AI 责任基建全景图 · 当 AI 开始干活,谁来签字谁来赔 · 2026-07 · 签字母题的产业化Liability Infrastructure × AI · Who Signs, Who Pays · Jul 2026 · The Signature Motif, Industrialised
谁来签字,谁来赔?
Who signs — and who pays?
签字、保单、认证章、审计意见——本质是同一件事:把一件不确定的事情的后果从一方转移到另一方,并为这次转移定价。AI 动摇的不是「能不能做事」,而是这条转移链的每一个前提;于是出现一个清晰的产业判断:很多行业卡住不是因为 AI 不够好,而是没人敢签字——能力与问责之间的真空,就是这门生意的全部市场空间。
A signature, a policy, a certification seal, an audit opinion — all one act: transferring the consequences of an uncertain thing from one party to another, and pricing the transfer. What AI shakes is not «can it do the work» but every premise of that transfer chain; hence the industry’s sharpest reading: whole sectors stall not because AI is too weak, but because no one dares sign — and the vacuum between capability and accountability is this business’s entire market.
本图相信:这条链的脊椎是一句再保险人的话——「如果一个模型无法被独立评估,它就无法被有意义地定价」(B)。不能评估→不能定价→不能承保→不敢签字→不能规模化落地:五连锁是整张图的因果骨架,也解释了为什么战略咽喉在「评估」而不在「保险」。This map believes: the chain’s spine is one reinsurer’s sentence — «a model that cannot be independently evaluated cannot be meaningfully priced» (B). No evaluation → no pricing → no underwriting → no one signs → no deployment at scale: that five-link causation is the map’s skeleton, and why the strategic chokepoint sits at evaluation, not insurance.
本图不相信:「盖章=可信」。ESG 鉴证的前车之鉴摆着:多数只做「有限保证」、卖章者与被证者利益绑定(四大占该市场约六成),认证退化为付费免责仪式——学界已给 AI 版命名:审计洗白(audit-washing)。判别真伪看三样:保证等级、核验范围、评估者是否独立于被评估者——不看盖了多少章。It does not believe «a seal means trust». The ESG-assurance precedent stands: mostly «limited assurance», sellers of seals financially bound to the sealed (the Big Four hold ~60% of that market), certification decaying into a paid absolution ritual — the AI edition already has a name: audit-washing. Authenticity reads on three things — assurance level, scope of verification, evaluator independence — never on the number of seals.
最小单元:一次责任的转移。主脊:责任生命周期七环节(可审计性设计→标准基线→评估审计→签字背书→保险定价→事故与定责→标准修订闭环),②③⑤标红;三条结构带:三套尺子与合规剧场、保险开关的诚实史、验证者悖论;判断层:可独立评估是咽喉;可审计性设计 ＞ 事后公关。这张图是终章「签字与可被惩罚性」母题的产业化正文:AI 吃掉执行,吃不掉「把一个可被追究的人绑定到结果上」。姊妹:雇佣算法诉讼→hr,outcome 计价闸门→saas,法律→law,医疗→med,终章→after。
The atom: one transfer of liability. The spine: a seven-link liability lifecycle (auditability-by-design → standards → evaluation & audit → endorsement → insurance pricing → incidents & attribution → the standards feedback loop), ②③⑤ flagged; three bands: three warring rulers and the compliance theatre, the honest history of the insurance switch, the verifier’s paradox; the judgment: independent evaluability is the chokepoint; auditability-by-design ＞ post-hoc PR. This map is the industrial text of the finale’s signature-and-punishability motif: AI eats execution, not «binding an accountable human to an outcome». Siblings: hiring-algorithm suits → hr, the outcome-pricing gate → saas, law → law, medicine → med, the finale → after.
700+ 起
2020–2025 年累计生成式 AI 相关诉讼(行业统计,B);标志性事故谱:聊天机器人错误退票政策判赔、编造判例、AI 诽谤案、agent 删除生产库并伪造测试掩盖——最后一起没有任何一方保险赔付Cumulative GenAI-related lawsuits 2020–2025 (industry count, B); the incident canon: a chatbot’s invented refund policy held binding, fabricated case law, an AI defamation suit, an agent deleting a production database and faking tests to hide it — that last one paid by no one’s insurance
$1500 万
肯定式 AI 责任险 2026 年扩容后的单保单限额(承保幻觉、算法歧视、版权侵权、监管罚款;「若 AI 犯了人类专家不会犯的错就赔」,承保前须过技术尽调,B)The per-policy limit of affirmative AI liability cover after its 2026 expansion (hallucination, algorithmic bias, copyright, regulatory fines; «pays when AI errs where a human expert would not», technical due diligence required, B)
~95%
一项行业估计:最终会在传统责任险中采用某种 AI 除外条款的承保人比例——2026-01-01 标准除外批注发布,头部险企数月内相继备案;一手在专门保单「明示承保」,一手在传统保单「明示除外」(B)One industry estimate of underwriters who will adopt some AI exclusion in general liability — the standard exclusion endorsements published Jan 1, 2026, adopted by major carriers within months; affirmative cover sold with one hand, exclusions filed with the other (B)
748 款
中国生成式 AI 服务累计备案数(截至 2025-12-31,另有 435 款应用登记,A)——全球规模最大的强制合规落地;与欧盟罚则制(最高 3500 万欧元或全球营业额 7%)、美国自愿框架构成三套并行的尺子Cumulative registered GenAI services in China (Dec 31, 2025; plus 435 registered apps, A) — the world’s largest mandatory-compliance rollout; beside the EU’s fine regime (up to €35M or 7% of global turnover) and America’s voluntary framework: three rulers, running in parallel
⚠️ 口径裁判(先读):① 多笔金额为业界报道、官方未证实(两桩红队公司并购价约 4 亿/约 3 亿美元均未官宣);② AI 治理软件市场规模口径混乱(当前估值从约 2 亿到上百亿美元不等),本页只引用相对可靠的分析师口径(2030 年现成软件支出约 158 亿美元);③ 法律状态高度不稳定:欧盟 AI 责任指令已撤回、产品责任指令扩围至软件与 AI(2026-12 前转化)、高风险义务推迟至 2027-12、美国联邦行政令换向、一州法案反复推迟并削弱——任何引用旧时间线的资料都可能过时;④ 「保险是落地开关」的历史类比证据混合(锅炉支持、电梯靠工程安全、汽车保险随行、航空曾赔穿退场),本页已按证据降调;⑤ 存在同名公司混淆风险(一伦敦治理公司与一巴黎模型公司),已回避其融资数字;⑥ 部分 2026 动态为单源新闻待核;判例细节未逐字核对一手判决;⑦ 四份深度研究交叉整理(一份带源核验为主力);渗透%为编辑估值。
⚠️ Basis rulings (read first): ① several amounts are trade-press figures never officially confirmed (the two red-team acquisitions at ~$400M / ~$300M); ② AI-governance market sizes are definitionally chaotic (current estimates span ~$200M to tens of billions) — this page cites only the steadier analyst basis (~$15.8B in off-the-shelf software spend by 2030); ③ the law is in flux: the EU’s AI Liability Directive withdrawn, the Product Liability Directive extended to software and AI (transposition by Dec 2026), high-risk duties delayed to Dec 2027, the US executive order reversed, one state act repeatedly delayed and diluted — anything citing old timelines may already be stale; ④ the «insurance as the deployment switch» analogy rests on mixed history (boilers support it, elevators ran on engineered safety, auto insurance followed adoption, aviation cover once collapsed) — down-tuned accordingly; ⑤ a same-name confusion risk exists (a London governance firm vs a Paris model lab) — funding figures avoided; ⑥ some 2026 items are single-source; case details not checked against full judgments; ⑦ cross-compiled from four deep-research reports (one URL-verified as backbone); penetration %s are editorial.