萧咏仪、姚君源:AI在调解中的应用(Artificial Intelligence in Mediation)

Supporting Reframing through Neutral Language

透过中性语言支持重新表述

Neutral language is a cornerstone of effective mediation. It emphasises facts, avoids blame, and promotes constructive dialogue. One of the key skills of a mediator is the ability to reframe heated or accusatory statements into neutral language that reduces tension and encourages collaboration. Ideally, issues should be expressed in positive and balanced terms wherever possible.

中性语言是有效调解的基石。中性语言强调事实，避免指责，促进建设性对话。调解员的关键技能之一是能够将激烈或指责性的陈述重新表述为中性语言，从而缓解紧张气氛，鼓励合作。在理想情况下，争议事项应尽可能以正面而平衡的方式表达。

This task often requires mediators to think quickly under pressure, which can be particularly challenging for less experienced mediators when parties are highly emotional. LLMs can provide valuable assistance by neutralising language almost instantaneously.

这项任务通常要求调解员在压力下迅速思考，对经验较浅的调解员而言，当事人情绪高涨时，这项工作尤其具挑战性。大型语言模型可以几乎即时将带有对抗性的语句转化为较中性的表述，因而能为调解员提供有价值的支援。

For example, if Party A were to state: “Your company completely lied to us and wasted our money. You never intended to deliver on time, and now my whole business is suffering because of your dishonesty.”

例如，如果甲方想说：「贵公司完全欺骗了我们，浪费了我们的钱。你们根本就没打算按时交货，现在我盘生意都因为你们的不诚实而遭受损失。」

An LLM could reframe the statement as: “What I hear is that you are concerned about delays in delivery and the impact this has had on your business operations. You would like clarity on whether the original commitments can still be met and how future cooperation can avoid similar issues.”

大型语言模型可以把这句话重新表述为：「我听到的是，您关注交货延误，以及这对贵公司业务运作造成的影响。您希望厘清原先的承诺是否仍能履行，并了解未来合作可以如何避免再次出现类似问题。」

In this way, AI can serve as an additional resource to alleviate the cognitive burden on mediators and keep discussions on track.

这样AI可以作为额外的资源，减轻调解员的认知负担，使讨论保持正轨。

Assisting Mediators in Understanding the Case

协助调解员了解案件

 Mediation cases can range widely, from family disputes to complex corporate conflicts. Some matters may involve highly technical or professional subject matter, such as medical, technological, or legal terminology, which can be difficult for a mediator without specialised training to fully grasp. This can make it challenging for the mediator to follow the discussion or to respond effectively during the session.

调解案件的类型非常广泛，从家庭纠纷到复杂的公司冲突均包含在内。有些案件可能涉及高度专业或技术性的内容，例如医学、科技或法律术语。对未受相关专门训练的调解员而言，要全面掌握这些内容往往并不容易，从而增加其在调解过程中跟进讨论及作出有效回应的难度。

LLMs can assist by helping mediators prepare for such cases and by simplifying complex terminology into plain language. With this support, mediators are better equipped to grasp technical issues and to engage with the parties. Importantly, if a party introduces specialised terms during the session, the LLMs can provide immediate clarification, allowing the mediator to stay on track and guide the discussion. 

大型语言模型可以协助调解员为这类案件做好准备，并将复杂术语转化为浅白易明的语言。在此支援下，调解员能更有效理解技术性议题，并与各方进行有意义的沟通。更重要的是，如任何一方在调解过程中使用专业术语，大型语言模型亦可即时提供释义，协助调解员掌握讨论脉络，并引导对话有序推进。

For example, consider the following party statement: “The defendant breached the licensing agreement by reverse-engineering our encryption algorithm, which violates the non-compete clause and exposes us to GDPR compliance risks.” 

例如，考虑以下当事人的陈述：「被告通过逆向工程破解了我们的加密演算法，违反了许可协议，这不仅违反了禁止竞争条款，亦使我们面对GDPR合规风险。」

An LLM might render this into simpler terms for the mediator: “The concern is that the other side copied and used your data security system in a way that may breach the contract and could also raise legal issues about data protection.” 

大型语言模型可能会用更通俗易懂的方式向调解员解释：「问题在于，对方复制并使用了你方的数据安全系统，而该做法可能违反合约，亦可能引起有关数据保障的法律问题。」

By translating technical concepts into accessible language, AI enables mediators to maintain control of the process and to ensure that discussions remain clear and productive, even in highly specialised disputes.

透过将技术概念转化为易于理解的语言，AI可协助调解员在高度专业化的争议中仍然掌握调解程序，并确保讨论保持清晰及具成效。

Facilitating the Right Questions 

协助提出正确的问题

Another critical skill in mediation is the ability to ask the “right” question. Such questions help uncover the parties’ underlying interests, test the reality of their positions, and encourage them to consider alternative options. However, identifying and framing the right question is not straightforward. A question is “right” not only if it is objectively relevant, but also if it is phrased in a way that the party can subjectively understand and meaningfully respond to. In practice, the effectiveness of a question often depends on the party’s education level, age, professional background, and communication style. A question that resonates with a university professor may not work for an investment banker, and vice versa. 

调解的另一项关键技能，是提出「恰当」问题的能力。这类问题有助揭示各方的深层利益、测试其立场在现实中的可行性，并鼓励他们考虑其他可行方案。然而，要识别并妥善表达恰当的问题，并不容易。问题是否「恰当」，不仅取决于其在客观上是否相关，亦取决于其措辞是否能让当事人从其自身角度理解，并作出有意义的回应。在实务上，问题是否有效，往往取决于当事人的教育程度、年龄、专业背景及沟通方式。 一条能引起大学教授共鸣的问题，未必适用于投资银行家；反之亦然。

LLMs can provide valuable assistance in this area. As Judge Newsom observed in Snell v. United Specialty Insurance Co., 102 F.4th 1208 (11th Cir. 2024), these models are trained on vast amounts of ordinarylanguage input, ranging from Hemingway novels and Ph.D. dissertations to gossip rags and comment threads. This broad training enables them to tailor the style and tone of a question to suit the audience, while preserving its substance. 

大型语言模型在这方面可以提供宝贵帮助。正如 Newsom法官在 Snell v. United Specialty Insurance Co.，102 F.4th 1208 (11th Cir. 2024) 案指出，大型语言模型接受过大量日常语言输入训练，涵盖范围从海明威的小说、博士论文到八卦杂志和评论区。这种广泛的训练使它们能够根据受众调整问题的风格和语气，同时保留问题的实质内容。

For example, when a mediator is asking a university professor to uncover underlying interests, a more reflective and academic tone may be effective: “Is your main concern the immediate financial shortfall, or how this delay affects your reputation and future cooperation?”   

例如，当调解员要求大学教授揭露潜在利益时，采用更具反思性和学术性的语气可能会更有效：「您主要关注的，是眼前的资金短缺，还是这次延误对您的声誉及日后合作所造成的影响？」

By contrast, when a mediator is asking a trader to uncover underlying interests, a more direct and practical approach may be preferable: “What’s more important now—quick cash in, or keeping the counterparty for future deals?”

相较之下，当调解员向交易员提问，以发掘其深层利益时，较直接和务实的表达方式可能更为合适：「现在什么更重要 : 快速套现，还是留住对手以备将来交易？」

In this way, LLMs can support mediators by generating tailored questions that align with the parties’ backgrounds and communication styles, which increases the likelihood of constructive dialogue. 

透过这种方式，大型语言模型可协助调解员拟定切合各方背景及沟通风格的提问，从而提高建设性对话的可能性。

Generating Settlement Options 

生成和解方案

AI can also support mediators in generating settlement options. Because LLMs are trained on vast amounts of data, including examples of past disputes and resolutions, they can help mediators brainstorm a broad range of possible solutions. This can encourage parties to move beyond preliminary positions and consider creative or alternative approaches that they might not otherwise have identified on their own.

AI亦可协助调解员构思和解方案。由于大型语言模型是以大量资料训练而成，其中包括过往争议及其解决方式的例子，因此可协助调解员就可能的解决方案进行更广泛的构思。这有助鼓励各方跳出原有的初步立场，并考虑一些原本未必会想到的创新或替代方案。

Drafting Settlement Agreements

起草和解协议

 An ideal outcome in mediation is for the parties to sign a settlement agreement promptly once consensus has been reached. In practice, however, delays often arise because of the time needed to prepare and refine the written agreement. LLMs, particularly legal-focused products such as Thomson Reuters CoCounsel, can assist mediators by generating draft settlement agreements and suggesting edits. This support helps streamline the process and allows the parties, after obtaining proper legal advice and carefully considering the terms, to finalise and sign the agreement without unnecessary delay.

调解的理想结果是各方在达成共识后立即签署和解协议。然而，在实务中，由于草拟和完善书面协议需时，故常常会出现延误。大型语言模型（特别是Thomson Reuters CoCounsel等法律专用产品）可以协助调解员生成和解协议草案，并提出修改建议。这种支援有助简化流程，使各方在获得适当的法律建议及仔细考虑条款后，能够最终确定并签署协议，从而避免不必要的延误。

While AI offers promising benefits, its use in mediation is not without risks. Mediators must exercise caution and remain alert to potential challenges when integrating AI into the mediation process. The following four key risks deserve particular attention:-

尽管AI具有可观的潜在助益，但其在调解中的应用并非毫无风险。调解员在将AI纳入调解过程时，必须审慎行事，并对可能出现的挑战保持警觉。以下四项关键风险尤其值得注意：

1. AI Hallucinations 

AI幻觉

Language models can generate answers that appear plausible but are in fact false. These “hallucinations” may arise even in response to simple questions and cannot be eliminated by prompting alone. If left unchecked, hallucinations may produce unreasonable or misleading statements, or even questions that inadvertently appear to favour one party. Such outputs could inflame emotions, damage trust, and ultimately derail the mediation. 

语言模型可能会产生看似合理实则错误的答案。即使面对简单问题，亦可能出现此类「幻觉」，而且不能单靠提示语（prompting）加以消除。 1如不加以查核，这些幻觉可能产生不合理或具误导性的陈述，甚至提出无意间令人感到偏袒某一方的问题。此类输出可能激化情绪、破坏信任，最终令调解偏离正轨。

2. Bias in AI Algorithms 

AI演算法中的偏见

LLMs are trained on vast datasets that may reflect historical or structural biases. As a result, their outputs can unintentionally reinforce stereotypes or discriminatory patterns. If a mediator were to adopt such outputs uncritically, this could compromise the impartiality of the process. Maintaining neutrality requires that mediators critically review and adapt any AI-generated suggestions rather than relying on them wholesale. 

大型语言模型基于庞大的资料集进行训练，这些资料集可能反映出历史或结构性偏见。因此，其输出结果可能会无意中强化刻板印像或歧视模式。如果调解员不加批判地采纳这些输出结果，可能会损害调解过程的公正性。为了保持中立，调解员需要批判性地审查和调整AI生成的任何建议，而不是全盘依赖AI。

3. Loss of Human Interaction 

人际互动减少

AI tools currently lack the ability to fully capture or respond to human emotion. A skilled mediator can recognise subtle expressions of frustration, anger, or anxiety and adapt their approach accordingly. Over-reliance on AIgenerated questions or statements risks overlooking these emotional dimensions,weakening the human connection that is essential to building trust and rapport. Mediation is ultimately a human-centred process, and no AI tool can replace the importance of empathy and interpersonal sensitivity. 

AI工具目前尚不具备全面捕捉或回应人类情感的能力。经验丰富的调解员能够辨识出细微的沮丧、愤怒或焦虑情绪，并据此调整调解方式。过度依赖AI生成的问题或陈述，可能会忽略这些情感层面，从而削弱建立信任和融洽关系所必需的人际连结。调解始终是以人为本的过程，任何AI工具均无法取代同理心和人际敏感度的重要性。

4. Confidentiality and Disclosure 

保密和披露

Confidentiality is a core principle of mediation, enshrined in Section 8 of the Mediation Ordinance (Cap. 620), which prohibits disclosure of mediation communications except in limited circumstances set out in subsections 8(2) and 8(3). It remains unclear whether sharing mediation communications with an LLM, particularly when the model is hosted on cloud servers, would constitute a breach of these provisions. To avoid any doubt, parties’ express consent should be obtained before AI is used to assist in mediation. 

保密是调解的核心原则，《调解条例》（香港）（第620章）第8条订明，除按第8(2)或8(3)款的规定，任何人不得披露调解通讯。将调解通讯内容输入大型语言模型（尤其是模型设于云端伺服器时），是否会构成违反上述规定，目前仍未有明确答案。为免产生疑问，在以AI辅助调解前，应先取得各方当事人的明示同意。

The IBA Guidelines provide useful reference points. Part Two of the Guidelines stresses that users of AI must take reasonable steps to ensure that confidential information is not compromised, noting that data entered into proprietary or open-source LLMs may be vulnerable to data breaches and unintended disclosure.The Guidelines recommend measures such as anonymising inputs, limiting the amount of information entered to what is necessary to achieve the desired outputs, and reviewing the privacy policies of any AI tools before use.3 They also include a sample disclosure statement in Part Three, which mediators or parties can adopt to inform others that AI tools are being used and the potential risks.  

IBA指引提供了具参考价值的实务指引。指引第二部分强调，AI使用者必须采取合理措施，确保机密资讯不受损害，并指出输入至专有或开源大型语言模型的资料，可能面临资料外洩及非预期披露的风险。该指引并建议采取多项措施，例如将输入资料匿名化、把输入内容限制在达致所需输出所必须的范围内，以及在使用前审阅相关AI工具的私隐政策。指引第三部分提供披露声明范本，调解员或当事人可采用此声明通知他人正在使用AI工具及其潜在风险。

Most of the technical limitations of LLMs can be managed by the human mediator. It is essential to remember that AI serves only as an assistance tool, not as a substitute for the mediator. From this perspective, the “Two Guiding Rules” set out in the Guidelines on the Use of Generative Artificial Intelligence for Judges and Judicial Officers and Support Staff of the Hong Kong Judiciary provide a useful reference point for mediators, even though they were drafted for the judiciary rather than mediation practice.

大多数大型语言模型的技术限制都可以由人类调解员来应对。必须牢记，AI仅作为辅助工具，而非调解员的替代品。从这个角度来看，《香港司法机构法官及司法人员和支援人员使用生成式人工智能的指引》列出的「两项指导原则」虽然是为司法工作而非调解实务而制定，仍可作为调解员的有用参考。

The first rule is that mediator functions must never be delegated to AI. The second rule is that mediators must remain vigilant about the output generated by the AI chatbots, in particular, factual accuracy, potential bias, infringement of intellectual property rights, and use it at the mediator’s own risk. The mediator should take responsibility for using AI and for the end product. 

第一个原则是，调解员的职能绝不能委讬给AI。第二个原则是，调解员必须对AI聊天机器人生成的输出保持警惕，尤其要关注事实准确性、潜在偏见、知识产权侵权等问题，并自行承担使用风险。调解员应对AI的使用及其最终结果负责。

AI has significant potential to assist mediators, but its limitations, including hallucinations, bias, loss of human interaction, and confidentiality concerns, demand vigilance and sound professional judgment. While the IBA has already issued guidelines setting out both applications and safeguards, Hong Kong has yet to develop a comparable framework, and establishing such guidance would help ensure that AI is integrated responsibly, in line with the principles of neutrality, confidentiality, and party autonomy that underpin mediation. The challenge ahead lies in striking the right balance between embracing innovation and safeguarding the integrity of the mediation process.

AI在协助调解方面具有巨大潜力，但亦有其局限性，包括可能出现幻觉、偏见、缺乏人际互动及保密性问题，故调解员必须保持警觉及运用专业判断力。IBA 已发布相关指引，阐述AI的应用和相关保障措施；相对而言，香港尚未建立相应的本地框架。制定此类指引有助确保AI能够以负责任的方式融入调解工作，并符合中立、保密和当事人自主等调解原则。未来的挑战在于如何在拥抱创新与维持调解程序的完整性之间，取得适当平衡。

来源：多元化纠纷解决机制