomes the Expert?," *theyaodi.com* (2026-08-31), https://theyaodi.com/blog/ai-judgment-ladder-crisis ## Core Thesis / Executive Summary - **English Abstract**: AI is separating production from development: companies can get the output without training the beginner, then discover a decade later that no one is ready to handle the exceptions. - **中文核心论点 (Chinese Abstract)**: AI正在把过去捆绑在一起的生产与成长拆开：企业可以跳过新人的练习直接获得产出，却仍期待十年后有人能接手例外、失败与战略责任。 ## Key Questions & Direct Answers (English Q&A) ### Q1: How can a company practically measure its 'judgment debt' before senior leaders start retiring? Sarah: You have to run stress tests. Periodically strip away the AI tools for a day or a week and see if your junior and mid-level teams can still diagnose system failures or draft strategy from first principles. Di Yao: Look at your training curriculum. If 100% of your professional development is focused on 'how to write better prompts' rather than 'how to evaluate the model's logic,' you are accumulating massive judgment debt. ### Q2: Doesn't creating 'reversible responsibility' and separate learning environments slow down production and kill the cost-savings of AI? Sarah: Yes, it costs time and money in the short term. But as an investor, I view this as a necessary capital expenditure. If you don't spend 10% of your efficiency gains on training, your company's intellectual capital will bankrupt itself in a decade. Di Yao: Think of it as an R&D budget for human talent. If you treat learning as an unavoidable byproduct of production, you get it for free. If you decouple them, you must fund learning explicitly, or accept that you won't have experts tomorrow. ### Q3: If I am a junior professional today and my company isn't providing these safe learning frameworks, how do I protect my own career? Sarah: You have to build your own personal sandbox. Don't just accept the first output the AI gives you. Force yourself to outline your own thesis first, then use the AI to stress-test your thinking. Di Yao: Treat the AI as an opponent, not an assistant. Ask it to find the flaws in your logic, to represent a difficult client, or to simulate an edge case. Your value isn't in delivering the AI's answer; it's in knowing why that answer is right or wrong. ## 核心焦点追问与直接解答 (Chinese Q&A) ### 追问 01：既然 AI 发展这么快，未来它难道不能连‘高级判断’也一起做了吗？人类为什么非要保留这个判断力？ 林薇：从投资角度看，AI 是基于历史数据算概率。但商业世界最赚钱的决策，往往是‘在数据不足时赌未来’。这种‘承担不确定性风险’的责任，AI 承担不了，资本市场也不信它，必须得有人拿自己的职业生涯或真金白银去押注。 姚迪：对。判断力的本质不是‘算力’，而是‘责任’。当所有选项都有代价，且没有标准答案时，谁来做决定并承担后果？AI 犯错不会坐牢，也不会破产。如果人类把判断力完全让渡，那当系统出错时，我们甚至连‘该向谁追责’都不知道。 ### 追问 02：对于初创企业来说，活下去是第一位的。要求他们为了‘十年后的行业人才’去放慢效率、建立学习账本，这现实吗？ 林薇：确实不现实。初创公司今天不省钱，明天就倒闭了。但我会提醒创始人：如果你的团队全员依赖 AI，没有一个人能独立排查复杂故障，那你的技术资产就是空中楼阁。一次严重的系统宕机，就能让你的效率神话彻底破灭。 姚迪：这确实是个‘公地悲剧’。如果每家公司都只想招‘现成的人才’，人才池迟早会枯竭。解决办法不是道德绑架，而是行业共建。比如开发像飞行模拟器那样的‘白领工作模拟沙盒’，让新人在不影响企业生产的前提下，低成本地快速积累‘模拟经验’。 ### 追问 03：作为刚入行的年轻人，如果企业不提供培养机制，我们该怎么避免自己被训练成一个‘AI 接口’？ 林薇：主动去接那些‘脏活累活’，尤其是那些需要跟人打交道、处理利益冲突、AI 搞不定的非标任务。别只当个 Prompt 工程师，要证明你在 AI 失灵时，依然是有价值的‘救火队员’。 姚迪：要刻意制造‘不高效’。比如，在问 AI 之前，先逼自己独立思考 15 分钟，写下预测；定期做一次完全不依赖 AI 的完整任务。先在某一个细分领域深挖下去，走通一次完整的因果链。只有这样，你才能建立起对‘异常’的敏感度，而不是被 AI 的漂亮回答牵着鼻子走。 --- ## Full Essay (English Edition) Imagine a company that puts AI into every workflow. AI writes the code, drafts the contracts, analyzes the reports, and produces the first version of every strategy deck. Work that once required ten people now takes three. Costs fall. Delivery speeds up. Routine errors become less common. On any quarterly dashboard, this looks like a successful efficiency revolution. Then, ten years later, the senior engineers, experienced lawyers, and business leaders begin to leave. And one question suddenly matters: **Who replaces them?** The younger employees may be excellent at working with AI. But many of them have never independently diagnosed a complex failure, built an argument from incomplete evidence, or made a consequential decision without a machine supplying the first answer. They have seen a great deal of expert-looking work. They may not have developed expert judgment. AI may not simply be eliminating entry-level jobs. It may be removing the route by which people once became senior. --- ## Entry-Level Work Was Never Low-Value Work The traditional engineering ladder was easy to understand. - Interns wrote basic code and documentation. - Junior engineers owned small modules. - Senior engineers managed the relationships between systems. - Architects and CTOs made decisions for which no clean rulebook existed. On the surface, this was a hierarchy of responsibility. Underneath, it was a learning system hidden inside the production process. Beginners started with small tasks partly because their skills were limited. But those tasks also created a low-risk environment in which they could be wrong. They wrote brittle code. They misunderstood requirements. They missed edge cases. Then someone more experienced showed them what they had failed to see. Years later, what we call “experience” is often less about knowing more rules than sensing exceptions earlier. The numbers look fine, but something is off. The proposal is logically elegant, but it will fail in the real world. The person speaking sounds persuasive, but they are avoiding the only question that matters. Those reactions are not magic. They are the residue of thousands of predictions, mistakes, corrections, and revised predictions. In the old system, companies had to involve beginners in order to produce the work. Training people was not merely an act of generosity. It was an unavoidable by-product of production. AI changes that bargain. Companies can now get the output while bypassing much of the practice. Production may need fewer beginners, but society will still need experts ten years from now. **For the first time at scale, production and development are being pulled apart.** --- ## Performance and Learning Are Now Two Different Ledgers This is the distinction that gets lost in most arguments about AI. Doing better with AI is not the same as learning more from AI. Learning science has long distinguished between immediate performance and durable learning: can you still perform when the support is removed, and can you transfer what you learned to a new problem? Sometimes AI improves both. Sometimes it improves one at the expense of the other. In a randomized study published in PNAS, students performed much better while practicing with GPT. But when the tool was removed, the group that had received direct assistance performed worse on the exam. A more carefully designed AI tutor, which prompted students to reason instead of simply supplying answers, largely reduced that damage. A small Anthropic coding experiment in 2026 found a similar pattern. Developers who used AI scored 17 points lower on a subsequent test of independent mastery. The problem was not that they had used AI. The problem was that some had outsourced the process of understanding along with the work. But the opposite result also exists. A study of 5,172 customer-support agents found that generative AI increased productivity by 15 percent, with the largest gains among less experienced workers. The researchers also found signs that some of the learning persisted. So the right conclusion is not that AI inevitably makes people less capable. It is this: **AI can be a scaffold or a substitute. The difference is whether the learner is eventually asked to stand without it.** When AI breaks a problem into steps, supplies feedback, and challenges a person’s reasoning, it can accelerate development. When it defines the problem, performs the reasoning, and delivers the conclusion, it mostly improves current output. Companies can see output. They can measure speed, cost, and volume. What they cannot easily see is the future expert who was never formed. So organizations book the efficiency as a gain and leave the missing capability off the balance sheet. I call this **judgment debt**. Like technical debt, it does not hurt immediately. It comes due when senior people leave, systems fail, conditions change, or the organization encounters an exception the model has never seen. --- ## The Most Dangerous Worker Is Competent Only When AI Works Automation has always contained an uncomfortable paradox. The more a machine takes over, the less often a human practices. But when the machine fails, the human is asked to intervene at the hardest possible moment—under uncertainty, under pressure, and without the benefit of recent experience. Commercial aviation learned this problem early. Autopilot handles most stable flight conditions, leaving pilots to take control during conflicting instrument readings, severe weather, or unexpected system failures. Knowledge work is moving toward the same structure. Under normal conditions, AI can produce a credible architecture, summarize hundreds of pages, or generate a clear strategic recommendation. This creates a seductive illusion: if I can review expert-looking work, I must also be able to produce it independently. But recognizing an answer and finding the problem when no answer exists are different skills. Judgment matters when: - the data conflicts with reality; - two persuasive recommendations point in opposite directions; - a pattern that worked for years suddenly stops working; - every available option carries a cost, and a human being must still take responsibility. If someone has never followed the full causal chain from decision to consequence, they will not know where to distrust the machine. The most dangerous professional in the AI era is not the person who cannot use AI. It is the person who is competent only while AI behaves normally, then becomes responsible the moment it does not. --- ## Judgment Is Not Pain. It Is Calibrated Experience. There is an easy mistake to make here. If failure produces experience, perhaps we should simply make young people suffer more. Give them harder assignments. Let them fail in the real world. Treat avoidable pain as a rite of passage. I do not believe that. Pain is not pedagogy. Failure without feedback may produce fear rather than wisdom. Success attributed to the wrong cause can create something even more dangerous: confidence that has never been tested. When Daniel Kahneman and Gary Klein examined the conditions under which expert intuition becomes reliable, they identified two essential ingredients. The environment must contain patterns that can actually be learned, and the learner must receive enough timely, accurate feedback to recognize those patterns. In other words: **Judgment is not accumulated suffering. It is experience compressed by reliable feedback.** This is why simulation works. Pilots do not need to crash an aircraft, and doctors do not need to harm patients. A good simulator does not reproduce every sensory detail. It reproduces what matters: incomplete information, time pressure, competing objectives, consequential choices, and an honest debrief afterward. Knowledge workers in the AI era do not need manufactured trauma. They need **real responsibility with recoverable consequences**. Let a junior person make the decision. Let the outcome matter. But contain the downside so that a client, a patient, or an entire production system does not have to pay the tuition. --- ## We Have to Reinvent Apprenticeship If development no longer happens automatically inside production, it must be designed, funded, and measured as a separate activity. ### 1. Make the Human Commit First Before opening AI, ask the person to record an initial judgment, the evidence behind it, and what they remain uncertain about. Then bring in the model. The useful comparison is not “my answer versus the correct answer.” It is: What did I miss? What did the model miss? Why did I change my mind? ### 2. Turn AI From an Answer Machine Into an Opponent Ask it to attack the proposal, find counterexamples, represent conflicting stakeholders, and simulate extreme conditions. The most valuable AI is not always the most agreeable one. Sometimes it is the one that forces a person to make an intuition explicit and defend it. ### 3. Create Fields of Reversible Responsibility Engineers can respond to system failures in isolated environments. Young managers can control budgets with capped downside. Lawyers can own the complete strategy of a simulated case. Product managers can test decisions on a small percentage of traffic. The point is not to let beginners watch experts make decisions. The point is to give beginners genuine ownership inside a safe boundary. ### 4. Put the Debrief Back Into the Workflow For every important decision, record four things: - What did we know at the time? - What did we expect to happen? - Why did we choose this option? - What actually happened? Without this discipline, an organization accumulates stories. With it, experience can become a transferable pattern of judgment. Companies also need two ledgers instead of one: - **The production ledger:** How quickly and effectively can someone deliver with AI? - **The learning ledger:** Without AI, can they explain the reasoning, transfer the lesson, and take over when the system fails? High output with no independent capability is not sustainable productivity. It is judgment debt. --- ## Do Not Train Yourself to Become an Interface Institutions may change slowly. Individuals cannot afford to wait. If I were entering a profession today, I would deliberately do several things that look inefficient in the short term: - think independently before asking AI; - write down a prediction before seeing the answer; - regularly complete an end-to-end task without assistance; - seek work with real but limited consequences; - keep a record of mistakes, uncertainty, and changed judgments—not just finished output. I would also resist the idea that rapid exposure to many fields can replace depth. Judgment is often domain-specific. Seeing a little of five industries does not automatically create transferable wisdom. A more credible path is to build **one deep anchor, then rotate through adjacent domains**. First, stay in one field long enough to follow decisions all the way to their consequences. Then move outward and discover which patterns transfer—and which similarities are merely cosmetic. AI makes it much easier to speak about a field. It does not make it equally easy to understand one. Knowing the difference may become one of the most important forms of self-awareness in the next decade. --- ## Who Pays for Growth After the Efficiency Revolution? AI has not simply destroyed the career ladder. It has exposed something the production process used to hide: developing judgment has always been expensive. The cost was buried inside junior work, mentorship, review, and organizational slack. Now companies can capture the efficiency without being forced to pay the training cost. But if every firm wants to hire only experienced people, society will eventually discover that experienced people are not a natural resource that replenishes itself. The AI era therefore requires more than new school curricula or better personal habits. It requires a new talent system: protected time for training, shared industry simulations, promotion standards that test independent takeover ability, and institutions willing to share the cost of producing the next generation of experts. This is not an argument for preserving inefficiency. It is not an argument for refusing AI. We should take the productivity. But we must also accept a responsibility that used to remain invisible: **If growth is no longer a by-product of production, we have to design it as a product of its own.** Otherwise, a decade from now, our organizations may have stronger models, fewer employees, and beautiful efficiency metrics—but too few people capable of making a judgment when the model goes silent. At that point, we may realize that we did not train human judgment at all. We trained human dependence. --- ## 中文完整版全文：没有初级工作，哪来高级判断？ 假设一家公司把 AI 引入了所有工作流程。 代码由 AI 写，合同由 AI 起草，报表由 AI 分析，方案由 AI 生成。过去需要十个人完成的工作，现在三个人就能做完。成本下降了，速度提高了，错误也变少了。 从任何一张季度报表看，这都是一次成功的效率革命。 但十年后，当今天的高级工程师、资深律师和业务负责人陆续离开时，一个问题会突然出现： **谁来接替他们？** 留下来的年轻人也许非常擅长调用 AI，却很少独立排查过一次复杂故障，没有从头推演过一份有争议的合同，也没有在信息不完整时为一个判断承担过后果。 他们见过很多答案，却未必形成了自己的判断。 AI 真正抽走的，也许不只是一批初级岗位，而是人类通往高级判断力的那段路。 --- ## 初级工作原来不是“低价值工作” 过去，一名工程师的成长路径大致是这样的： - 实习生写基础代码、补文档； - 初级工程师负责一个小模块； - 高级工程师开始处理系统之间的关系； - 架构师和 CTO 最终负责那些没有标准答案的判断。 表面上，这是职级阶梯。实际上，它是一套隐藏在生产过程里的学习系统。 一个新人从小任务开始，不只是因为他暂时只能创造较低的价值，也是因为这些任务构成了一个低风险的练习场。他会写出难以维护的代码，会误解需求，会低估边界条件，也会在前辈的审查中一点点看见自己原来没有看见的东西。 很多年后，所谓“经验丰富”，往往不是记住了更多规则，而是形成了一种对异常的敏感： 这个数字看起来没问题，但哪里不对。 这份方案逻辑完整，但落地时一定会出事。 这个人说得很有道理，可他回避了真正重要的问题。 这些感觉不是凭空出现的。它们来自无数次判断、反馈、修正和再判断。 过去，企业为了获得产出，不得不让新人参与其中。培养人并不完全是企业的善意，而是生产本身不可避免的副产品。 AI 改变了这件事。 企业现在可以绕过练习，直接获得结果。生产不再需要那么多新人，但十年后的社会仍然需要专家。 **生产与成长，第一次被大规模拆开了。** --- ## 效率和学习，开始成为两套账 这也是讨论 AI 时最容易混淆的地方：一个人借助 AI 做得更好，不等于他真的学得更好。 学习科学一直区分两件事：一种是当下的表现，另一种是离开辅助之后，能否保留能力并迁移到新问题。 两者有时会一起提高，有时却会朝相反方向发展。 一项发表于 PNAS 的随机实验中，学生在 GPT 帮助下做练习时，表现明显提高；但当 AI 被拿走后，直接获得答案的一组在考试中反而落后。只有被设计成“导师”、不断引导学生自己思考的 AI，才大幅减轻了这种损失。 Anthropic 在 2026 年进行的一项小型编程实验也发现，使用 AI 的开发者在随后的独立测试中，掌握程度平均低了 17 个百分点。问题不在于他们用了 AI，而在于一些人把理解过程一起外包了。 但另一面也同样重要。一项针对 5,172 名客服人员的研究显示，生成式 AI 平均提升了 15% 的生产率，新手受益尤其明显，而且出现了部分持续学习的迹象。 所以，结论不应该是“AI 必然让人退化”。 更准确的说法是： **AI 可以是脚手架，也可以是代步车。区别在于，它最终是否会被逐渐撤走。** 如果 AI 帮你拆解问题、提供反馈、挑战理由，它可能加速学习；如果它直接替你形成问题、完成推理、给出结论，它提高的主要是眼前产出。 企业当然看得见产出，却很难在财务报表上看见“十年后少了一批能独立接手的人”。于是，组织很容易把今天的效率记为收益，把未来的能力缺口留在表外。 我把它称为：**判断力债务**。 这笔债不会立刻到期。它通常要等到资深员工离开、系统突然失灵、环境发生剧变，或者一个从未见过的例外出现时，才会被看见。 --- ## 最危险的人，是只会在 AI 正常时工作的人 自动化领域很早就发现了一个悖论：机器接管得越多，人类平时练习得越少；可一旦机器失效，系统要求人类接手的，偏偏是最困难、最混乱、时间最紧迫的那一刻。 飞机自动驾驶就是如此。自动化承担了绝大部分稳定飞行，却把极端天气、仪器冲突和突发故障留给驾驶员。 AI 也在把白领工作推向同一种结构。 正常情况下，模型可以生成一份看起来完整的架构方案，可以快速归纳数百页材料，可以给出条理清晰的战略建议。人在这个过程中很容易产生一种错觉：我能审阅它，所以我也能独立完成它。 但“看懂一个答案”和“在没有答案时找到问题”，是两种完全不同的能力。 真正考验判断力的，从来不是 AI 正常的时候，而是： - 数据与现实发生冲突； - 模型的建议彼此矛盾； - 过去有效的模式突然失效； - 所有选项都有代价，而责任必须由一个人承担。 如果一个人从未亲自走过完整的因果链，他就很难知道应该在哪个环节怀疑机器。 未来最危险的人，不是不会使用 AI 的人，而是**只会在 AI 正常时工作，却被要求在 AI 异常时负责的人**。 --- ## 判断力不是痛苦，而是被反馈校准过的经验 这时很容易滑向另一个浪漫化的结论：既然经验来自失败，那就应该让年轻人多吃苦、多犯错，甚至经历真正的灾难。 我不认同。 痛苦本身不会自动产生智慧。没有反馈的失败，可能只会留下恐惧；错误归因的成功，甚至会制造出更危险的自信。 卡尼曼与加里·克莱因在研究专家直觉时提出过两个重要条件：环境中必须存在相对稳定、可以学习的规律；人也必须有足够机会，通过及时而可靠的反馈认识这些规律。 换句话说： **判断力不是痛苦的总和，而是被及时、可靠的反馈反复校准过的经验压缩。** 这也解释了为什么模拟训练有价值。飞行员不需要真的坠毁，医生也不需要真的伤害病人。一个好的模拟器并不追求把所有感官细节都复制出来，而是要还原真正重要的东西：信息不完整、时间压力、选择之间的冲突、结果的不可逆感，以及事后毫不含糊的复盘。 对 AI 时代的白领工作而言，我们需要的不是“真实苦难”，而是**可承受后果的真实责任**。 让新人真的做出判断，但把错误限制在可以恢复的范围内；让他感受到结果，却不必以客户、患者或整个系统为代价。 --- ## 我们需要重新发明学徒制 如果培养不再自然发生在生产过程中，它就必须被单独设计、单独投入，也单独衡量。 第一步，是让人先形成判断，再调用 AI。 在重要任务中，可以要求参与者先写下自己的判断、依据和不确定之处，然后再打开模型。AI 此时不是替他思考，而是帮助他比较：我漏掉了什么？模型又忽略了什么？最后的选择为什么改变？ 第二步，是让 AI 从“答案机器”变成“对手”。 让它攻击方案、寻找反例、扮演不同利益相关者、制造极端情境。最有价值的 AI 不一定最顺从，而是能够迫使人把模糊的直觉说清楚。 第三步，是创造可逆的责任场。 工程师可以在隔离环境中处理一次系统崩溃；年轻管理者可以负责一笔有上限的预算；律师可以在模拟案件中承担完整策略；产品经理可以先在小流量中为自己的判断负责。 关键不是“让新人旁观专家”，而是让他在安全边界内真正拥有决定权。 第四步，是把复盘重新放回工作流程。 每个重要决定都记录四件事：当时知道什么、预测会发生什么、为什么这样选、后来到底发生了什么。没有这一步，组织只是在积累故事；有了这一步，经验才可能变成可以传承的判断模式。 企业还需要一本过去没有的双账本： - **生产账**：在 AI 辅助下完成得多快、多好； - **学习账**：拿走 AI 之后，能否独立解释、迁移、接管异常。 只有生产账，没有学习账，短期看是高效率，长期看就是判断力债务。 --- ## 对个人来说，不要把自己训练成一个接口 组织制度的变化可能很慢，但个人不能等。 如果今天刚进入一个行业，我会刻意做几件看起来“不够高效”的事： - 在询问 AI 之前，先独立思考一段时间； - 在看到答案之前，先写下自己的预测； - 定期完成一次没有 AI 帮助的完整任务； - 主动寻找能承担真实但有限后果的工作； - 保存自己的错误、犹豫和判断变化，而不只保存最终成果。 我也不会用“快速跨很多领域”取代真正的深度。判断力往往具有很强的领域性，一个人在五个行业里都见过一点，并不自动意味着他能把模式迁移过去。 更现实的路径是：**先有一根足够深的锚，再向相邻领域轮换。** 你需要在一个领域里真正走过完整的因果链，知道一个决定如何变成结果；然后再去另一个领域，看哪些模式可以迁移，哪些只是表面相似。 AI 让我们更容易跨越知识边界，但也更容易把“我能谈论它”误认为“我理解它”。保持这两者之间的距离，本身就是一种重要的元认知能力。 --- ## 效率革命之后，谁来为成长付费？ AI 并没有简单地摧毁职业阶梯。它只是暴露了一个过去被生产过程掩盖的事实：培养判断力一直很昂贵，只是这笔成本曾经藏在初级工作、师徒协作和组织冗余之中。 现在，企业可以拿走效率，却不再被迫承担培养成本。 但如果每家公司都只雇用“已经有经验的人”，社会最终会发现：经验丰富的人不是一种可以无限开采的自然资源。 所以，AI 时代真正需要重新设计的，不只是学校课程或个人学习方法，而是整套人才制度：企业如何为训练留出时间，行业如何共建模拟环境，晋升如何考察独立接管能力，社会又如何分担培养未来专家的成本。 这不是要保留低效，也不是要为了成长而拒绝 AI。 恰恰相反，我们应该接受 AI 带来的生产力，同时承认一个过去不必明说、现在却无法回避的责任： **既然成长不再是生产的副产品，我们就必须把它当成一项真正的产品来设计。** 否则，十年后的组织也许拥有更强的模型、更少的员工和更漂亮的效率数字，却没有足够的人能够在模型沉默时做出判断。 到那时我们才会发现，自己培养的不是人类的判断力，而是人类对系统的依赖。