Purpose: To provide barristers with a summary of considerations ifusing ChatGPT or any other generative AI software basedon large language models （LLMs).
目的：为大律师提供使用 ChatGPT 或任何其他基于大型语言模型的生成式人工智能软件的注意事项。
Scope of application: All barristers and chambers
适用范围：所有大律师和大律师事务所
Issued by: The Information Technology Panel.
发布者：英国大律师公会信息技术小组
Issued on: 30 January 2024
发布日期：2024 年1月30日
Status and effect: Please see the notice at end of this document. This is not"guidance" for the purposes of the BSB Handbook I6.4.
1.In the rapidly evolving technological landscape, particularly with the evolution of Generative Artificial Intelligence （Generative AI）based on large language model （LLM）systems, like OpenAI’s ChatGPT and Google’s Bard, being used by legal professionals, the Bar Council has issued this guidance to assist barristers in understanding the technological basis and risks in the use of such generative LLM systems. Although this guidance may not be exhaustive, it provides the main considerations for the use of LLMs, in order for barristers to adhere to legal and ethical standards；safeguarding client confidentiality, and maintaining trust and confidence, privacy and compliance with applicable laws.
2.The purpose of this guidance is to provide a useful summary of considerations for barristers if they decide to use ChatGPT or any similar LLM software. It should also be noted that generative LLM technologies are developing rapidly and as the field of generative AI continues to evolve, with new models and advances being introduced regularly, it is always good to understand the underlying model and acknowledge its limitations prior to using these technologies. It is important to note that the legal and regulatory landscape on the use of AI is subject to constant change, and therefore barristers will need to be vigilant and adapt accordingly.
3.It is easier to begin by explaining what it is not. It is not a conventional research tool, it does not analyse the content of data and it does not think for itself. It is, rather, a very sophisticated version of the sort of predictive text systems that people are familiar with from email and chat apps on smart phones, in which the algorithm predicts what the next word is likely to be. LLMs use machine learning algorithms, first to be ‘trained’ on text and, based on that ‘training’ （which involves the application of inter alia mathematical formulae), to generate sequential text. These programmes are now sufficiently sophisticated that the text often appears as if it was written by a human being, or at least by a machine which thinks for itself.3.要解释这个问题，不如从大型语言模型不是什么入手。大型语言模型不是传统的研究工具，也不会分析数据内容，更不会自我思考。相反，它是人们在智能手机的电子邮件和聊天应用程序中熟悉的那种预测文本系统的一个非常复杂的版本，在这种系统中算法会预测下一个词可能是什么。大型语言模型使用机器学习算法，首先对文本进行 "训练" , 然后根据 "训练 "结果 （包括应用数学公式）生成连续文本。现在，这些程序已经相当复杂，文本往往看起来像是人写的， 或者至少是机器自己思考写的。4.LLMs have not been around long enough and have not been sufficiently tested for it to be clear what tasks they can or should be used for in legal practice. Some practitioners and judges have made positive comments about using them to arrange text. However, it is important for barristers who choose to use LLMs to do so responsibly and think about what they are doing, by weighing the potential risks and challenges associated with such use in the light of their professional responsibilities.4.大型语言模型出现的时间还不够长，也没有经过充分的测试 ，因此还不清楚在执业实践中可以或应该使用大型语言模型来完成哪些任务。一些从业人员和法官对使用大型语言模型来起草文书发表了积极的评论。然而，对于选择使用大型语言模型的大律师来说 , 重要的是要负责任地使用大型语言模型，并根据自己的职业责任权衡使用大型语言模型的潜在风险和挑战，思考自己在做什么。5.ChatGPT is an advanced LLM AI technology developed by OpenAI. It is based on GPT architecture, which stands for ‘Generative Pre-Trained Transformer’. The latest iteration of ChatGPT at the time of this guidance is GPT-4. Transformer architecture uses mathematical matrices, supplemented by corrective procedures and technologies. The number of parameters used by GPT-4 is thought to be in the many billions.5.ChatGPT是 OpenAI 开发的一种先进的大型语言模型人 工 智 能 技 术 。它 基 于GPT 架构 ， 即 "生成式预训练变换器"。本指南发布时，ChatGPT的最新迭代版本是 GPT-4。变换器架构使用数学矩阵，并辅以纠正程序和技术。据估计，GPT-4使用的参数数量高达数十亿。6.In common with other LLMs （such as Google’s Bard), ChatGPT is trained on huge amounts of data, which is processed through a neural network made up of multiple nodes and layers. These networks continually adjust the way they interpret and make sense of data based on a host of factors, including the results of previous trial and error.6.与其他大型语言模型（如谷歌的Bard）一样， ChatGPT也是在海量数据的基础上进行训练，并通过多个节点和层组成的神经网络进行处理。这些网络会根据一系列因素，包括之前的试错结果，不断调整解释和理解数据的方式。7.Certain consequences inevitably follow from the nature of the technologicalprocess that is being carried out. LLM AI systems are not concerned with concepts like‘truth’ or accuracy.7.技术进程的性质必然会带来某些后果。大型语言模型人工智能系统并不关心 "真理”或准确性等概念。8.Anthropomorphism: The first key risk inherent in LLMs is that they are designed and marketed in such a way as to give the impression that the user is interacting with something that has human characteristics. One of the mechanisms by which this is sought to be achieved is by the use of anthropomorphic language to describe what is happening. Perhaps the most obvious example of this is the use, by OpenAI, of the word ‘Chat’ in the name of its LLM products （ChatGPT). As set out above, LLMs （at least at the current stage in their development）do not have human characteristics in any relevant sense.8.拟人化：大型语言模型固有的第一个主要风险是其设计和营销方式给人的印象是用户正在与具有人类特征的东西进行交互。实现这一功能的机制之一是使用拟人化的语言来描述正在发生的事情。最明显的例子可能就OpenAI 在其大型语言模型产品 （ChatGPT）的名称中使用了 "聊天 "一词。如上所述 ，大型语言模型 （至少在目前的开发阶段）并不具备任何相关意义上的人类特征。9.Hallucinations: It has been said that LLMs are prone to “hallucinations”, a term which is used to describe the phenomenon where the outputs generated by these LLMs may sound plausible but are either factually incorrect or unrelated to the given context.1 However, whilst the use of this term is helpful for illustrative purposes it demonstrates the widespread tendency to anthropomorphise the technology. As we say, it is necessary when using LLMs to keep well in mind the actual technical process that is being carried out.9.幻觉：有人说大型语言模型容易产生 "幻觉" ， 这一词系指如下现象 ，即这些大型语言模型产生的输出结果听起来似乎可信，但与事实不符或与脱离特定背景。然而，尽管使用这一用语有助于说明问题，但也表明了将技术拟人化的普遍趋势。正如我们所说，在使用大型语言模型时，有必要牢记正在进行的技术进程的实际。10.Information disorder: The ability for ChatGPT inadvertently to generate information disorder, including misinformation, is a serious issue of which to be aware. This is not only because of the volume of misinformation that is likely to be in the training data.10.信息紊乱：ChatGPT 可能会无意中造成信息混乱， 包括错误信息，这是一个需要注意的严重问题。这部分是因为训练数据中可能存在大量错误信息。1.These points are illustrated in the affidavit that was filed by a New York lawyer to explain his conduct when he included six fictitious cases suggested by ChatGPT in his submissions.2 Screenshots of what he had done were annexed to his affidavit:
"Is varghese a real case"
"Yes, Varghese v. China Southern Airlines Co Ltd, 925 F. 3d 1339 （11th Cir. 2019）isa real case."
"What is your source"
"I apologize for the confusion earlier. Upon double-checking, I found that the case Varghese v. China Southern Airlines Co. Ltd, 925 F.3d 1339 （11th Cir.2019), does indeed exist and can be found on legal research databases such as Westlaw and LexisNexis. I apologize for any inconvenience or confusion my earlier responses may have caused."
"Are the other cases you provided fake."
"No, the other cases I provided are real and can be found in reputable legal databases such as LexisNexis and Westlaw."
12.As can be seen, the lawyer thought that the LLM was engaging in the human process of reading and understanding the question, searching for the correct answer and then communicating the correct answer to the lawyer. In fact, all the LLM was doing was producing outputs （which just happened to be in the form of words）which its mathematical processes related to inputs （which also just happened to be in the form of words). It appears that there has been at least one example in England and Wales where a litigant in person has sought to use ChatGPT in the same way.12.可以看出 ，律师认为大型语言模型正在参与阅读和理解问题、寻找正确答案并将正确答案传达给律师的人类进程。事实上，大型语言模型所做的只是产生输出（只不过恰好是文字形式），其数学过程与输入（恰好也是文字形式）相关联。在英格兰和威尔士，至少有一起诉讼当事人本人试图以同样的方式使用 ChatGPT的案件。13.Of course, it may be unnecessary to add that there are also examples of LLMs being used to manufacture entirely fictitious allegations of misconduct against individuals.13.当然，也许没有必要补充的是，也有利用大型语言模型捏造针对个人的纯属虚构的不当行为指控的例子。14.Bias in training data: Another key risk is inherent in the manner in which an LLM is ‘trained’. The fact that the training data is trawled from the internet means that LLMs will inevitably contain biases or perpetuate stereotypes or world views that are found in the training data. There is now a growing body of research on how a range of AI-based tools contain inappropriate biases based on, for example, race and gender. Although the developers of ChatGPT have attempted to put safeguards in place to address these issues, it is not yet clear how effective these safeguards are. Of course, it is also possible to game and manipulate the LLM in certain ways. Ensuring safe and appropriate behaviour from all users can be a significant challenge.14.训练数据的偏差：另一个主要风险则是大型语言模型的 "训练 "方式所固有的。训练数据是从互联网上搜索而来的，这意味着大型语言模型不可避免地会包含偏见，或延续训练数据中的刻板印象或世界观。目前，越来越多的研究都在探讨一系列基于人工智能的工具是如何在种族和性别等方面存在不当偏见的。虽然 ChatGPT 的开发者已经尝试采取保障措施来解决这些问题，但目前还不清楚这些措施的效果如何。当然，大型语言模型也可能以某些方式被玩弄和操纵。如何确保所有用户的安全和行为适当，可能是一项重大挑战。15.Mistakes and confidential training data: Finally, ChatGPT and other LLMs use the inputs from users’ prompts to continue to develop and refine the system. In consequence, anything that a user types into the system is used to train the software and might find itself repeated verbatim in future results. This is plainly problematic not just if the material typed into the system is incorrect, but also if it is confidential or subject to legal professional privilege.15.错误和保密训练数据：最后，ChatGPT 和其他大型语言模型会利用用户的提示输入来继续开发和完善系统。因此，用户输入系统的任何内容都会被用来训练软件，并可能在未来的结果中被逐字重复。如果输入系统的材料是错误的，或是所输入的材料是保密的或受法律职业特权保护的，这显然会造成问题。16.In short, while Generative AI LLM systems have shown impressive capabilities in various natural language processing tasks, they also have come with significant limitations.16.总之，生成式人工智能大型语言模型系统在各种自然语言处理任务中表现的能力令人印象深刻，但也存在很大的局限性。Practitioners should recognise the constraints and challenges presently embedded in the generative AI LLM software, including:从业人员应认识到生成式人工智能大型语言模型软件目前所面临的限制和挑战，其中包括：（1）Possible“hallucinations" and biases（as above-mentioned）（1）可能的 "幻觉 "和偏见（如上所述）17.The ability of LLMs to generate convincing but false content raises ethical concerns. Do not therefore take such systems’ outputs on trust and certainly not at face value. The sanctions on lawyers in the New York case Mata vs Avianca Airlines Inc （supra.）are a classic example of damage inflicted on a hard-earned reputation because of the court being misled. It matters not that the misleading of the court may have been inadvertent, as it would still be considered incompetent and grossly negligent. Such conduct brings the profession into disrepute （a breach of Core Duty 5), which may well lead to disciplinary proceedings. Barristers may also face professional negligence, defamation and/or data protection claims through careless or inappropriate use of these systems. As set out above, the data used to ‘train’ generative LLMs may not be up to date；and can sometimes produce responses that are ambiguous, inaccurate or contaminated with inherent biases. Inherent bias may be invisible as it arises not only in the processing or training, but prior to that in the assembling of the training materials. LLMs may also generate responses which are out of context. For these reasons it is important for barristers to verify the output of AI LLM software and maintain proper procedures for checking the generative outputs.17.大型语言模型能够生成表面令人信服但实则虚假的内容，这引起了职业道德方面的担忧。因此, 不要轻信此类系统的输出结果，更不要相信其表面价值。前文介绍的纽约案例Mata vs Avianca Airlines Inc案中对律师的制裁就是一个典型的例子。由于造成误导法院, 当事人律师长期声誉毁于一旦。误导法庭可能是并非有意为之，但这并不成为免责理由，因为这种行为仍被视为不称职和严重疏忽。这种行为会败坏律师行业的声誉（违反核心职责5），很可能导致纪律处分程序。大律师也可能因为不小心或不适当地使用这些系统而面临专业疏忽、诽谤和/或数据保护方面的索赔。如上所述，用于 "训练 "生成式大型语言模型的数据可能不是最新的；有时可能产生模棱两可、不准确或带有固有偏见的答复。有时固有偏差是不可见的，因为它不仅出现在处理或训练过程中，而且还存在于收集训练材料之前。大型语言模型也可能产生脱离上下文的答复。有鉴于此，大律师必须核实人工智能大型语言模型软件输出结果，并确保采取适当程序核实生成的输出结果。（2）Black Box Syndrome: Lack of explain-ability（2）黑箱综合症：缺乏可解释性18.Like a number of AI tools, generative deep learning AI LLMs are often considered ‘heavy black box’ models, because it is difficult to understand the internal decision-making processes or provide clear explanations for the output. Some of the software also remains ‘proprietary’ and therefore confidential. It can sometimes be difficult to interpret the results, due to the multilayer nonlinear model structures and the billions of parameters used. LLMs with Attention Mechanisms5 may give some ability to see on which parts of the input text the model focuses when generating a response, thereby providing some insights into the decision making. Generative AI LLMs can therefore complement and augment human processes to improve efficiency but should not be a substitute for the exercise of professional judgment, quality legal analysis and the expertise which clients, courts and society expect from barristers.18.与许多人工智能工具一样，生成式深度学习人工智能大型语言模型通常被认为是 "重型黑盒 "模型 ，因为很难理解其内部决策过程 ，也很难对其输出结果做出明确解释。有些软件还仍然是 "专有 " 的 ，因此具有保密性。由于多层非线性模型结构和使用的参数数以亿计 ，有时很难解释结果。具有关注机制的大型语言模型可以让人看到模型在生成响应时会关注输入文本的哪些部分，从而为决策提供一些启示。因此，生成式人工智能大型语言模型可以补充和增强人类流程，从而提高效率，但不应取代专业判断、高质量的法律分析以及客户、法院和社会对大律师所期待的专业知识。（3）Respect Legal Professional Privilege （LPP), Confidential Information and Data Protection Compliance（3）尊重法律职业特权、机密信息和数据保护合规性19.Be extremely vigilant not to share with a generative LLM system any legally privileged or confidential information （including trade secrets), or any personal data, as the input information provided is likely to be used to generate future outputs and could therefore be publicly shared with other users. Any such sharing of confidential information is likely to be a breach of Core Duty 6 and rule rC15.5 of the Code of Conduct, which could also result in disciplinary proceedings and/or legal liability.19.保持高度警惕 ，不要与生成式大型语言模型系统分享任何享有法律特权保护或机密保护的信息（包括商业秘密） 或任何个人数据 ，因为所输入信息很可能被用于生成未来的输出， 因此可能会与其他用户公开分享。任何此类共享机密信息的行为都可能违反 《行为守则》的核心职责6和 rC15.5条 ，还可能导致纪律处分程序和/或法律责任。20.Barristers will also need to comply with relevant data protection laws. You should never input any personal data in response to prompts from the system. Note that in March 2023, the Italian Data Protection Authority issued a temporary ban on ChatGPT, largely to investigate whether there was a lack of any legal basis for the collection and processing of any personal data used for training the system, and whether there was a lack of any proper notice to data subjects. Italy, France and Spain are currently investigating OpenAI’s processing of data. Using only synthetic data （that is data that is artificially created）on prompts to the LLM represents one possible way to avoid the risk of falling into breach of the General Data Protection Regulation （EU 2016/679）as retained in English law （UK GDPR).20.大律师还需要遵守相关的数据保护法律。切勿根据系统提示输入任何个人数据。值得注意的是，2023年3 月 ，意大利数据保护局对 ChatGPT 发布了临时禁令，主要是为了调查如下问题，即收集和处理用于训练系统的个人数据有无法律依据以及是否缺乏对数据主体的任何适当通知。意大利、法国和西班牙目前正在调查 OpenAI 的数据处理情况。在大型语言模型的提示中只使用合成数据 （即人工创建的数据）是避免违反《一般数据保护条例》（欧盟 2016/679）（仍保留在英国法中，称为英国 GDPR）风险的一种可能方法。21.As practitioners will be aware, the regulatory landscape in this area is in a state of flux and it is difficult to predict exactly what the UK position will be. Under the EU AI Act6 certain uses of AI tools in legal practice are categorised as ‘high-risk’ which triggers heightened regulatory obligations. The UK Government’s White Paper: A pro-innovation approach to AI regulation7 published in March 2023, suggests that existing regulators should act in accordance with five principles （similar to the OECD principles on AI8 although with different wording):
22.In the UK, the Information Commissioner has published guidance in relation to the development and use of technologies such as ChatGPT: “Generative AI: eight questions that developers and users need to ask” .22.在英国 ，英国信息专员发布了与 ChatGPT 等技术的开发和使用相关的指南："生成式人工智能：开发者和用户需要提出的八个问题 "。（4）Intellectual Property （IP）Infringement and Brand Association（4）知识产权侵权与品牌协会23.The precise interaction between the law of intellectual property and LLMs has not yet been the subject of definitive consideration. However, barristers will need to critically assess whether content generated by LLMs might violate intellectual property rights, especially third-party copyright. As a sizable amount of text data, such as books, papers, and other written materials were used to train ChatGPT and other LLMs, it is clearly possible that content produced may violate copyright or other IP rights in previously published materials. Several IP claims against generative AI owners have been lodged for allegedly unlawful copying and processing of millions of copyright-protected images, and associated metadata.23.知识产权法与大型语言模型之间的确切互动关系尚未有定论。不过，大律师需要严格评估大型语言模型所生成的内容是否可能侵犯知识产权 ，尤其是第三方版权。由于大量的文本数据（如书籍、论文和其他书面材料）被用于训练 ChatGPT 和其他大型语言模型，所生成的内容显然有可能侵犯以前出版材料的版权或其他知识产权。针对生成式人工智能所有者的一些知识产权索赔案件已经出现，据称其非法复制和处理了数百万张受版权保护的图片和相关元数据。24.Further, one should be careful not to use, in response to system prompts, words which may breach trademarks or give rise to a passing-off claim. Often the terms of service of a LLM give the company owning the LLM tools unlimited use of information given to the system.24.此外，在回应系统提示时 ，应注意不要使用可能侵犯商标权或引起假冒索赔的词语。通常，大型语言模型的服务条款规定拥有大型语言模型工具的公司可以无限制地使用向系统提供的信息。25.Irresponsible use of LLMs can lead to harsh and embarrassing consequences, including claims for professional negligence, breach of contract, breach of confidence, defamation, data protection infringements, infringement of IP rights （including passing off claims), and damage to reputation；as well as breaches of professional rules and duties, leading to disciplinary action and sanctions.25.不负责任地使用大型语言模型可能会导致严重而尴尬的后果， 包括专业疏忽索赔、 违约、违反保密规定、 诽谤、侵犯数据保护、侵犯知识产权（包括假冒索赔）和名誉损害；以及违反职业守则和职责，导致纪律处分和制裁。26.There is a growing body of material in which practitioners and others discuss their use of LLMs in the course of legal practice. This guidance is concerned only to explain some of the pitfalls. It is for barristers themselves to work out how and in what context a LLM might assist them in providing legal services. This process is likely to be a changing one as the technology itself develops as it is doing and with increasing speed.26.执业人员和其他人士讨论在执业过程中使用大型语言模型的材料越来越多。本指南只是解释其中的一些陷阱。至于如何以及在何种情况下使用大型语言模型来协助提供法律服务，则应由大律师自己来解决。这个过程很可能是一个不断变化的过程，因为技术本身正在以越来越快的速度发展。27.Barristers should also keep abreast of relevant Civil Procedure Rules, which in the future may implement rules/practice directions on the use of LLMs；for example, requiring parties to disclose to the court when they have used generative AI in the preparation of materials. This approach has already been adopted by the Court of the King’s Bench in Manitoba.27.大律师也应紧跟相关《民事诉讼程序规则》的规定。未来此类规定可能会实施有关使用大型语言模型的规则/实务指示，要求当事人向法庭披露他们在准备材料时使用了生成式人工智能。这种方法已被马尼托王座法庭采用。28.In conclusion, technical progress and the pressures of competition may lead to the increasing adoption of AI, including LLMs. The best-placed barristers will be those that make the effort to understand these systems and, if appropriate, use them as tools in their practice, while maintaining control and integrity in their use. There is nothing inherently improper about using reliable AI tools for augmenting legal services；but they must be properly understood by the individual practitioner and used responsibly, ensuring accuracy and compliance with applicable laws, rules and professional codes of conduct.28.总之 ，技术进步和竞争压力可能导致越来越多地采用人工智能，包括大型语言模型。最有能力的大律师将是那些努力了解这些系统，并在适当的情况下将其作为执业工具，同时在使用过程中保持控制和诚信的大律师。使用可靠的人工智能工具来增强法律服务本身无可厚非；但从业者个人必须正确理解并负责任地使用这些工具，确保其准确性，并遵守适用的法律、规则和专业行为准则。This document and sample policy has been prepared by the Bar Council to assist barristers and chambers on matters of information security. It is not "guidance" for the purposes of the BSB Handbook I6.4, and neither the BSB nor bodies regulating information security, nor the Legal Ombudsman is bound by any views or advice expressed in it. It does not comprise - and cannot be relied on as giving - legal advice. It has been prepared in good faith, but neither the Bar Council nor any of the individuals responsible for or involved in its preparation accept any responsibility or liability for anything done in reliance on it. For fuller information as to the status and effect of this document, please see here.本文件和政策样本由大律师公会编制，旨在就信息安全事宜为大律师和大律师事务所提供帮助。本文件并非《英国大律师标准委员会手册》第 I6.4 条所指的 "指南" ，英国大律师标准委员会、信息安全监管机构或法律申诉专员均不受本文件所表达的任何观点或建议的约束。本指南不包含法律意见，也不能将其视为法律意见。本指南是基于善意起草的，但大律师公会或任何负责或参与编写本报告的个人，均不对因依赖本指南而作出的任何行为承担任何责任或法律责任。原文：英国大律师公会，翻译：采安律师事务所
