When AI Does the Work, Who Is Liable? / AI干活，谁担责？

AI now drafts pleadings and checks case law; in a single year it moved from experiment to daily routine. California, the first U.S. state to legislate on this, has set two hard rules: every court document submitted by a lawyer must bear the personal signature of at least one lawyer, and any document touched by generative AI must be disclosed to the court. Clients and law firms have not settled how the time saved should be divided. But the person who answers for the tool’s errors is already fixed: the lawyer who signs.

AI 写诉状、查案例，一年里从尝鲜变成日常。美国第一个就此立法的州（加州）定下了两条硬规矩：律师提交的每份法院文件，至少要有一名律师以个人名义签名；凡是经过生成式 AI 处理的文件，还要向法庭说明。省下来的时间怎么分，客户和律所还没谈拢。但替工具的错误担责的人，已经确定是签字的那个律师。

Adoption First, Rules Later

使用先到，规则后到

Law360 Pulse’s annual AI survey gives the speed of adoption: the share of lawyers at law firms using AI at least three times a week rose from 27% to 47% in a year, and in corporate legal departments it was higher, at 56%. Use nearly doubled in a year, but the rules governing its risks have not kept pace.

Law360 Pulse 的年度 AI 调查给出了使用端的速度：律所里每周至少用三次 AI 的律师，一年内从 27% 升到 47%；公司法律部更高，为 56%。使用率一年翻了近一倍，管理使用风险的规则却没有跟上这个速度。

Because the rules lag, the risk falls on identifiable individuals. Even where a firm has hired a specialist to promote AI and take over training, risk assessment remains the general counsel’s own job — and there is only more of it.

规则没跟上，风险就落到具体的人头上。即便律所请了专人推广 AI、接手培训，风险评估依然是总法律顾问自己的事，而且只会更多。

David Roth has been general counsel at Bradley since 2022. He says he has never seen the practice of law change so sharply in his career. Several matters sit with him: teaching lawyers how to use AI in compliance with the rules, setting policies and processes, bringing clients’ expectations about AI into line, and going through the terms of contracts with AI vendors clause by clause. He also flags an easily missed detail: as cyberattack methods improve, the means of defense have to improve with them.

Bradley 律所的总法律顾问 David Roth 从 2022 年起担任这个职务。他说，律师执业方式变化之剧烈，他从业以来没有见过。他手上压着几件事：教律师怎么合规地使用，制定政策和流程，把客户对 AI 的期待拉齐，再逐条看清与 AI 供应商签的合同条款。他还提到一个容易被忽略的细节：网络攻击的手法在进步，防守的手段也只能跟着升级。

AI brings law firms two things: output, and responsibility that a specific person must own. Who approves its use, who reviews its work, and who answers when something goes wrong — these three questions used to be settled by a firm’s internal policies; now the statute requires the signing lawyer to answer them one by one.

AI 给律所带来两件事：一是产出，二是必须由具体的人认领的责任。谁批准用、谁负责复核、出了差错谁担责，这三个问题原本由律所内部制度规定，现在法条要求由签字的律师逐项负责。

A Signature Becomes a Legal Duty

签字成了法定义务

On September 30, the governor of California signed Senate Bill 574 (SB 574), which takes effect on January 1, 2027.

9 月 30 日，加州州长签署参议院第 574 号法案（SB 574），2027 年 1 月 1 日生效。

Other states govern lawyers’ use of AI through court rules and bar association guidance; California is the first to do it by statute. A statute is passed by the state legislature and signed by the governor, so its force comes from law rather than professional self-regulation alone.

其他州管律师用 AI，靠的是法院规则和律师协会的指引；加州是第一家用成文法的州。成文法由州议会通过、州长签署，约束力来自法律，而不只是行业自律。

The statute’s requirements fall into several layers:

法条的要求可以拆成几层：

• A lawyer may not delegate the practice of law wholesale to AI;
• Confidential information may not be entered into public AI tools;
• For filings submitted to a court, every citation — the cases and statutes cited — must be checked personally by the lawyer responsible for the filing, including citations generated by AI;
• Documents processed by generative AI must be disclosed to the court;
• Every court document submitted by a lawyer acting for a party must bear the personal signature of at least one lawyer;
• Arbitrators are subject to the same constraints: they may not delegate adjudication to AI, and must tell the parties when they rely on AI-generated information.

• 律师不得把执业工作整体委托给 AI；
• 保密信息不得输入公开的 AI 工具；
• 递交法院的文件，每条引注（引用到的判例和法条）都要由负责提交的律师亲自核对，AI 生成的引注同样不能免；
• 经过生成式 AI 处理的文件，要向法庭披露；
• 代理律师提交的每份文件，至少有一名律师以个人名义签名；
• 仲裁员受同一套约束：不得把裁断交给 AI，依赖 AI 生成的信息要向各方说明。

Violations can be sanctioned by the court. The statute did not come out of nowhere. In March of this year, a California appellate court fined a lawyer $5,000 in a custody dispute because the materials filed with the court contained citations invented by AI.

违规的，法院可以制裁。这条法也不是凭空来的。今年 3 月，加州一家上诉法院就对一起抚养权纠纷中的律师罚了 5000 美元，原因是提交给法院的材料里有 AI 编出来的引注。

Bloomberg Law’s October 2 report puts the difficulty of implementation most plainly: the hardest question is what counts as using AI.

执行端的难处，Bloomberg Law 在 10 月 2 日的报道里点得最清楚：最难的一条是「什么算用了 AI」。

Scott Meyers, chair of the board at Akerman, says AI is already embedded in the research and drafting tools lawyers use every day; which situations count as using AI is hard to define, but every firm must define it before the law takes effect.

Akerman 律所董事会主席 Scott Meyers 说，AI 已经内嵌在律师每天使用的检索和文书工具里，哪些情形算用了 AI 不好界定，但每家律所都必须在生效前界定清楚。

Peter Salib, an assistant professor of law at the University of Houston, notes that the disclosure duty applies equally to every document filed with a court, with no exemption for “minor use.” Firms therefore need two things: standard disclosure language, and a register recording which documents were processed by AI.

休斯顿大学法学助理教授 Peter Salib 提醒，披露义务对递交法院的全部文件一视同仁，没有「只是轻微使用」的豁免。律所因此需要两样东西：一套标准的披露文本，一本台账，记下哪些文件经过 AI 处理。

Ariel House, a partner at Baker Botts, points to another cost: checking every citation personally shifts work that support staff used to do under a lawyer’s supervision onto the signing lawyer. Solo practitioners and small firms carry the heaviest burden.

Baker Botts 合伙人 Ariel House 看到的是另一项代价：逐条亲自核对，等于把原本由辅助人员在律师监督下完成的核对工作，转由签字律师本人承担。个人执业者和小律所的负担最重。

When Something Goes Wrong, Who Is Responsible

出错之后，责任怎么分

These matters fall to Bradley’s general counsel David Roth, and the first pressure comes from outside. People representing themselves in court have begun using AI to draft filings. Twenty years ago, he says, most pro se submissions were handwritten complaints that judges had to strain to make out the claims; now the same group generates motions in batches with AI and files them in volume, and the firms responsible have to add staff.

这些事落到 Bradley 律所总法律顾问 David Roth 头上，先要应付的是外部。自己打官司的普通人开始借 AI 起草文书。他说，二十年前的自诉材料多半是一份手写诉状，法官得费劲辨认诉求；如今这批人用 AI 批量生成动议，成批往法院递，负责的律所只能加人手。

Inside the firm, a different set of questions arises. Who is responsible for checking AI-generated content? Roth cannot review every pleading in every one of the firm’s cases, and no one expects him to. What he can do is set the requirements for lawyers together with his colleagues at the firm and pass them down to everyone.

律所内部是另一面。核对 AI 生成的内容该算谁的责任？ Roth 审不完这家律所每一起案件里的每一份诉状，也没有人指望他这样。他能做的，是和律所里的同事一起把对律师的要求定下来，再传到每个人。

Clients’ attitudes have reversed over two years: they used to come in warning firms not to use AI; now they push firms to find ways to use it well and deliver work more accurately and faster. Whether to use the tool is no longer the question; how responsibility is assigned when it goes wrong is the new task firms have taken on.

客户的态度在两年里转了个方向：以前客户上门是叮嘱别用 AI，现在催着律所想办法把它用好，把服务做得更准、更快。工具要不要用已经不是问题，出了错怎么追责，才是各律所新添的功课。

Responsibility is not divided only inside the firm. At this year’s annual meeting of the Association of Corporate Counsel, Seyfarth partner David Baffa said AI is no longer just a question of which tool to pick; it has changed the division of work between clients and outside counsel. The questions listed at the meeting all pointed to responsibility: whose model is used, where the data comes from, how human review and escalation work, and how the money saved and the risk assumed are shared. The Axiom survey cited in the same report shows that one in five corporate legal departments already require outside firms to say when AI was used.

责任也不只在律所内部划分。今年美国企业法律顾问协会的年会上，Seyfarth 合伙人 David Baffa 说，AI 已经不只是「挑哪个工具」的事，它改变了客户与外部律师之间的分工。会上列出的问题逐条指向责任：用谁的模型，数据从哪里来，人工复核和升级怎么走，省下来的钱和担下的风险怎么分。同一篇报道引用的 Axiom 调查显示，已有五分之一的企业法务部门要求外部律所说明什么时候用了 AI。

As Regulation Lands, the Role Comes Closer

监管落地一步，岗位近一步

In another October 2 report, law.com described a role that is heating up: the “AI lawyer.” Companies are eager to hire lawyers who can advise on AI, under a jumble of job titles and with entry requirements that vary widely.

law.com 在 10 月 2 日的另一篇报道里，提到一个正在变热的岗位：「AI 律师」。企业急着招能就 AI 出具意见的律师，岗位名目五花八门，门槛深浅不一。

Rebekah Stafford, a vice president at Epiq, expects this demand to fade as the skills spread. Transactional and corporate lawyers will all need to understand how AI fits into the business; it is basic competence, not a separate profession.

Epiq 副总裁 Rebekah Stafford 认为，这股需求会随技能普及而消退。交易律师、公司律师都得懂 AI 怎么嵌进业务，它是基本功，不是一个独立的行当。

Jared Coseglia, founder of TruLegal, draws on the history of the industry: e-discovery and privacy both became established lines of work because regulation drove them to maturity, with the rules arriving first and demand settling into roles afterward. AI is still far from that point. He also warns that lawyers who cannot use AI will gradually slide toward contract work.

TruLegal 创始人 Jared Coseglia 拿行业史做对照：电子取证和隐私这两个行当，都是被监管催熟的，监管先立起来，需求才固化成岗位。AI 离那一步还远。他还提醒，不会用 AI 的律师会慢慢滑向合同工的位置。

California supplies exactly the other half of the facts: a statute now exists, and the signature requirement places responsibility on a specific person. Each step that regulation takes moves the role closer to a formal headcount and away from a line in an internal memorandum.

加州提供的恰好是另一半事实：成文法已经出现，签名义务把责任落到了具体的人身上。监管每落地一步，这个岗位就从一份备忘录向一个编制挪近一步。

China has no statute dedicated to lawyers’ use of AI; local bar associations began issuing guidance this year.

中国没有专门管律师用 AI 的法律，地方律协今年开始出指引。

The Ningxia Lawyers Association’s March Guidelines on the Use of Generative AI by Lawyers put it most directly: the lawyer is ultimately responsible, and fault liability arising from AI use cannot be transferred; firms must set up internal systems, control tool selection, and provide training.

宁夏律师协会 3 月发布的《宁夏律师使用生成式人工智能指引》说得最直：律师是最终责任人，AI 使用中的过错责任不可转移；律所须建内部制度、管住工具选型、做培训。

In July the Shanghai Lawyers Association adopted its Operational Guidelines for AI Applications in Corporate Legal Counsel Services, making clear that lawyers bear ultimate professional judgment over legal opinions, documents, and analytical conclusions produced with AI assistance.

上海市律师协会 7 月通过《律师企业法律顾问服务 AI 应用操作指引》，明确律师对 AI 辅助形成的法律意见、文件和分析结论，承担最终的专业判断责任。

The operational guidance issued by the Kunming Lawyers Association in late September goes into greater detail: acts with legal effect, such as sending legal documents to third parties or filing litigation materials, must be confirmed item by item by the lawyer personally; where an AI agent acts without such confirmation, the consequences still fall on the lawyer and the firm.

昆明市律师协会 9 月底发布的操作指引走得更细：对外发送法律文书、提交诉讼材料这类会产生法律效果的行为，必须由律师本人逐项确认；未经确认、由智能体自主完成的，后果仍由律师和其所在律所承担。

None of the three sets of guidance sets up a signature procedure, but they point the same way: using AI does not transfer professional responsibility.

三份指引都没有设「签名」这道程序，但指向一致：用了 AI，执业责任不转移。

A discussion of legal ethics published by China Lawyers Network offers a more concrete practice: lawyers should record the time of each AI use, the prompts entered, and the raw output received, and keep the working files with the final version. If only the final version is kept, the AI-generated draft cannot be traced when something goes wrong. As for who approves and who reviews inside a firm, that still rests with each firm’s own systems and its partners.

中国律师网刊出的一篇执业伦理讨论提了更具体的做法：律师应记下每次使用 AI 的时间、输入的提示词和拿到的原始输出，过程性文件一并留档。只存定稿，出了问题就查不到 AI 生成的那一版。至于律所内部谁批准、谁复核，眼下还是落在各家自己的制度和合伙人身上。

The California statute, the practices of large U.S. firms, and three sets of local guidance in China all point to the same question: whether AI governance works is no longer measured by how many documents a firm has issued, but by whether a specific person can be found when something goes wrong.

加州的法条、美国大型律所的做法、中国三地的指引，指向同一个问题：AI 治理做得好不好，不再按发了多少份文件来算，而按「出了事能不能找到人」来算。

For most large firms, the statute does not change how they use AI. Daren Orzechowski, global co-head of the technology practice at A&O Shearman, puts it this way: the California law largely turns into statutory text what lawyers were already supposed to do; what is genuinely new is the disclosure duty and the way verification is pressed onto the name that signs.

对多数大型律所来说，这个法条不改变它们用 AI 的方式。A&O Shearman 全球科技业务联席负责人 Daren Orzechowski 的说法是，加州这条法基本是把律师本来就该做的事定为条文，真正新增的是披露义务，以及把核验压到签字的那个名字上。

Sources: law.com (reports of October 2 and October 20, 2026); Law360 Pulse (2026 AI survey and October 2026 coverage of the Association of Corporate Counsel annual meeting); Bloomberg Law (report of October 2, 2026); China Lawyers Network (legal-ethics discussion); Ningxia Lawyers Association, Guidelines on the Use of Generative AI by Lawyers; Shanghai Lawyers Association, Operational Guidelines for AI Applications in Corporate Legal Counsel Services (2026) (Trial); Kunming Lawyers Association, Operational Guidelines on the Use of AI Assistance in Legal Practice (Trial).

来源：law.com（2026年10月2日、10月20日各一篇报道）；Law360 Pulse（2026年AI调查及10月企业法律顾问协会年会报道）；Bloomberg Law（2026年10月2日报道）；中国律师网（执业伦理讨论）；宁夏律师协会《宁夏律师使用生成式人工智能指引》；上海市律师协会《律师企业法律顾问服务AI应用操作指引（2026）（试行）》；昆明市律师协会《关于律师使用人工智能（AI）辅助执业操作指引（试行）》。