纯文本行业为何最难被吞噬 · AI 与法律产业 · 2026-07Why a pure-text industry is hardest to swallow · AI & law · Jul 2026

按理说,法律是 AI 最好啃的行业:输入是文本,输出是文本。现实是它啃得最慢——「怎么做」被重写了,而执照、签字、地方性知识三道墙后面那件事,AI 一个字也改不动:「谁负责」 On paper, law should be AI's easiest meal: text in, text out. In practice it chews slowest — the «how» has been rewritten, while behind the three walls of licence, signature and local knowledge sits the one thing AI cannot edit: «who answers for it»

最小单元=一次「法律三段论」:事实(小前提)+ 规范(大前提)→ 结论。三段里 AI 真正拿下的只有中间那段——「规范」(GPT-4 过了统一律考);两头的「事实认定」与「为结论签字」,被执照与责任制度焊死在人类手里。这就是中心悖论:瓶颈从来不是算力,是制度。The atomic unit is one «legal syllogism»: fact (minor premise) + rule (major premise) → conclusion. Of the three parts, AI has truly taken only the middle — the «rule» (GPT-4 passed the bar); the two ends, «finding the facts» and «signing for the conclusion», are welded to human hands by licence and liability. That is the central paradox: the bottleneck was never compute — it is the institution.

诚实底色=赢得判决 ≠ 拿到钱。中国 2025 执行到位率约 50%,美国民事金钱判决估计 65–80% 收不回;而低收入者 92% 的民事法律问题根本得不到足够帮助。两个数字夹出同一个结论:对大多数纠纷,最优解在法庭之外——事前的合同与证据,胜过事后的雄辩。The honest baseline: winning a judgment ≠ getting paid. China's 2025 collection rate is ~50%; an estimated 65–80% of US civil money judgments are never collected; and 92% of low-income civil legal problems receive no adequate help at all. Two numbers, one conclusion: for most disputes, the best solution lies outside the courtroom — contracts and evidence upfront beat eloquence after the fact.

主脊是纠纷生命周期六节点(风险预防→咨询评估→取证→调解仲裁→诉讼判决→执行)+ 刑事平行轨,每节点标 AI 渗透深度:合同最深、执行最浅——AI 的力量从「事前」到「事后」逐级衰减。另含三道制度护城河、中美「次序倒置」、幻觉零容忍(Mata 案)、产品指南(2C 当事人可用 / 2B 执业侧)。本页是批判性行业解剖:提供法律信息,不构成法律意见。签字担责三兄弟:注会→ledger、医生→med;法学教育→edu。 The spine: the six-node dispute lifecycle (prevention → intake → evidence → mediation/arbitration → litigation → enforcement) plus a criminal rail, each node tagged by AI depth: contracts deepest, enforcement shallowest — AI's force decays from «before» to «after». Plus the three institutional moats, the China-US «inversion», zero-tolerance for hallucination (Mata), and a product guide (2C for parties / 2B for practice). A critical dissection: legal information, not legal advice. The signature-liability trio: the CPA's → ledger, the doctor's → med; legal education → edu.

传统节点Traditional

AI 已重写 · 生产方式AI rewrote · production

制度护城河 · 未被重写Institutional moat · intact

幻觉零容忍 · 执行难Zero-tolerance · enforcement

判断层 · 法庭之外Judgment · beyond court

44.32%

中国 2025 一季度执行完毕率(到位率 56.74%)。判决书上的正义要打对折才是账户里的钱——AI 提得了效率,变不出可执行财产。⚠️完毕率≠到位率,季度≠全年,不可直比China's Q1 2025 enforcement clearance (collection 56.74%). Justice on paper halves before it reaches the account — AI speeds the process, it cannot conjure collectable assets. ⚠️Clearance ≠ collection; quarter ≠ year

110亿美元

Harvey 估值(2026-03,GIC/Sequoia 领投,ARR 约 1.9 亿)。资本给法律 AI 的定价,买的全是「怎么做」——合同、检索、起草;「谁负责」一寸未让Harvey's valuation (Mar 2026, GIC/Sequoia, ~$190M ARR). What capital is pricing is entirely the «how» — contracts, research, drafting; the «who answers» has not yielded an inch

$5,000 × N

Mata v. Avianca:律师提交 ChatGPT 虚构判例,法官按第 11 条对两名律师及律所各罚 5000 美元,斥其「胡言乱语」。一份假引用=一次制裁——法律对幻觉零容忍;仅 2025 就发现 294 起 pro se 幻觉案Mata v. Avianca: ChatGPT-fabricated precedents drew $5,000 Rule 11 sanctions each on two lawyers and their firm, the output called «gibberish». One fake citation = one sanction — law is zero-tolerance; 294 pro-se hallucination cases in 2025 alone

92%

美国低收入者 92% 的民事法律问题得不到足够帮助(LSC《正义鸿沟》)。AI 把咨询压到近零成本,是弥合鸿沟的机会,也是诉讼洪水的引信——取决于系统接不接得住US low-income people get inadequate help for 92% of civil legal problems (LSC Justice Gap). AI drives advice to near-zero cost — a chance to close the gap, and a fuse for a litigation flood; it depends on whether the system can hold

口径警告:本页提供关于行业趋势的法律信息(legal information),不构成针对具体情形的法律意见(legal advice);AI 输出与本页结论都必须经持证专业人士核验后才可用于实际决策。执行率三口径并存(完毕率 44.32%/到位率 56.74%,2025Q1;全年到位率 50.59%;2023 完毕率 41.96%)——互不可直比。GPT-4 律考「第 90 百分位」经 MIT 复核降为「首考第 62、通过者第 48、论述题第 15」(版图 C 并列展示)。Harvey 用户数官方「逾 10 万律师」(D)与 Sacra「14.2 万」(C)并存;厂商自述一律打 D。「AI 幻觉第一案」杭州(AI 公司责任)≠ 北京通州(律师伪造案例),勿混。卡片右上角 A/B/C/D=证据强度。 Basis warning: this page offers legal information about industry trends, not legal advice for any specific situation; AI output and these conclusions must be verified by a licensed professional before real decisions. Enforcement rates carry three bases (clearance 44.32% / collection 56.74%, Q1 2025; full-year collection 50.59%; 2023 clearance 41.96%) — none directly comparable. GPT-4's bar «90th percentile» falls, on MIT's re-check, to «62nd first-takers, 48th among passers, 15th on essays» (shown side by side in Map C). Harvey's users: official «100k+ lawyers» (D) vs Sacra's «142k» (C); vendor claims are graded D throughout. The «first AI-hallucination case»: Hangzhou (AI-firm liability) ≠ Beijing Tongzhou (lawyer fabrication). Card badges A/B/C/D = evidence strength.

◆ 中心悖论 · 纯文本行业为何最难被吞噬The central paradox · why pure text resists

技术这边门槛早已越过;制度那边,门都没开过On the technical side the bar was cleared long ago; on the institutional side the door never opened

法律的输入输出都是文本,推理链条高度结构化——按吞噬公式,它本该第一个沦陷。可它偏偏落地最慢。原因不在模型:GPT-4 早就过了统一律考;在于三道制度护城河把「事实认定」与「结论负责」整段焊死。所以要看清这张图,先分清两个对象:AI 冲击的是法律工作的「生产方式」,不是律师、法官、检察官这些制度角色——前者是技术问题,后者是宪制问题。Law's inputs and outputs are text and its reasoning is rigidly structured — by the consumption formula it should have fallen first. Yet it lands slowest. The reason is not the model: GPT-4 passed the bar long ago; it is that three institutional moats weld shut «finding the facts» and «answering for the conclusion». To read this map, separate two objects: AI strikes the «production» of legal work, not the institutional roles of lawyer, judge and prosecutor — the former is an engineering question, the latter a constitutional one.

▚ 能力越界 vs 制度锁死Capability spills over vs the institution locks down

▸ 技术:门槛早已越过Tech: the bar is cleared

GPT-4 通过统一律考GPT-4 passed the bar exam

原始口径「接近第 90 百分位、约 75% 原始分」,远超 GPT-3.5 的第 10 百分位;合同 / 检索 / 初稿 / 摘要 / 类案的中后台生产已被整体重写。⚠️但别急着欢呼——MIT 复核把「90%」打回原形:论述题只有第 15 百分位(版图 C 有全账)。The raw claim: «~90th percentile, ~75% raw score», far past GPT-3.5's 10th; the back office of contracts / research / drafts / summaries / precedent has been rewritten wholesale. ⚠️Hold the applause — MIT's re-check returned the «90%» to earth: 15th percentile on essays (full ledger in Map C).

◆ 制度:三道墙纹丝不动Institution: three walls hold

改不了「谁负责」Can't rewrite «who answers»

过律考 ≠ 有执业资格 ≠ 能出庭。只要签字的无限责任与伪证 / 失职的惩戒,最终必须落在一个持有执照的物理人类身上,AI 的天花板就被钉死在「超级助理」——再聪明,也是助理。Passing the bar ≠ a licence ≠ standing in court. As long as the signature's unlimited liability and the sanctions for perjury / malpractice must land on a licensed physical human, AI's ceiling is nailed at «super-assistant» — however brilliant, still an assistant.

① 执照垄断① Licence monopoly

出庭资格 · UPL 无证执业禁令Standing · UPL bans

② 签字责任② Signature liability

谁签字谁负无限责任 · 不可转嫁算法Whoever signs bears unlimited liability

③ 地方性知识③ Local knowledge

隐性规则 · 法官偏好 · 关系网Tacit rules · judge preferences · networks

▸ 第一层张力 · 请得起 vs 请不起的正义Tension 1 · justice you can afford vs can't

「法律面前人人平等」,账单面前未必Equal before the law — not before the bill

现实里法律更像一种特权:LSC《正义鸿沟》——低收入美国人 92% 的民事法律问题得不到足够帮助,74% 的家庭一年至少遇到一个;WJP 估全球约 50 亿人司法需求未被满足;同一个中国,北京律师年均创收约 84 万,内蒙古约 16 万。AI 第一次把咨询的边际成本压到近零(幂律「吾律」月费 88 元、通义法睿 / DeepSeek 免费)——普通人第一次拿到了「看起来专业」的法律帮助。「看起来」三个字,是机会,也是陷阱。In practice law behaves like a privilege: LSC's Justice Gap — 92% of low-income Americans' civil problems get inadequate help, 74% of households face one a year; the WJP counts ~5 billion people with unmet justice needs; within one China, a Beijing lawyer bills ~¥840k a year against ~¥160k in Inner Mongolia. AI has, for the first time, driven the marginal cost of advice to near zero (PowerLaw's «Wulü» at ¥88/month, Tongyi Farui / DeepSeek free) — ordinary people's first access to «professional-looking» legal help. The word «looking» is both the opportunity and the trap.

▸ 第二层张力 · 完美想象 vs 执行的冷数学Tension 2 · the fantasy vs enforcement's cold math

赢得判决 ≠ 拿到钱Winning ≠ getting paid

执行率是司法系统最诚实的一栏账:中国 2025 一季度执行完毕率 44.32%、到位率 56.74%(全年到位率约 50.59%);美国民事金钱判决估计 65–80% 永远收不回来。判决书写得再漂亮,债务人没有财产,AI 也变不出来。而当维权成本崩塌,被激活的不只是被压抑的正义——还有诉讼洪水:仅 2025 就发现 294 起 pro se 当事人提交 AI 幻觉材料。一句话:AI 已进场,系统还没准备好。The enforcement rate is the system's most honest ledger line: China's Q1 2025 clearance was 44.32%, collection 56.74% (full-year ~50.59%); an estimated 65–80% of US civil money judgments are never collected. However elegant the judgment, if the debtor has no assets, AI cannot conjure them. And when the cost of claiming rights collapses, what awakens is not only suppressed justice but a litigation flood: 294 pro-se filings with AI-hallucinated material found in 2025 alone. In one line: AI has arrived; the system has not.

◆ 判断层杠杆排序 · 结论卡The leverage stack · the conclusion

事前的合同与证据留存Upfront contracts & evidence ＞事后的诉讼技巧after-the-fact litigation

和解的期望值Expected value of settlement ＞情绪化的坚持emotional persistence

管辖权与程序设计Jurisdiction & procedure ＞实体法庭雄辩courtroom eloquence

最优解在法庭之外,而且可以算出来:和解期望值 = 可执行金额 × 收回概率 − 诉讼成本 − 时间成本——这个数常常大于「赢了官司收不到钱」。制度也在往同一方向走:中国 2024 办结执行保全 344.4 万件(+45.23%,「以保促执」),2023 诉前调解成功 1204 万件(占民事行政案件 40.3%)。程序是骨架,实体是血肉——骨架搭错,血肉无处附着。The best solution is outside the courtroom — and it can be computed: settlement EV = collectable amount × recovery probability − litigation cost − time cost, a figure that regularly beats «won the case, collected nothing». The institutions lean the same way: China closed 3.44M preservation cases in 2024 (+45.23%, «preserve to enforce») and succeeded in 12.04M pre-litigation mediations in 2023 (40.3% of civil/administrative filings). Procedure is the skeleton, substance the flesh — build the skeleton wrong and the flesh has nowhere to hang.

从这张图带走的五条规律Five patterns to take away