# Why Is AI Transforming Coding Faster Than Legal Work?
> **Analytical Scope**: Formal Verification, RLVR & Legal AI vs. Software Engineering
# 为什么 AI 改变编程比改变法律工作更快？ — 形式化验证、可验证奖励与法律 AI 演进速度对比

## Citation & Metadata
- **Author**: Yao Di (姚迪 / Di Yao) — Stanford University Alumnus
- **Published Date**: 2026-09-16
- **Canonical URL (English)**: https://theyaodi.com/blog/ai-coding-law-verifiability
- **Canonical URL (Chinese / 中文全文)**: https://theyaodi.com/blog/ai-coding-law-verifiability?lang=zh
- **Pillars**: LEX, TECHNE, SOPHIA
- **Keywords**: AI coding vs legal work, law and coding, legal coding, verifiability in AI, RLVR reinforcement learning verifiable rewards, legal AI benchmarks, compiler verification vs legal judgment
- **Preferred Citation**: Di Yao (姚迪), "Why Is AI Transforming Coding Faster Than Legal Work?," *theyaodi.com* (2026-09-16), https://theyaodi.com/blog/ai-coding-law-verifiability

## Core Thesis / Executive Summary
- **English Abstract**: Why has AI already reshaped coding while its impact on legal work seems less dramatic? The key may be how easily each field can verify its results.
- **中文核心论点 (Chinese Abstract)**: 最近和朋友聊到一个问题：为什么 AI 在 coding 上已经带来结构性变化，而在 legal 上还没有那么明显？我觉得关键不在于谁更‘聪明’，而在于谁更容易被验证。

## Key Questions & Direct Answers (English Q&A)

### Q1: Why is 'verifiability' the ultimate bottleneck for AI adoption in non-technical fields?
Sarah: If you can't cheaply and automatically verify if an AI's output is correct, a human expert has to review every single line. That completely erases the cost and speed benefits of using AI in the first place. Di Yao: In software, the compiler acts as a free, instant auditor. In law, medicine, or strategic consulting, the 'compiler' is a real-world crisis, a lawsuit, or a market failure. The cost of a bad verification there is catastrophic.

### Q2: How can legal departments prepare for the 'hollowing out' of junior roles?
Sarah: Law firms and corporate legal departments need to completely abandon the billable hour model for junior tasks and shift toward active apprenticeship models, where juniors shadow senior partners on judgment-heavy tasks early on. Di Yao: We have to treat judgment as a deliberate, structured curriculum. Since juniors won't be spending thousands of hours doing manual document review anymore, we must design simulated crisis scenarios to accelerate their decision-making experience.

### Q3: What does 'context reconstruction' mean for the limits of LLMs?
Sarah: LLMs only know what is explicitly written down in their context window or training data. They are blind to the 'dark data' of human organizations—the unwritten alliances, the CEO's unstated priorities, or the mood of a regulator. Di Yao: Yes, senior advisory work is less about information retrieval and more about reading between the lines. Until AI can capture the tacit, relational knowledge embedded in human networks, its strategic advice will remain shallow.

## 核心焦点追问与直接解答 (Chinese Q&A)

### 追问 01：既然编程更容易形成闭环、更容易被 AI 替代，那程序员是不是会比律师更早面临大规模失业？
林薇：从投资角度看，是的。未来一个项目可能只需要一个懂业务的 PM 和几个 AI Agent，纯粹靠写代码谋生的初级程序员，生存空间会被极速压缩。 姚迪：我不这么悲观。编程门槛降到零，意味着软件需求会呈指数级爆发。程序员不会消失，而是会转型为‘系统架构师’和‘问题定义者’，他们要管理成百上千个 AI 帮手。

### 追问 02：你们说法律面对的是复杂的人性，但现在 AI 已经可以通过大数据预测法官的判决倾向，准确率甚至超过人类，这难道不算‘可验证’吗？
林薇：这绝对算验证！数据不会撒谎。只要有足够多的历史判例，AI 就能算出胜率。在诉讼策略上，AI 很快就会比人类律师更精准。 姚迪：预测不等于说服。法官的判决不仅看历史数据，还要考虑当下的社会舆论、政策风向甚至庭审时的情感共鸣。AI 能预测概率，但无法在法庭上进行有温度的辩护和利益博弈。

### 追问 03：如果初级律师的‘生产性工作’被 AI 掏空，律所不招新人了，那这个行业的传承不就断裂了吗？
林薇：律所必须重构商业模式。不能再按小时计费了，而是要按‘结果’和‘判断的价值’收费。律所甚至需要建立‘模拟法庭’或‘虚拟案例库’来人工加速训练新人。 姚迪：没错，这是教育体制的倒逼。未来的法学教育不能再教学生怎么背法条、写格式合同，而必须从第一天起，就训练他们如何在模糊、复杂的商业冲突中做决策和承担责任。

---

## Full Essay (English Edition)

I recently had a conversation with a friend about a question that I find increasingly interesting: why has generative AI already produced such visible, and arguably structural, change in engineering—especially in coding—while the impact of AI on legal work still feels comparatively incremental?

One intuitive answer is that coding may simply be easier for AI to replace, while many parts of legal work still require things machines cannot do.

There is some truth in that. But I do not think the deepest distinction is that programmers do simpler work than lawyers, or that law somehow requires more intelligence than software engineering.

The more fundamental difference may be this:

**Coding is much easier to turn into a closed loop in which AI can act, test its own output, detect failure, and correct itself.**

That may be one of the main reasons why the AI revolution has arrived in software engineering first.

## 1. The key advantage of coding is not simplicity. It is verifiability.

A coding agent today can increasingly perform an entire sequence of work:

read a repository, understand an issue, change the code, run the program, execute tests, detect errors, revise the code, test again, and eventually submit the result.

The workflow looks roughly like this:

**Generate → Run → Test → Fail → Fix → Run Again**

The most important feature is not that the AI produces perfect code on the first try.

The most important feature is that **the system itself can tell the AI that it is wrong.**

Does the code compile?

Do the tests pass?

Does the API return the expected result?

Does performance stay within the required threshold?

These questions often generate relatively clear feedback.

As a result, even if the coding agent fails on the first attempt, it can continue iterating until the system accepts the result.

Software engineering, in other words, provides an unusually strong feedback loop.

Legal work does not.

Suppose an AI system advises an in-house lawyer:

“In this situation, the company should consider making a voluntary disclosure to the regulator.”

How do you test that?

There is no button that runs a legal-advice test and returns either “PASS” or “FAIL.”

The correctness of legal advice is often not binary.

It may depend simultaneously on black-letter law, enforcement practice, regulatory probability, corporate risk appetite, government relations, business interests, reputation, global policy, internal precedent, and future developments.

Two excellent lawyers looking at exactly the same facts may reach different recommendations.

And neither may necessarily be wrong.

That creates one of the most important structural constraints on legal AI:

**Generation is becoming cheap. Verification remains expensive.**

AI is becoming increasingly capable of producing a legal answer.

But deciding whether that answer should be trusted still requires costly human judgment.

## 2. Software engineering and legal work operate in fundamentally different environments

There is another important difference: the outputs of coding and legal work ultimately face different kinds of systems.

Code ultimately has to face a machine.

Machines are relatively honest.

They will not accept a broken API because you have a good relationship with them.

They will not reinterpret a piece of Python six months later because the political environment has changed.

Code must eventually run.

That means a large portion of engineering output can be evaluated through system behavior.

Legal work ultimately faces something very different:

**people and institutions.**

How will a regulator interpret the issue?

How will a judge decide?

Will the counterparty accept the position?

How much risk is the board willing to bear?

How will the public react?

Will the same government agency take the same view six months from now?

There is no stable runtime environment for these questions.

So the distinction can be stated quite simply:

**Coding operates in a machine-native environment.**

**Legal work operates in a human-and-institution-native environment.**

Those two environments have very different levels of AI-friendliness.

## 3. Legal problems often begin before the objective is even clearly defined

Many coding tasks are relatively easy to specify.

For example:

build an API endpoint;

accept a user ID;

query the database;

return structured data;

keep response time below 200 milliseconds;

add automated tests.

This does not mean the task is easy.

It means that “done” can usually be described with reasonable precision.

Legal work is often very different.

An in-house lawyer may receive a question such as:

“The government called this morning. What do you think we should do?”

Behind that sentence may sit a dozen separate questions.

Is this a formal investigation or an informal inquiry?

Why are they asking now?

Has the company handled a similar situation before?

Does headquarters know?

Do we want to cooperate?

Will cooperation create a precedent?

Will refusal lead to escalation?

What does Chinese law require?

Is there a conflict with U.S. law?

What does the business actually want to achieve?

Who should speak with the government?

Much of the most important context may not exist in any database.

It may exist in someone’s head.

It may exist in organizational history.

It may come from an understanding of a particular regulator.

It may be embedded in tacit knowledge built up over the last ten meetings.

Senior legal work is therefore often not primarily an information-retrieval problem.

It is a problem of:

**context reconstruction.**

And that is still something AI has difficulty obtaining in full.

## 4. Legal work is often not about finding the answer. It is about deciding what the answer should optimize for.

AI is particularly strong at a certain type of problem:

**Given X, find Y.**

But much of senior legal work is closer to:

**What should X be?**

Consider a company asking:

“Can we launch this product?”

At a relatively junior level, the legal question may be:

“Does the law permit it?”

At the GC or senior-counsel level, the real question is often:

“How much risk are we willing to take across law, regulatory relationships, commercial opportunity, brand impact, and future precedent?”

At that point, the task is no longer purely legal reasoning.

It contains:

judgment,

value choices,

organizational authority,

and accountability.

AI can identify the options.

It can explain the risks associated with each one.

It may even be able to simulate likely regulatory responses.

But someone still has to say:

**We are willing to take this risk.**

That act is fundamentally one of agency.

This is why I increasingly think that legal AI cannot be understood only as a question of intelligence.

It must also be understood as a question of authority and responsibility.

Who has the authority to decide?

Who bears the consequences?

Who is accountable when things go wrong?

## 5. None of this means that legal work will be protected from AI

Quite the opposite.

I think the legal industry may ultimately experience enormous disruption.

But the path will not look exactly like software engineering.

Many legal tasks are already highly susceptible to automation or compression.

For example:

clause extraction;

document comparison;

due diligence;

case chronology;

standard legal research;

regulatory mapping;

first drafts;

policy comparison;

discovery review;

large volumes of standardized contract work.

What do these tasks have in common?

**The input is relatively bounded.**

**The output is relatively specified.**

**The relevant sources are relatively identifiable.**

**Verification is relatively easy.**

In other words, they increasingly resemble coding.

So the better question is not:

“Will AI replace legal work?”

It is:

**How code-like is this legal task?**

The more a legal task can be reduced to clear inputs, clear outputs, verifiable results, and iterative correction, the faster AI will penetrate it.

By contrast, the more a task depends on relationships, ambiguous goals, institutional power, judgment, and irreversible consequences, the slower autonomous AI execution will be.

## 6. A simple framework for AI substitutability

One way to think about this is to evaluate any task across several dimensions.

**Formalizability**

Can the task be clearly specified?

**Verifiability**

Can the result be checked cheaply and reliably?

**Digital availability**

Is the relevant context and data already digitized and accessible?

**Reversibility**

If the system makes a mistake, can the error be easily rolled back?

And finally:

**Judgment density**

How much normative, institutional, interpersonal, or commercial judgment is required?

Coding often has:

high formalizability;

very high verifiability;

very high digital availability;

relatively high reversibility;

and moderate judgment density.

That makes it unusually receptive to AI.

Legal work often looks very different.

Of all these variables, I suspect that the most important is:

**verifiability.**

Whether AI can replace a person does not depend primarily on whether the AI gets the answer right on the first attempt.

It depends on whether, when AI gets the answer wrong, **the system can tell it that it is wrong.**

That is the great advantage of coding.

And it is one of the great constraints on legal AI.

## 7. The deeper risk may not be that lawyers disappear, but that the production layer of the profession gets hollowed out

Traditionally, the legal profession has looked something like a pyramid:

senior lawyers;

mid-level lawyers;

junior lawyers;

paralegals.

Junior lawyers perform large amounts of research, review, drafting, and due diligence.

Through that work, they gradually develop judgment.

AI may change the structure.

Senior lawyers may increasingly focus on judgment, agency, and relationships, while a large AI layer performs much of the production work below them.

Research, drafting, review, comparison, and summarization could increasingly be handled by machines.

That means AI may not eliminate senior lawyers first.

It may first compress the production work traditionally performed by junior and mid-level lawyers.

But this creates a much deeper problem:

**If judgment was historically learned through production, where will future senior judgment come from once machines perform most of that production?**

A lawyer may historically have developed intuition by reading one hundred cases, revising fifty contracts, and writing thirty memoranda.

If young lawyers increasingly spend their time merely approving AI output, will they lose the opportunity to develop the instincts that senior judgment requires?

That suggests that the long-term challenge of legal AI may not be:

“How do we replace lawyers?”

It may instead be:

**How do we train human judgment in a world where machines perform most of the production?**

I do not think the profession has fully confronted that question yet.

## 8. Engineering may simply be one generation ahead of law

Coding is already moving from:

“Help me write this code.”

to:

“Here is the problem. Go solve it.”

That is the transition from assistance to agency.

Legal AI is still, in most cases, operating at an earlier stage:

help me research;

help me summarize;

help me draft;

help me review;

help me analyze.

The real legal AI revolution may begin when the instruction becomes:

**“Here is the matter. Run the workflow.”**

Imagine a future legal agent that can:

receive a regulatory request;

find relevant historical matters inside the company;

research the applicable law;

retrieve internal policy;

identify relevant stakeholders;

generate a risk matrix;

draft the response;

recommend whether escalation is required;

update the matter-management system;

track deadlines;

and finally submit the package to a lawyer for approval.

At that point, legal work would truly begin to enter the agentic era.

So perhaps software engineering and legal work are not following fundamentally different paths.

Engineering may simply be showing us part of the future earlier.

## 9. What AI is really changing is not who is smarter, but what becomes cheap

Historically, high-quality production was expensive.

Code was expensive.

Legal research was expensive.

Contract drafting was expensive.

Analysis was expensive.

AI is making production cheaper.

And when production becomes cheap, professional value moves elsewhere.

In engineering, value may increasingly shift toward architecture, product judgment, system design, and orchestration.

The same may happen in law.

Once research, drafting, comparison, and summarization become inexpensive, the scarce resources may increasingly become:

**judgment, agency, trust, and accountability.**

That is why I increasingly find the question “Will AI replace lawyers?” too narrow.

A better question may be:

**When answers become cheap, what remains worth having humans do?**

My tentative answer is this:

AI will become increasingly good at producing answers.

But humans will still have to decide:

which answers deserve to be trusted;

which risks are worth taking;

which objectives are worth pursuing;

and ultimately, where we want those answers to take the organization.

That may be the deeper transformation of professional work in the AI era:

**AI makes production cheap. Human value moves toward judgment and agency.**

---

## 中文完整版全文：为什么 AI 改变编程比改变法律工作更快？

最近和朋友聊天，聊到一个很有意思的问题：为什么生成式人工智能对工程领域，尤其是编程，已经带来了如此明显、甚至可以说是结构性的变化，而在法律行业，人工智能虽然同样很热，但似乎还没有产生同等程度的“翻天覆地”？

一个直觉性的回答是：也许编程更容易被人工智能取代，而法律工作中还有很多事情机器做不了。

这个判断有一定道理，但如果继续往下想，我觉得真正的区别并不是“程序员的工作比律师简单”，也不是“法律比代码更需要智力”。

更根本的区别可能是：

**编程比法律工作更容易形成一个人工智能可以自己完成、自己检查、自己纠错的闭环。**

而这件事，可能决定了为什么人工智能革命首先在软件工程领域发生。

## 一、编程最大的优势，不是容易，而是可以验证

今天，一个编程智能体已经可以完成这样的流程：

读取代码库，理解任务，修改代码，运行程序，执行测试，发现报错，重新修改，再测试，最后提交。

它的工作流程大致是：

**生成 → 运行 → 测试 → 失败 → 修复 → 再运行**

这里最重要的，并不是人工智能第一次生成的代码有多好。

真正重要的是，**系统本身会告诉人工智能，它错了。**

代码能不能编译？

测试有没有通过？

接口有没有返回预期结果？

性能有没有超过设定阈值？

这些问题往往存在相对明确的反馈。

所以，一个编程智能体即使第一次做错，也可以不断试错、迭代，直到通过测试。

换句话说，软件工程天然提供了一个非常强的反馈闭环。

而法律工作恰恰缺少这样的机制。

比如，一个人工智能给公司法务提出建议：

“在这种情况下，我们可以考虑向监管机构主动披露。”

问题来了。

怎么测试？

没有一个按钮可以运行“法律建议测试”，然后返回“通过”或者“不通过”。

法律建议的“正确”往往不是二元的。

它可能同时涉及法律文本、监管实践、执法概率、公司风险偏好、政府关系、商业利益、声誉、全球政策、内部先例，以及未来可能发生的事情。

两个同样优秀的律师，面对同一组事实，完全可能提出不同的建议。

而且两个人都未必是错的。

这就构成了法律人工智能最核心的结构性难题：

**内容生成正在变得越来越便宜，但验证仍然昂贵。**

人工智能越来越容易生成一个法律答案。

但确认这个答案是否值得依赖，仍然需要昂贵的人类判断。

## 二、软件工程和法律工作面对的是两个不同的世界

如果进一步比较，会发现编程和法律工作的最终产出本身就非常不同。

代码最终必须面对机器。

机器相对“诚实”。

它不会因为你和它关系好，就接受一个无法工作的接口。

也不会因为政治环境变化，就突然重新解释一段程序。

代码最终会被执行。

所以，大量工程工作的结果可以通过系统行为来检验。

但法律工作最终面对的是：

**人和制度。**

监管机关会怎么理解？

法官会怎么判断？

交易对手会不会接受？

董事会愿意承担多少风险？

公众会怎么看？

某个政府部门今天和六个月后是不是同一个态度？

这些问题没有稳定的运行环境。

所以可以这样概括：

**编程运行在一个以机器为核心的环境中。**

**法律工作运行在一个以人和制度为核心的环境中。**

这两种环境对人工智能的友好程度差别很大。

## 三、还有一个容易被忽略的差别：法律问题经常连目标都没有定义清楚

很多编程任务的目标相对容易明确。

比如：

建立一个接口；

输入用户编号；

查询数据库；

返回结构化结果；

响应时间低于两百毫秒；

同时加入自动测试。

这不是说任务简单。

而是说，“完成”是什么，通常比较容易描述。

但公司律师收到的问题经常是：

“政府今天打电话来了，你觉得我们应该怎么办？”

这句话背后可能隐藏着十几个问题。

这是正式调查还是非正式询问？

为什么是现在？

之前公司有没有处理过类似情况？

总部知道吗？

我们是否希望合作？

如果合作，会不会形成先例？

如果不合作，会不会升级？

中国法怎么说？

美国法有没有冲突？

业务团队实际上希望达到什么结果？

谁应该去和政府沟通？

很多关键背景甚至根本不在任何数据库里。

它存在于人的脑子里。

存在于组织历史里。

存在于对某个监管机构的理解里。

存在于过去十次会议形成的默会知识里。

所以，高级法律工作很大一部分并不是信息检索。

而是：

**重新构建背景。**

而这是今天人工智能仍然很难完整获得的东西。

## 四、更深一层：法律经常不是在寻找答案，而是在决定“我们想要什么答案”

人工智能最擅长的一类问题是：

**给定条件，寻找答案。**

但是很多高级法律工作实际上是在问：

**条件本身应该如何定义？**

例如，公司问：

“这个产品能不能上线？”

一个比较初级的法律问题是：

“法律允不允许？”

但真正到了总法律顾问或者高级法务层面，问题往往变成：

“在法律风险、监管关系、商业机会、品牌影响和未来先例之间，我们愿意承担多少风险？”

这时候已经不是单纯的法律推理问题。

它包含：

判断，

价值取舍，

组织授权，

以及责任承担。

人工智能可以告诉你有哪些选择。

可以告诉你每个选择有哪些风险。

甚至可以模拟监管机构的反应。

但最终仍然需要有人说：

**我们愿意承担这个风险。**

这个动作，本质上是主体性。

所以我越来越觉得，理解法律人工智能不能只讨论智能水平。

还要讨论：

谁有权作出决定？

谁承担后果？

出错以后谁负责？

## 五、所以法律工作并不是不会被人工智能改变

恰恰相反。

我认为法律行业受到的冲击可能非常大。

只是它的路径不会完全复制软件工程。

很多法律工作已经非常适合被人工智能大幅压缩。

比如：

合同条款提取；

文件比对；

尽职调查；

案件时间线整理；

标准化法律研究；

监管规则梳理；

初稿起草；

政策比较；

证据审阅；

大量标准化合同工作。

这些工作的共同特征是什么？

**输入相对有限。**

**输出相对明确。**

**信息来源相对可识别。**

**验证相对容易。**

换句话说：

它们越来越像编程。

所以真正应该问的，并不是：

“法律会不会被人工智能替代？”

而是：

**这项法律任务有多像编程任务？**

一个法律任务越接近“输入明确—输出明确—可以验证—可以纠错”，人工智能介入的速度就会越快。

反过来，如果一个任务高度依赖关系、判断、模糊目标、组织权力和不可逆后果，那么人工智能自主执行的速度就会更慢。

## 六、可以建立一个简单的人工智能可替代性框架

我觉得可以用几个变量来判断一个任务有多容易被人工智能接管：

**可形式化程度**

任务能不能被清楚描述？

**可验证程度**

人工智能做完以后，能不能低成本判断它对不对？

**数字化程度**

完成任务需要的背景和数据是否已经数字化？

**可逆程度**

犯错以后是不是容易回滚？

以及：

**判断密度**

完成任务需要多少规范性、制度性、人际和商业判断？

编程往往具有这样的特征：

可形式化程度高；

可验证程度很高；

数字化程度很高；

可逆程度较高；

判断密度中等。

所以，人工智能的采用非常快。

法律往往恰恰相反。

而在这些变量里面，我认为最重要的可能是：

**可验证性。**

人工智能能不能取代一个人，很大程度上不取决于人工智能能不能第一次做对。

而取决于：

**人工智能做错以后，系统能不能告诉它，它错了。**

这是编程的巨大优势。

也是法律的巨大障碍。

## 七、真正危险的可能不是律师消失，而是法律职业的“生产层”被掏空

过去，法律职业大致是一个金字塔：

资深律师，

中级律师，

初级律师，

律师助理。

初级律师做大量研究、审阅、起草和尽职调查。

然后在这些工作中逐渐形成判断。

但人工智能出现以后，一个可能的新结构是：

资深律师负责判断、主体性和关系管理，

下面不再需要那么多人工生产，

而是一个很大的人工智能层。

研究、起草、审阅、比对、总结，大量由人工智能完成。

这意味着，人工智能不一定首先让资深律师消失。

反而可能首先压缩的是初级和中级律师的生产性工作。

但这又带来一个非常深的问题：

**如果判断过去是通过生产工作训练出来的，那么当生产工作被人工智能做掉以后，未来的高级判断从哪里来？**

一个律师过去可能要读一百个案例、改五十份合同、写三十份备忘录，才慢慢形成“感觉”。

如果未来年轻律师只负责批准人工智能的输出，他们会不会失去形成这种直觉的机会？

所以，法律人工智能的长期挑战可能不是：

“我们怎么用人工智能替代律师？”

而是：

**在机器承担大部分生产工作的时代，我们如何训练人的判断力？**

这个问题，我觉得目前远远没有被充分讨论。

## 八、工程可能只是比法律提前了一代

今天，编程已经越来越从：

“帮我写这段代码。”

走向：

“这是问题，你去解决。”

也就是从辅助工具进入智能体阶段。

而法律人工智能目前大多数时候还停留在：

帮我查；

帮我总结；

帮我写；

帮我审阅；

帮我分析。

真正的法律人工智能革命，可能要等到它跨过下一条线：

**“这是这个案件或者事项，你来运行整个流程。”**

比如，未来一个法律智能体可以：

收到监管机构的要求；

自动找到公司过去的类似案例；

查询相关法律；

调取内部政策；

识别相关利益方；

生成风险矩阵；

起草回复；

提出升级建议；

更新案件管理系统；

追踪截止日期；

最后交给律师批准。

到了这里，法律工作才真正开始进入智能体时代。

所以，也许软件工程和法律工作并不是走在两条不同的路上。

只是工程已经提前让我们看到了一部分未来。

## 九、人工智能真正改变的，可能不是“谁更聪明”，而是什么东西变得便宜

过去，好的生产能力很贵。

代码很贵。

法律研究很贵。

合同起草很贵。

分析也很贵。

人工智能正在快速让这些生产活动变便宜。

一旦生产变便宜，职业价值就会向其他地方移动。

在工程领域，价值越来越可能向系统架构、产品判断、系统设计和任务编排移动。

在法律领域，也可能发生类似变化。

当研究、起草、比较、总结都越来越便宜之后，真正稀缺的东西可能是：

**判断、主体性、信任和责任承担。**

所以，我越来越不愿意用“人工智能会不会取代律师”来理解这个问题。

更准确的问题可能是：

**当答案越来越便宜以后，什么东西还值得由人来做？**

而我的初步答案是：

人工智能可以越来越高效地生产答案。

但人仍然需要决定：

哪个答案值得相信；

哪个风险值得承担；

哪个目标值得追求；

以及最终，我们希望这个答案把组织带向哪里。

这也许才是人工智能时代专业工作的核心变化：

**人工智能让生产变得便宜，而人的价值将更多地转向判断与主体性。**