首页 / 财富中文网 / 正文

AI能否适用于法庭?律师称它已经在改变法律执业方式

财富中文网 2025-06-30 20:30:05

AI能否适用于法庭?律师称它已经在改变法律执业方式
图片来源:Illustration by Andrei Cojocaru; Original photos from Getty Images

2019年,一架从萨尔瓦多飞往纽约的哥伦比亚航空公司(Avianca)的航班上,一辆金属手推车沿过道滑下,撞到了罗伯托·马塔的膝盖。马塔聘请了两位人身伤害律师起诉该航空公司。该案件开创了法律先例,但这并非是由于马塔或其律师所期望的原因。相反,这两位律师在2023年声名狼藉,当时审理此案的法官因两人在提交给法庭的一份法律文书中引用了六个由人工智能软件虚构的案例,而对两人处以5,000美元的罚款。

该处罚开创了先例,并在法律界被广泛报道,告诫律师们依赖人工智能的风险。然而,这一警示故事并未得到重视。自该事件发生后的两年内,已有至少139起包含人工智能虚构引用的法律文书案例,仅今年5月就有超过30起。尽管其中许多案例来自自我代理的诉讼人,但也包括由知名全国性律师事务所提交的法律文书。而且,在一个案例中,法官排除了由人工智能巨头Anthropic公司的一位数据科学家提交的专家证词,理由是证词中包含了Anthropic自有软件生成的“幻觉”内容。

这些人工智能在法庭上出现失误的故事,成为律师事务所茶水间热议的话题。然而,这些案例也属特例。实际上,绝大多数使用人工智能的律师都在以负责任的方式使用它,包括在将人工智能生成的引文提交给法官之前,核实其是否真实存在。

许多人也对人工智能巨大的实用性感到欣喜。律师们表示,人工智能在研究和文件审阅方面表现出色,并且最近他们最近还发现它能够完成更具挑战性的任务,如起草法律文书。所有这些进展对法律职业的未来走向和法律业务本身都将产生重大影响。

日益增多的AI工具套件

与其他人一样,大多数律师在2022年11月才开始认识到人工智能的变革力量。当时,OpenAI向公众发布了ChatGPT的初始版本。不出所料,大约六个月后,第一批虚假引文开始出现在法庭上。

事实上,利用人工智能辅助法律研究的想法已存在十多年之久。该领域的先驱包括Ironclad和DISCO等公司,它们分别提供自动创建和管理合同以及加速证据开示流程的方法。

如今的不同之处在于,与其他行业一样,过去三年里,法律界以比以往更广泛、更快的速度获取和部署人工智能技术,尤其是在生成式人工智能变得更易获取且更强大之后。

今年2月,美国联邦律师协会(Federal Bar Association)发布了一项对2,800多名法律专业人士的调查:21%的受访者表示他们在2024年的执业中使用了人工智能,而在大型律师事务所的律师中,这一比例更是高达39%。法律转录服务公司Rev在4月发布的另一项律师调查发现,人工智能的整体采用率要高得多,48%的受访者表示他们使用人工智能进行研究。

不出所料,面向法律专业的人工智能产品数量持续增长。其中包括像Harvey这样的服务(它基于OpenAI、Anthropic和谷歌(Google)等公司的人工智能模型构建,专精于文书起草和研究);以及总部位于加拿大不列颠哥伦比亚省温哥华的Clio(提供律所管理工具)。Clio在2024年融资9亿美元,公司估值达到30亿美元;6月下旬,Harvey在一轮融资中筹集了3亿美元,公司估值达到50亿美元。

目前,尚无明确数据显示在工作中部署人工智能的律师使用这些专业工具的普遍程度,或者他们是否依赖由OpenAI和谷歌等公司开发的流行大语言模型,这些模型已输入了大量普通法法规和判例。

精通法律技术的斯坦福大学(Stanford)知识产权教授马克·莱姆利认为,大多数法学院学生和初级律师很可能在使用像ChatGPT这样的通用服务,因为这些服务是最容易获取的。

在大型律所中流行的一个人工智能工具是CoCounsel,它由法律初创公司Casetext开发。数据巨头汤森路透(Thomson Reuters)于2023年收购了Casetext,现在将CoCounsel纳入Westlaw服务(一项广泛用于所谓的“大型律所”、流行但昂贵的服务)。

全国性律所摩根路易斯(Morgan Lewis)的诉讼律师达里尔·兰迪表示,他喜欢CoCounsel的部分原因在于,它的设计考虑到了法律行业在客户保密和数据保护方面的特殊义务。在实践中,这意味着兰迪及其同事创建的人工智能提示不会成为训练数据集的一部分,从而无法为律所以外人士的研究提供信息。在从业者通常能接触到客户高度敏感信息的法律行业,这是一个重要问题。

职业规范的变迁

几十年来,大多数律师事务所都严格遵循“发现者、维护者和操作者”的模式。这个说法描述了一种金字塔结构:顶端的少数案源发现者负责招揽客户,并将大部分工作移交给二级的高级律师,后者则负责监督人数更多的初级律师群体。

人工智能智能体在这个金字塔结构中处于什么位置?一段时间以来显而易见的是,人工智能非常适合取代大量枯燥的“操作者”工作。事实上,在文件审阅、合同起草和基础研究等任务上,这种情况正在发生。这在许多方面是件好事。人工智能已经在消除法律工作中一些单调乏味的工作,其中许多传统上是为了追求计费工时而产生的。这是该行业一个臭名昭著的方面,常常让客户质疑发票上的工时有多少代表的是无谓的忙碌,而非不可或缺的任务。

从这个角度看,很难反驳一项能大幅减少繁琐任务用时、降低成本、并有望让律师腾出时间处理更深入和更具战略性事务的技术。然而,其中也存在弊端,尤其是在培训下一代律师方面最为明显。

莱姆利表示:“培养良好法律本能、进行逻辑论证的一种途径是通过反复实践训练,尝试各种方法,并反复地去做。”莱姆利担心,随着人工智能的更广泛采用,年轻律师可能无法获得一些传统上通过大量阅读案例法和法规而积累的分析技能。

莱姆利还观察到,人工智能正越来越多地承担起诸如撰写法律文书和构建论证等工作,而这些工作在许多律师看来是工作的乐趣所在,并且通常是由更资深的律师负责。

所有这些不仅引出了律师未来将如何安排时间的问题,也引出了他们将如何收取服务费用。多年来,客户们一直在努力(通常成效有限)引入替代性收费方式,以摆脱按小时支付初级律师钻研法律难题所耗费的时间。

莱姆利表示,未来这种情况可能会逆转:客户意识到AI可以在几分钟而非数小时内完成许多法律工作,因而可能要求按固定时间段收费,而律师事务所则会提议固定费用和其他方案。

长期以来,法律行业一直精于守护其商业模式,因此这类变革近期内不太可能发生。但目前可以明显看出:人工智能已经在为律师提供某种极为珍贵的东西:更多的时间。

根据汤森路透的调查,人工智能的广泛采用可能为普通律师每周节省出四小时,即每年200小时。这些时间可以用于拓展客户群体、培养额外技能,当然,也可以用于增加计费工时。

AI律师的法治规则

巴勃罗·阿雷东多是广受欢迎的人工智能工具CoCounsel的创始人,现任汤森路透副总裁。他经常为法官举办有关法律与人工智能交叉领域的研讨会。

在工作中,他常常思考一句拥有近百年历史的表述,这段表述出现在《联邦民事诉讼规则》(Federal Rules of Civil Procedure)开篇第一小节,即法官应鼓励“公正、迅速且低成本地”解决每一起诉讼。

阿雷东多略带讽刺地说道:“并非所有人都会用这三个词来描述美国的诉讼过程。”

如果法院能广泛且有效地采用人工智能工具,这种情况可能会发生改变。事实上,这种采用方式已经开始,例如马萨诸塞州的法院系统正在采用自动化工具,帮助面临驱逐的租户处理复杂的法律文件。

阿雷东多表示:“只要使用这项技术能带来更好的结果,就有道义责任去使用它。”不过,他也告诫道,无论是法院还是律师部署人工智能,都必须在最严格的监督下进行。

抱有这种想法的不止阿雷东多一人,因为法律行业与许多其他行业一样,正试图在人工智能带来的效率提升与安全问题之间取得平衡。2023年,首席大法官约翰·罗伯茨在思考该技术将如何影响法官工作时写道:“任何人工智能的使用,都需要保持审慎和谦逊的心态。” (*)

译者:刘进龙

审校:汪皓

2019年,一架从萨尔瓦多飞往纽约的哥伦比亚航空公司(Avianca)的航班上,一辆金属手推车沿过道滑下,撞到了罗伯托·马塔的膝盖。马塔聘请了两位人身伤害律师起诉该航空公司。该案件开创了法律先例,但这并非是由于马塔或其律师所期望的原因。相反,这两位律师在2023年声名狼藉,当时审理此案的法官因两人在提交给法庭的一份法律文书中引用了六个由人工智能软件虚构的案例,而对两人处以5,000美元的罚款。

该处罚开创了先例,并在法律界被广泛报道,告诫律师们依赖人工智能的风险。然而,这一警示故事并未得到重视。自该事件发生后的两年内,已有至少139起包含人工智能虚构引用的法律文书案例,仅今年5月就有超过30起。尽管其中许多案例来自自我代理的诉讼人,但也包括由知名全国性律师事务所提交的法律文书。而且,在一个案例中,法官排除了由人工智能巨头Anthropic公司的一位数据科学家提交的专家证词,理由是证词中包含了Anthropic自有软件生成的“幻觉”内容。

这些人工智能在法庭上出现失误的故事,成为律师事务所茶水间热议的话题。然而,这些案例也属特例。实际上,绝大多数使用人工智能的律师都在以负责任的方式使用它,包括在将人工智能生成的引文提交给法官之前,核实其是否真实存在。

许多人也对人工智能巨大的实用性感到欣喜。律师们表示,人工智能在研究和文件审阅方面表现出色,并且最近他们最近还发现它能够完成更具挑战性的任务,如起草法律文书。所有这些进展对法律职业的未来走向和法律业务本身都将产生重大影响。

日益增多的AI工具套件

与其他人一样,大多数律师在2022年11月才开始认识到人工智能的变革力量。当时,OpenAI向公众发布了ChatGPT的初始版本。不出所料,大约六个月后,第一批虚假引文开始出现在法庭上。

事实上,利用人工智能辅助法律研究的想法已存在十多年之久。该领域的先驱包括Ironclad和DISCO等公司,它们分别提供自动创建和管理合同以及加速证据开示流程的方法。

如今的不同之处在于,与其他行业一样,过去三年里,法律界以比以往更广泛、更快的速度获取和部署人工智能技术,尤其是在生成式人工智能变得更易获取且更强大之后。

今年2月,美国联邦律师协会(Federal Bar Association)发布了一项对2,800多名法律专业人士的调查:21%的受访者表示他们在2024年的执业中使用了人工智能,而在大型律师事务所的律师中,这一比例更是高达39%。法律转录服务公司Rev在4月发布的另一项律师调查发现,人工智能的整体采用率要高得多,48%的受访者表示他们使用人工智能进行研究。

不出所料,面向法律专业的人工智能产品数量持续增长。其中包括像Harvey这样的服务(它基于OpenAI、Anthropic和谷歌(Google)等公司的人工智能模型构建,专精于文书起草和研究);以及总部位于加拿大不列颠哥伦比亚省温哥华的Clio(提供律所管理工具)。Clio在2024年融资9亿美元,公司估值达到30亿美元;6月下旬,Harvey在一轮融资中筹集了3亿美元,公司估值达到50亿美元。

目前,尚无明确数据显示在工作中部署人工智能的律师使用这些专业工具的普遍程度,或者他们是否依赖由OpenAI和谷歌等公司开发的流行大语言模型,这些模型已输入了大量普通法法规和判例。

精通法律技术的斯坦福大学(Stanford)知识产权教授马克·莱姆利认为,大多数法学院学生和初级律师很可能在使用像ChatGPT这样的通用服务,因为这些服务是最容易获取的。

在大型律所中流行的一个人工智能工具是CoCounsel,它由法律初创公司Casetext开发。数据巨头汤森路透(Thomson Reuters)于2023年收购了Casetext,现在将CoCounsel纳入Westlaw服务(一项广泛用于所谓的“大型律所”、流行但昂贵的服务)。

全国性律所摩根路易斯(Morgan Lewis)的诉讼律师达里尔·兰迪表示,他喜欢CoCounsel的部分原因在于,它的设计考虑到了法律行业在客户保密和数据保护方面的特殊义务。在实践中,这意味着兰迪及其同事创建的人工智能提示不会成为训练数据集的一部分,从而无法为律所以外人士的研究提供信息。在从业者通常能接触到客户高度敏感信息的法律行业,这是一个重要问题。

职业规范的变迁

几十年来,大多数律师事务所都严格遵循“发现者、维护者和操作者”的模式。这个说法描述了一种金字塔结构:顶端的少数案源发现者负责招揽客户,并将大部分工作移交给二级的高级律师,后者则负责监督人数更多的初级律师群体。

人工智能智能体在这个金字塔结构中处于什么位置?一段时间以来显而易见的是,人工智能非常适合取代大量枯燥的“操作者”工作。事实上,在文件审阅、合同起草和基础研究等任务上,这种情况正在发生。这在许多方面是件好事。人工智能已经在消除法律工作中一些单调乏味的工作,其中许多传统上是为了追求计费工时而产生的。这是该行业一个臭名昭著的方面,常常让客户质疑发票上的工时有多少代表的是无谓的忙碌,而非不可或缺的任务。

从这个角度看,很难反驳一项能大幅减少繁琐任务用时、降低成本、并有望让律师腾出时间处理更深入和更具战略性事务的技术。然而,其中也存在弊端,尤其是在培训下一代律师方面最为明显。

莱姆利表示:“培养良好法律本能、进行逻辑论证的一种途径是通过反复实践训练,尝试各种方法,并反复地去做。”莱姆利担心,随着人工智能的更广泛采用,年轻律师可能无法获得一些传统上通过大量阅读案例法和法规而积累的分析技能。

莱姆利还观察到,人工智能正越来越多地承担起诸如撰写法律文书和构建论证等工作,而这些工作在许多律师看来是工作的乐趣所在,并且通常是由更资深的律师负责。

所有这些不仅引出了律师未来将如何安排时间的问题,也引出了他们将如何收取服务费用。多年来,客户们一直在努力(通常成效有限)引入替代性收费方式,以摆脱按小时支付初级律师钻研法律难题所耗费的时间。

莱姆利表示,未来这种情况可能会逆转:客户意识到AI可以在几分钟而非数小时内完成许多法律工作,因而可能要求按固定时间段收费,而律师事务所则会提议固定费用和其他方案。

长期以来,法律行业一直精于守护其商业模式,因此这类变革近期内不太可能发生。但目前可以明显看出:人工智能已经在为律师提供某种极为珍贵的东西:更多的时间。

根据汤森路透的调查,人工智能的广泛采用可能为普通律师每周节省出四小时,即每年200小时。这些时间可以用于拓展客户群体、培养额外技能,当然,也可以用于增加计费工时。

AI律师的法治规则

巴勃罗·阿雷东多是广受欢迎的人工智能工具CoCounsel的创始人,现任汤森路透副总裁。他经常为法官举办有关法律与人工智能交叉领域的研讨会。

在工作中,他常常思考一句拥有近百年历史的表述,这段表述出现在《联邦民事诉讼规则》(Federal Rules of Civil Procedure)开篇第一小节,即法官应鼓励“公正、迅速且低成本地”解决每一起诉讼。

阿雷东多略带讽刺地说道:“并非所有人都会用这三个词来描述美国的诉讼过程。”

如果法院能广泛且有效地采用人工智能工具,这种情况可能会发生改变。事实上,这种采用方式已经开始,例如马萨诸塞州的法院系统正在采用自动化工具,帮助面临驱逐的租户处理复杂的法律文件。

阿雷东多表示:“只要使用这项技术能带来更好的结果,就有道义责任去使用它。”不过,他也告诫道,无论是法院还是律师部署人工智能,都必须在最严格的监督下进行。

抱有这种想法的不止阿雷东多一人,因为法律行业与许多其他行业一样,正试图在人工智能带来的效率提升与安全问题之间取得平衡。2023年,首席大法官约翰·罗伯茨在思考该技术将如何影响法官工作时写道:“任何人工智能的使用,都需要保持审慎和谦逊的心态。” (*)

译者:刘进龙

审校:汪皓

In 2019 a metal cart rolled down the aisle of an Avianca flight from El Salvador to New York and struck Roberto Mata on the knee. Mata retained two personal injury lawyers to sue the airline, resulting in a case that made legal precedent—but not for a reason Mata or his attorneys would have hoped for. Instead, the two lawyers achieved notoriety in 2023 when the judge in the case fined the pair $5,000 for citing six fictitious cases, invented by artificial intelligence software, in a brief to the court.

The punishment was the first of its kind and was widely reported in legal circles as a cautionary tale for lawyers about the risk of relying on AI. That cautionary tale has not been heeded. In the two years since the incident, there have been no fewer than 139 instances of legal briefs containing AI-invented citations—including over 30 this May alone. While many of these came from self-represented litigants, they also include briefs filed by prominent national law firms. And, in one instance, a judge struck expert testimony submitted by a data scientist from AI giant Anthropic—on the grounds it contained a “hallucination” created by Anthropic’s own software.

These stories of AI gone wrong in the courtroom make for a lively source of chatter around law firm watercoolers. Yet these cases are also outliers. In reality, the vast majority of lawyers using AI are doing so in responsible fashion—including by verifying that AI-produced citations actually check out before putting them before a judge.

Many are also delighted by how useful it is. Lawyers say AI excels at research and document review and that, more recently, they have found it to be capable of more challenging tasks like drafting legal briefs. And all of this progress has big implications for both the future trajectory of legal careers and for the business of law.

A growing suite of AI tools

Most lawyers came to appreciate the transformative power of AI at the same time as everyone else: November 2022. That was when OpenAI released the initial version of ChatGPT to the general public, and, unsurprisingly, it was about six months later that the first wave of fake citations began showing up in courtrooms.

In fact, the idea of using AI to inform legal research has been around for over a decade. Pioneers in the field include the likes of Ironclad and DISCO, which respectively offer automated ways to create and manage contracts and to expedite the discovery process.

The difference today is that, as in other industries, the past three years have seen the legal profession acquire and deploy AI tech at a broader and faster pace than ever before—especially as generative AI has become more easily available and more powerful.

In February, the Federal Bar Association published a survey of over 2,800 legal professionals: 21% said they used AI in their practice in 2024, and the number was significantly higher—39%—for attorneys at large law firms. A separate attorney survey published in April by the legal transcription service Rev found much higher overall adoption, with 48% of respondents saying they used AI for research.

Unsurprisingly, the number of AI offerings aimed at the legal profession is growing. There are services like Harvey, which is built on top of AI models from companies including OpenAI, Anthropic, and Google and specializes in drafting and research; and Vancouver, British Columbia–based Clio, which offers practice management tools. Clio raised $900 million in 2024, valuing the company at $3 billion; in late June, Harvey raised $300 million in a funding round that values the company at $5 billion.

For now, there is no clear data on how widely the lawyers deploying AI in their jobs are using these specialized tools, or if they are instead relying on popular LLMs created by the likes of OpenAI and Google, which have all ingested reams of common-law statutes and decisions.

According to Mark Lemley, a Stanford intellectual property professor well versed in legal tech, most law students and junior lawyers are likely using general services like ChatGPT, since those are the most accessible.

In large firms, a popular AI tool is CoCounsel, which was developed by the legal startup Casetext. The data giant Thomson Reuters purchased Casetext in 2023 and now offers CoCounsel as part of Westlaw, a popular but expensive service widely used by so-called Big Law firms.

Daryl Landy, a litigator with the national firm Morgan Lewis, says he likes CoCounsel in part because it is designed to take account of the legal profession’s special obligations regarding client confidentiality and data protection. In practice, this means that AI prompts created by Landy and his colleagues do not become part of a training set that can inform the research of people outside his firm—an important issue in a profession in which practitioners often have access to highly sensitive information about their clients.

Changing professional norms

For decades, most law firms have fit neatly into the “finders, minders, and grinders” model. The phrase describes a pyramid structure where a handful of rainmakers at the top bring in clients and hand over the bulk of the work to a second tier of senior lawyers, who in turn oversee a larger pool of associates.

Where does an AI agent fit into the pyramid? It’s been obvious for some time that AI is a good fit to displace large swaths of mundane “grinder” work—and indeed that is occurring with tasks like document review, contract drafting, and basic research. This is a good thing in many respects. AI is already eliminating some of the drudgery of legal work, much of which has traditionally come about in pursuit of billable hours—a notorious aspect of the profession that often leaves clients wondering how many hours on an invoice represent busywork rather than indispensable tasks.

In this light, it’s hard to argue with a technology that vastly reduces time spent on tedious tasks, reduces costs, and promises to free up lawyers for deeper and more strategic matters. Still, there are drawbacks, most notably when it comes to training the next generation of lawyers.

“One way you get good legal instincts for logical arguments is by getting the reps in—trying things, and doing them over and over,” says Lemley. He worries that wider adoption of AI means young lawyers may not be gaining some of the analytic skills that have traditionally been acquired through extensive reading of case law and statutes.

Lemley also observes that AI is increasingly taking on tasks like writing briefs and crafting arguments, which many attorneys view as the fun part of the job, and which typically fall to more senior attorneys.

All of this raises the question of not just how lawyers will spend their time in the future, but also how they will charge for their services. For years, clients have fought—typically with limited success—to introduce alternative fee arrangements in order to escape paying by the hour for junior associates to go down legal rabbit holes.

In the future, Lemley says, the situation may be reversed as clients—realizing that AI can perform many legal tasks in minutes not hours—ask to be billed for fixed blocks of time, while law firms propose flat fees and other arrangements.

The legal profession has long been adept at guarding its business model, so such changes are likely not on the immediate horizon. But for now, it’s becoming clear that AI is already providing lawyers with something precious: additional time.

According to the Thomson Reuters survey, the widespread adoption of AI will likely free up four hours a week for the typical lawyer, or 200 hours a year—hours that could be used to expand their client base, develop additional skills or, of course, stack billable hours.

A rule of law for AI lawyers

Pablo Arredondo, who created the popular CoCounsel AI tool and is now a vice president at Thomson Reuters, frequently runs seminars for judges about the intersection of law and artificial intelligence.

In the course of his work, he often reflects on a nearly century-old phrase that appears in the very first section of the Federal Rules of Civil Procedure—namely, that judges should encourage the “just, speedy, and inexpensive” resolution of every action.

“Those three words are not how anyone would describe U.S. litigation,” Arredondo says dryly.

This could change if courts adopt AI tools in a widespread and effective fashion. And, indeed, such adoption is already beginning, including in Massachusetts where the court system is adopting automated tools to help tenants facing eviction navigate complex legal forms.

“To the extent this technology can be used for better outcomes, there is a moral imperative to use it,” says Arredondo. Still, he cautions that any deployment of AI by both courts and lawyers must occur with the strictest of oversight.

Arredondo is not alone in that sentiment as the legal profession, like so many other industries, tries to balance the efficiency gains offered by AI with safety. In 2023, reflecting on how the technology is poised to disrupt the work of judges, none other than Chief Justice John Roberts wrote, “Any use of AI requires caution and humility.”

*