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

法律理论上最该被 AI 吞噬，现实中却落地最慢——AI 冲击的是「怎么做」的生产方式，撞上的是执照、签字、地方性知识三道制度墙：它改不了「谁负责」 Law should be the easiest to swallow, yet it lands slowest — AI reworks the 'how' of production but hits three institutional walls of license, signature and local knowledge: it can't rewrite 'who's accountable'

最小单元＝一次「法律三段论」：事实（小前提）＋规范（大前提）→ 结论。AI 已能处理「规范」（GPT-4 过统一律考、原始分约 75%），但「事实认定」与「结论负责」两端被执照与责任制度锁死。这就是纯文本行业反而落地最慢的中心悖论——瓶颈不在算力，在制度。The atomic unit is one 'legal syllogism': fact (minor premise) + rule (major premise) → conclusion. AI already handles the 'rule' (GPT-4 passed the bar, ~75% raw), but 'fact-finding' and 'owning the conclusion' are locked by license and liability. That is the central paradox — the bottleneck isn't compute, it's 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；低收入者 92% 的民事法律问题得不到足够帮助。所以对绝大多数纠纷，最优解在法庭之外：事前的合同与证据留存＞事后的诉讼技巧，和解的期望值＞情绪化的坚持。The honest baseline: winning ≠ getting paid. China's 2025 collection rate is ~50%, US civil money judgments an estimated 65–80% uncollectable; 92% of low-income civil legal problems get no adequate help. So for most disputes the best solution is outside the courtroom: upfront contracts and evidence ＞ after-the-fact litigation skill; the expected value of settlement ＞ emotional persistence.

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。另含中心悖论、三道制度护城河、中美「次序倒置」、幻觉零容忍（Mata 案）、龙头 Harvey。这是一张批判性行业解剖，提供法律信息而非法律意见。相邻议题见姊妹图——签字担责三兄弟：注会签字→ledger、医生签字→med；法学教育 / 法考→edu。 The spine is the six-node dispute lifecycle (prevention → intake → evidence → mediation/arbitration → litigation → enforcement) plus a criminal rail, each tagged by AI penetration: contracts deepest, enforcement shallowest. Plus the central paradox, three institutional moats, the China-US 'inversion,' zero-tolerance for hallucination (Mata), and leader Harvey. A critical dissection — legal information, not legal advice. Adjacent topics on siblings — the signature-liability trio: the CPA's → ledger, the doctor's → med; legal education / the bar → 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%）——赢得判决 ≠ 拿到钱，AI 提效却不能凭空制造可执行财产。⚠️口径：完毕率≠到位率，年份 / 季度不同不可直比China Q1-2025 case-clearance rate in enforcement (collection rate 56.74%) — winning ≠ getting paid; AI speeds things up but can't conjure collectable assets. ⚠️Basis: clearance ≠ collection, periods differ

110亿美元

Harvey 估值（2026-03，GIC / Sequoia 领投）、ARR 约 1.9 亿美元——资本狂热与能力越界并存，但覆盖的是合同 / 检索 / 起草的「生产方式」Harvey's valuation (Mar-2026, GIC / Sequoia), ARR ~$190M — capital frenzy and capability alike, but it covers the 'production' of contracts / research / drafting

5000美元 × N

Mata v. Avianca 案：法官依《联邦民诉规则》第 11 条对两名律师及律所各罚 5000 美元，斥 AI 生成分析为「胡言乱语」——法律对幻觉零容忍；2025 一年发现 294 起 pro se 用 AI 幻觉Mata v. Avianca: Rule 11 sanctions of $5,000 each on two lawyers and the firm, calling AI output 'gibberish' — law is zero-tolerance on hallucination; 294 pro-se AI-hallucination cases found in 2025 alone

92%

美国低收入者 92% 的民事法律问题得不到任何或足够帮助（LSC《正义鸿沟》）——AI 把咨询成本压到近零，可能弥合鸿沟，也可能引发诉讼洪水US low-income people get no or inadequate help for 92% of their civil legal problems (LSC Justice Gap) — AI drives advice cost near zero, which may close the gap or trigger a litigation flood

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 输出与本页结论均须经持证专业人士核验后方可用于实际决策。执行率三口径打架（完毕率 44.32% / 到位率 56.74%，2025Q1；全年到位率 50.59%；2023 完毕率 41.96%）——完毕率≠到位率、季度≠全年，不可直比。GPT-4 律考原始「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」（并列展示见版图 C）。Harvey 用户数官方「逾 10 万律师」(D) vs Sacra「14.2 万」(C)；厂商自述数字统一打 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。每张卡片右上角 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 conclusions here must be verified by a licensed professional before real decisions. Enforcement rates conflict across three bases (clearance 44.32% / collection 56.74%, Q1-2025; full-year collection 50.59%; 2023 clearance 41.96%) — clearance ≠ collection, quarter ≠ year, not directly comparable. GPT-4's bar '90th percentile' was revised by MIT 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 '142k' (C); vendor figures graded D. The 'first AI-hallucination case' — Hangzhou (AI-firm liability) ≠ Beijing Tongzhou (lawyer fabrication) — don't conflate. Each card's top-right 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

技术早已越过门槛，制度门槛纹丝不动Technology cleared the bar; the institution didn't move

法律输入输出皆文本、推理链高度结构化，理论上最该被 AI 吞噬。但答案不在算法算力——GPT-4 已过统一律考；而在三道制度性护城河把「事实认定」与「结论负责」牢牢锁死。AI 首先冲击的是法律工作的「生产方式」，而不是律师 / 法官 / 检察官这些制度角色本身。Law's inputs and outputs are text and its reasoning is highly structured, so in theory it should be the easiest to swallow. But the answer isn't compute — GPT-4 already passed the bar; it's that three institutional moats lock down 'fact-finding' and 'owning the conclusion.' AI first hits the 'production' of legal work, not the institutional roles of lawyer / judge / prosecutor themselves.

▚ 能力越界 vs 制度锁死Capability spills over vs the institution locks it 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、高于人类平均 68%」，远超 GPT-3.5 的第 10 百分位。合同 / 检索 / 初稿 / 摘要 / 类案 / 在线办案的中后台生产方式已被重写。⚠️但 MIT 复核戳破「90%」神话（论述题仅第 15 百分位，见版图 C）。The raw claim: '~90th percentile, ~75% raw, above the human 68%,' far past GPT-3.5's 10th. The back-office production of contracts / research / drafts / summaries / precedent is already rewritten. ⚠️But MIT's re-check punctured the '90%' myth (essays only 15th percentile, see 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's accountable'

过律考 ≠ 获执业资格 ≠ 可出庭。只要签字责任与伪证 / 失职惩戒最终必须归属拥有执照的物理人类，AI 就永远被禁锢在「超级助理」的从属地位。Passing the bar ≠ a license ≠ standing in court. As long as the signature's liability and perjury / malpractice sanctions must land on a licensed physical human, AI stays confined to a subordinate 'super-assistant' role.

① 执照垄断① License 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

「人人平等」vs 高昂价格'Equal before the law' vs the price tag

法律现实中「更是一种特权」：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% 家庭一年至少遇到一个；WJP 估全球约 50 亿人司法需求未满足；中国北京律师年均创收约 84 万 vs 内蒙古约 16 万。AI 把咨询成本压到近零（幂律「吾律」月费 88 元、通义法睿 / DeepSeek 免费），让普通人第一次获得「看起来专业」的法律帮助。In practice law is 'more of a privilege': LSC's Justice Gap — 92% of low-income Americans' civil problems get inadequate help, 74% face one a year; WJP estimates ~5 billion people worldwide have unmet justice needs; a Beijing lawyer earns ~¥840k/yr vs ~¥160k in Inner Mongolia. AI drives advice cost near zero (PowerLaw's 'Wulü' at ¥88/mo, Tongyi Farui / DeepSeek free), giving ordinary people 'professional-looking' help for the first time.

▸ 第二层张力 · 完美想象 vs 执行难的冷酷数学Tension 2 · the fantasy vs the cold math of enforcement

赢得判决 ≠ 拿到钱Winning ≠ getting paid

胜诉判决书上的正义，不等于账户里的真金白银——执行率才是司法系统最诚实的底色。中国 2025 一季度执行完毕率仅 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约 65–80% 无法实际收回。维权成本崩塌可能触发诉讼洪水：2025 一年发现 294 起 pro se 用 AI 幻觉信息，系统「已进场，尚未准备好」。The justice on the winning judgment isn't cash in the account — the enforcement rate is the system's most honest baseline. China's Q1-2025 clearance was just 44.32%, collection 56.74% (full-year ~50.59%); US civil money judgments are an estimated 65–80% uncollectable. Collapsing enforcement cost may trigger a litigation flood: 294 pro-se AI-hallucination cases in 2025 alone — the system has 'arrived but isn't ready.'

◆ 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：和解期望值 = 可执行金额 × 收回概率 − 诉讼成本 − 时间成本，常高于「赢了官司却收不到钱」。佐证：中国 2024 办结执行保全案件 344.4 万件（同比 +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——骨架错了，血肉无从附着。The best solution is outside the courtroom: the expected value of settlement = collectable amount × recovery probability − litigation cost − time cost, often above 'won the case but got nothing.' Evidence: China closed 3.44M asset-preservation cases in 2024 (+45.23%, 'preserve to enforce'), 12.04M pre-litigation mediations succeeded in 2023 (40.3% of civil/administrative cases). Procedure is the skeleton, substance the flesh — a broken skeleton has nothing to attach to.

从这张图看到的五条规律Five patterns this map makes visible