Azerbaijan AI Ecosystem Automates Court Operations to Offset Judicial Deficit

阿塞拜疆利用AI生态推动法院自动化以弥补司法资源短缺 部署专业AI平台将法律案件准备时间缩短至数分钟，同时提升了相当于167名法官的有效司法运力。 本次演讲概述了阿塞拜疆的主权AI战略及其法律AI产品生态系统，旨在重塑公众获取法律服务的途径和法院运作模式。演讲重点介绍了面向公众的引擎e-Qanun AI，该引擎利用索引化的国家法规简化法律检索，并解答公民的自然语言咨询。为应对法院不断攀升的案件量及结构性法官短缺，阿塞拜疆推出了e-Court AI用于处理案卷、梳理判例并生成判决草案。该司法助手大幅缩短了审前准备时间，切实提升了全国的司法运力。 阿塞拜疆国家人工智能中心依托配备 NVIDIA GPU 的主权超算中心推进本土 AI 战略，构建了涵盖 AI Academy、Data Races、e-Qanun AI 和 e-Court AI 四大核心产品的生态系统。 (0:04) 面向公众的法律智能平台 e-Qanun AI 索引了超过 5.7 万份法律文件和 1250 万个句子，支持自然语言检索与精准法条引用，引用幻觉率低至 0.07%，已有逾 3.5 万名公民使用。 (3:49) 2022 至 2025 年间阿塞拜疆法院年案件量激增 23.8% 突破 60 万件，法官人均年办案 865 件，且该国每 10 万人仅拥有 6.8 名法官，远低于欧洲 22 名的中位数水平。 (8:18) 司法辅助系统 e-Court AI 可自动提取案件事实、检索先例并生成裁判文书草稿，将复杂案件的准备时间从数天缩短至 10-15 分钟，预计可节省超 30 万审判工时，生产力提升相当于增设 167 名法官。 (12:23)

To address severe judicial shortages and expanding caseloads without compromising judicial independence, the National Artificial Intelligence Center of Azerbaijan is deploying sovereign, locally trained AI systems to support both public legal literacy and court administration. Sovereign Infrastructure and Azerbaijan's AI Strategy (0:04) Rather than adopting foreign commercial models, Azerbaijan chose to build native AI capabilities aligned with a four-pillar national strategy: developing talent, establishing responsible governance, building local data infrastructure, and fostering public-private investment. To safeguard sensitive citizen data, the country operates a sovereign supercomputer center equipped with NVIDIA GPUs for localized model training. This infrastructure supports an interconnected four-part ecosystem: the AI Academy for engineering and applied mathematics training; the Data Races platform for public and private machine learning competitions; e-Qanun AI for citizen-facing statutory interpretation; and e-Court AI for judicial case preparation. Citizen Legal Access via e-Qanun AI (3:49) Recognizing that statutory complexity, technical terminology, frequent amendments, and the limitations of keyword search create barriers for non-lawyers, the center built e-Qanun AI. The platform allows citizens to ask questions in natural Azerbaijani, using semantic retrieval to cross-reference multiple statutes and deliver plain-language explanations tied directly to official legislation. The system has indexed over 57,000 consolidated legal documents and processed more than 12.5 million sentences for clause-level search, ingesting over 10,000 new or amended documents monthly. The center reports a 96.88% generative accuracy rate, 98% reference accuracy, a 0.07% citation hallucination rate, and an active user base exceeding 35,000 citizens. Caseload Expansion and Structural Judge Shortages (8:18) As legal accessibility increases, court filings are growing rapidly. Between 2022 and 2025, annual court cases in Azerbaijan grew 23.8%, from 485,000 to over 600,000. Over the same period, the bench increased 21.3%, from 554 to 672 active judges, resulting in an average annual load of 865 cases per judge across more than 242,000 civil cases and 143,000 administrative offenses. This volume highlights a deep structural deficit: Azerbaijan maintains 6.8 judges per 100,000 people, roughly 3.2 times below the European median of 22 judges per 100,000 people. e-Court AI Workflow and Capacity Projections (12:23) To absorb expanding caseloads, the center piloted e-Court AI as an administrative assistant rather than a decision-maker. The tool reviews case files, extracts material facts and party claims, identifies relevant precedents, maps facts to statutory provisions, and generates a structured draft decision for the judge to accept, modify, or reject with feedback that retrains the underlying model. The center reports that case preparation times fall from 15–30 minutes to approximately 5 minutes for simple cases, from 3 hours to 7–10 minutes for average cases, and from 4–5 days to 10–15 minutes for complex cases. Across the judiciary, pre-review processing speeds are estimated to improve by more than 18 times, projecting over 300,000 judicial hours saved—a capacity gain the center equates to adding 167 judges and raising effective capacity from 6.8 to 8.4 judges per 100,000 people. The center maintains that judicial independence is preserved because human judges retain sole authority to edit or reject AI draft decisions. Planned future updates include public-facing predictive tools that would allow citizens to submit case documents and evaluate potential legal outcomes prior to formal litigation.

阿塞拜疆国家人工智能中心表示，该国正通过构建主权人工智能生态系统应对司法资源短缺与案件激增难题，在恪守司法独立的前提下，通过公众端法律问答平台与法官端预审辅助系统提升司法系统的可及性与运作效率。 主权AI战略与四大支柱生态 (0:04) 阿塞拜疆国家人工智能中心指出，虽然人工智能正在重塑全球医疗、金融和制造业，但法律领域的敏感性引发了公众对透明度、信任度及算法理解复杂法条能力的审视。为避免依赖外国技术并保护公民敏感数据，阿塞拜疆选择依托国家人工智能战略自主研发本土能力。该战略围绕四大核心支柱展开：AI人才培育、负责任采用与公共服务透明度的AI治理、依托配备NVIDIA GPU主权超算中心的本土数据与基础设施，以及促进公私合作的创新投资生态。 为落实战略规划，中心整合开发了四款核心产品体系：开展应用数学到工程技能培训的人工智能学院（AI Academy）、面向政企实际需求提供机器学习挑战的竞赛平台（Data Races Platform）、面向普通公众的法律智能平台e-Qanun AI（Aigun AI），以及面向法官案卷分析、先例检索与裁决草案生成的试点系统e-Court AI。 e-Qanun AI与公民法律服务 (3:49) 针对公众在查阅法律时面临的“仅能获取分散条文而非直接解答、术语晦涩专业、法条频繁修订、传统关键词搜索无法匹配日常场景诉求”等四大痛点，中心推出了e-Qanun AI。该平台支持阿塞拜疆语自然语言交互，通过语义理解检索具体含义，结合法律推理生成清晰释义，并严格建立与官方条文源链接的一致对应，以保障结论的可验证性。 根据中心披露的运行指标，该平台已完成对阿塞拜疆5.7万余份法律文书的整合索引，处理了超过1250万个独立句子以实现条款级语义检索，每月持续吸收超过1万份新颁布或修订的法律文件。在性能测试中，平台展现出96.88%的生成准确率与98%的引用准确率，引用幻觉率控制在0.07%，并可在1分钟内输出法律分析摘要。目前该平台已有超过3.5万名活跃公民用户。 案件量激增与法官资源短缺 (8:18) 随着公众法律素养和维权渠道便利性的提升，阿塞拜疆法院受理的诉讼量显著增加。官方数据显示，2022年至2025年间，全国法院立案数量从48.5万件增至60万件以上，增幅达23.8%；同期在任法官人数由554名增加至672名，增幅为21.3%，这导致每位法官年均承办案件量高达865件。 除工作量攀升外，阿塞拜疆还面临相较于国际基准的法官编制缺口。目前该国每10万人口仅对应6.8名法官，约为欧洲中位数（每10万人对应22名法官）的三分之一。每年法院审理的案源结构高度集中，包括超过24.2万起民事案件和14.3万起行政违法案件，同时还需处理刑事及上诉审理程序，司法系统承载着极高的运转负荷。 e-Court AI预审架构与效益预测 (12:23) 为缓解审判压力，e-Court AI被设计为正式司法审查前的助理工具，协助法官完成案卷审阅、事实与诉求提取、司法先例检索、法律条文匹配以及自动生成结构化裁决草案。该系统保留了法官最终的决定权，法官可审核、修改或驳回草案；若法官选择驳回，必须输入驳回理由，该反馈将用于后续模型的重新训练与优化。 中心测试显示，AI辅助显著缩短了各类型案件的预审准备时间：简单案件从15至30分钟压缩至约5分钟，中等案件从约3小时减至7至10分钟，复杂案件则由4至5天缩短至10至15分钟。基于这一表现，中心预计该系统能将预审处理速度提升18倍以上，累计节约超过30万个法官工时，其生产力增益相当于为系统直接补充167名在任法官，有望在不增加实际编制的情况下，将该国有效司法运力等效提升至每10万人约8.4名法官。中心后续还计划支持公民自主上传文书并获取判决结果预测。 阿塞拜疆展示了利用主权数据中心与本土语义模型缓解司法资源供需矛盾的技术路径；然而，系统对实际判决质量的长远影响、法官在审判实践中对AI草案的实际采纳情况，以及未来面向公众开放诉讼结果预测时所伴随的程序正义与算法问责风险，仍有待更长时间维度的实践检验。

## Background 本内容由 Ai4 频道于 2026 年 8 月 15 日在 YouTube 平台发布，演讲者为阿塞拜疆国家人工智能中心首席商务官 Ulfan Aliyev。视频主要介绍了阿塞拜疆如何通过部署集成式 AI 解决方案来赋能司法基础设施，既面向公众简化晦涩法条，又在法庭端协助法官提炼案卷、检索先例与起草裁判文书。 ## TL;DR 面对法官运力严重不足与公民诉讼需求激增的双重挑战，阿塞拜疆基于主权算力基础设施打造了国家级法律 AI 生态闭环。面向公众的 e-qanun AI 平台通过句子级索引和语义理解，将 57,000 余部法规转化为通俗易懂且带有权威引注的解答，大幅降低了法律理解门槛；面向法官的 e-Court AI 则在坚守“辅助审判而非替代法官”的红线下，自动化处理案卷事实提炼与判决草案生成，预计每年可节省逾 30 万法官工时。整套体系通过法官裁决反馈持续迭代，并计划进一步向公众开放案件模拟与诉讼预期评估，实现司法效率与透明度的系统性升级。 ## Atmosphere & style 整场演讲呈现出严谨务实且极具前瞻性的国家级战略宣讲风格，演讲者逻辑清晰、数据详实，始终洋溢着对技术赋能公共治理的自信。叙事从公众对司法 AI 的天然疑虑与阿塞拜疆人均法官运力严重匮乏的现实痛点切入，随后层层递进地展开产品矩阵与架构设计，兼具宏观顶层设计与扎实的技术落地细节。听感上冷静专业、克制理性，在反复强调“算法辅助但司法独立不容妥协”的伦理边界时透露出深思熟虑的庄重感。 ## Key takeaways - [00:04](https://www.youtube.com/watch?v=DNz4wDeffNE&t=4s) 司法 AI 的核心信条在于扮演辅助与赋能的角色，在提升司法效率与透明度的同时，绝不替代法官的最终裁量权与司法独立性。 - [01:21](https://www.youtube.com/watch?v=DNz4wDeffNE&t=81s) 阿塞拜疆坚定选择自主研发的国家 AI 战略，通过布局 AI 人才、治理、主权超算基础设施及投资生态四大支柱，确保数据安全与核心技术自主。 - [02:31](https://www.youtube.com/watch?v=DNz4wDeffNE&t=151s) 国家人工智能中心构建了由 AI Academy、Data Race、e-qanun AI 与 e-Court AI 组成的产品矩阵，打通了从人才孵化、算法攻坚到司法落地的全链条闭环。 - [03:49](https://www.youtube.com/watch?v=DNz4wDeffNE&t=229s) 破解公众法律困境的关键在于转变设计思维：不再强求公民学习复杂的法学术语，而是利用自然语言技术让法律开口说大众听得懂的日常语言。 - [05:32](https://www.youtube.com/watch?v=DNz4wDeffNE&t=332s) 面向公民的 e-qanun AI 平台对 1,250 万个法条语句进行句子级深度切片索引，法条引用准确率达 98%，引用幻觉率低至 0.07%，且强制标注权威法律出处。 - [08:18](https://www.youtube.com/watch?v=DNz4wDeffNE&t=498s) 公众法律意识觉醒与维权门槛降低促使阿塞拜疆诉讼案件量 3 年激增 23.8%，法官人均年办案达 865 件，倒逼司法体系必须通过数字化手段提升吞吐能力。 - [11:03](https://www.youtube.com/watch?v=DNz4wDeffNE&t=663s) 阿塞拜疆每 10 万人仅拥有 6.8 名法官，远低于欧洲 22 名的中位数，标准化且数量庞大的民事与行政案件对传统法官运力构成了严重瓶颈。 - [12:23](https://www.youtube.com/watch?v=DNz4wDeffNE&t=743s) 专为法官打造的 e-Court AI 专注于案卷事实提炼与判决草案生成，预计每年可节省超过 30 万工时，相当于为全国司法系统增补 167 名法官的等效生产力。 - [15:31](https://www.youtube.com/watch?v=DNz4wDeffNE&t=931s) 系统建立了严格的人机协同反馈机制，法官在驳回或修改判决草案时填写的法理理由会被沉淀为高质量训练数据，用于大模型的持续迭代优化。 - [15:31](https://www.youtube.com/watch?v=DNz4wDeffNE&t=931s) 未来计划向公众开放纠纷诉讼模拟与走向预测功能，帮助当事人在起诉前理清法律权责，从源头上减少非理性诉讼并降低司法社会成本。 ## Detailed walkthrough ### [00:04](https://www.youtube.com/watch?v=DNz4wDeffNE&t=4s) 司法领域的 AI 探索：辅助而非替代 人工智能正在全球范围内重塑医疗、金融和制造业等各个行业，所有人都在思考“AI 如何让我们的工作变得更好”。然而，当这场变革触及司法体系时，讨论瞬间变得极为敏感。公众和法律界会本能地产生疑虑：“AI 真的值得信任吗？它能理解复杂的法律规则吗？”这些问题正是阿塞拜疆国家人工智能中心（National Artificial Intelligence Center）在启动项目时面临的核心追问。 在过去一年中，团队与法官、法律专家及公共机构紧密合作，共同探索一个核心命题：在不牺牲公众信任、透明度以及司法独立性的前提下，AI 能否让司法变得更加普惠与高效？这一探索最终促成了两大 AI 平台的诞生与一套完整的法律生态系统架构。中心的核心信条非常明确：AI 在司法系统中应当扮演“支持者”与“赋能者”的角色，而绝非取而代之。 ### [01:21](https://www.youtube.com/watch?v=DNz4wDeffNE&t=81s) 阿塞拜疆的国家 AI 战略与主权算力基础设施 面对 AI 重塑政府治理与经济结构的全球浪潮，阿塞拜疆面临着一个根本性抉择：是简单地采购和消费他国开发的技术，还是建立自主的本土能力？阿塞拜疆坚定地选择了自主研发的道路，并制定了目标明确的国家 AI 战略。 阿塞拜疆的国家人工智能战略构建在四大支柱之上： - **AI 人才（AI Talent）**：聚焦人力资本投资，建立从应用数学到 AI 工程的完整教育培养链。 - **AI 治理（AI Governance）**：推行负责任的 AI 应用，保障系统透明度，强化公共服务效能。 - **数据与基础设施（Data & Infrastructure）**：在本土建设现代数据中心，打造配备 NVIDIA GPU 的国家主权超级计算中心（Sovereign Supercomputer Center），确保公民敏感数据在本土环境安全受控地用于模型训练。 - **创新与投资生态（Innovation & Investment Ecosystem）**：搭建推动公私部门无缝协作的创新机制。 ### [02:31](https://www.youtube.com/watch?v=DNz4wDeffNE&t=151s) 核心产品矩阵：从人才培养到司法落地闭环 为了实现战略落地，国家人工智能中心构建了四大协同产品，形成从人才孵化、实战演练到端到端业务落地的完整生态闭环： ```mermaid flowchart LR accTitle: 阿塞拜疆 AI 产品矩阵架构 subgraph Talent["人才与实战演练"] A["AI Academy<br/>培养算法与工程人才"] --> B["Data Race 竞赛平台<br/>承接政企真实场景挑战"] end subgraph Infrastructure["主权技术底座"] C["主权超算中心<br/>NVIDIA GPU 本地算力"] end subgraph Applications["司法法律垂直应用"] D["e-qanun AI<br/>面向公民的法律智能平台"] E["e-Court AI<br/>面向法官的案卷分析与草拟系统"] end B --> Applications C --> Applications ``` - **AI Academy（AI 学院）**：专注于底层人力资本建设，为学生提供从应用数学到全栈 AI 工程的系统化培养。 - **Data Race 平台**：阿塞拜疆首个基于机器学习的算法挑战与竞赛平台，政府机构和私营企业将真实痛点发布于此，由学生及生态开发者实战攻坚。 - **e-qanun AI**：面向全体公民的法律智能交互平台，用自然语言化解复杂的法条壁垒。 - **e-Court AI**：专为法官设计的辅助审判平台（年内启动试点），协助法官调阅案卷、检索先例并自动生成结构化的判决书草案。 ### [03:49](https://www.youtube.com/watch?v=DNz4wDeffNE&t=229s) 破解公众法律困境：让法律说人民的语言 项目团队并未一开始就扎进算法模型中，而是退回原点思考最本质的问题：“既然现行法律法规全部对公众公开，为什么普通人运用法律依然如此艰难？”调研发现，症结并不在于法规的获取渠道，而在于“看得到却读不懂”。普通公民面临着四大结构性障碍： 1. **得到的是卷帙浩繁的文档而非直接答案**：咨询一个简单的法律问题，可能需要通读数部法律和几十条具体法条。 2. **法条语言天然晦涩**：立法语料专为专业律师和法学学者编写，对普通百姓而言犹如天书。 3. **法律修正追踪困难**：法律处于高频动态修订中，即使是法律专业人士也极难实时跟进最新条款。 4. **传统关键词搜索的局限**：公民不知道该用什么法学术语作为关键词进行检索。 普通人不会以“第几条第几款”来思考问题，他们脑海中浮现的是真实的生活场景：“我的雇主可以这么做吗？”、“我到底有没有这项权利？”、“我下一步该怎么办？”这促使团队彻底转变了设计思维——**不再要求公民学习法律的语言，而是让法律开口说人民的语言**。 ### [05:32](https://www.youtube.com/watch?v=DNz4wDeffNE&t=332s) e-qanun AI：高精度语义检索与可验证法律智能 基于上述理念打造的 e-qanun AI 并非另一个单纯存储法条的数据库，而是一个让法律知识易于理解、检索和应用的法律智能平台。公民只需使用日常阿塞拜疆语输入提问，系统在后台执行四项核心任务： ```mermaid flowchart TD accTitle: e-qanun AI 核心工作流程 Q["公民自然语言提问<br/>（阿塞拜疆语）"] --> NLU["自然语言理解"] NLU --> Semantic["语义级法律检索<br/>（基于句子粒度理解法律含义）"] Semantic --> Reasoning["法律推理与通俗解释生成"] Reasoning --> Verifiable["输出附带官方出处引注的答案"] ``` 在法律领域，“显得很有信心”毫无意义，真正的公信力来自于“精准展示答案为何正确”。因此，e-qanun AI 强制要求生成的每一条解释都必须直接锚定官方权威法条出处。 - **知识库规模**：收录并索引了阿塞拜疆统一立法库中的 **57,000 余部**法律文件。 - **分析粒度**：系统在**句子级（Sentence level）**而非文档级进行深度切片索引，总计索引超过 **1,250 万个法条语句**。 - **动态更新机制**：每月自动整合 **10,000 多份**新增或修订法律文件，确保答案永远基于最新现行法。 - **精度指标**：生成准确率达到 **96.88%**，法条引用准确率达到 **98%**，且在 **1 分钟内**即可生成带有完整引注的法律分析。 - **极低幻觉率**：法条引用幻觉率（Citation Hallucination Rate）仅为 **0.07%**，目前已有超过 **35,000 名公民**使用该系统。 ### [08:18](https://www.youtube.com/watch?v=DNz4wDeffNE&t=498s) 法律意识觉醒引发的法院案件量激增 随着 e-qanun AI 让普通公众能够轻松理解法律并明晰自身合法权益，阿塞拜疆社会的法律参与度显著上升，这直接推动了司法诉讼需求的激增。在过去三年中，全国诉讼案件量增长了 **20% 以上**。 虽然诉讼增长受到宏观经济、司法改革和社会变迁等多重因素综合影响，但这证实了一个必然趋势：当公众法律意识被唤醒、参与门槛降低时，司法系统必须同步提升吞吐能力与效率，以承受快速增长的案件负荷。 - **案件总量攀升**：全国法院受理案件从 2022 年的 **48.5 万件**激增至 2025 年的 **60 多万件**，三年增幅达 **23.8%**。 - **法官编制增长滞后**：同期法官人数从 **554 人**增加到 **672 人**（增长 **21.3%**），增幅明显落后于案件增速。 - **人均负荷极高**：目前阿塞拜疆每位法官年均需要处理 **865 起案件**，司法审判资源处于极度紧张状态。 ### [11:03](https://www.youtube.com/watch?v=DNz4wDeffNE&t=663s) 法官运力缺口与高频案件的效率瓶颈 为厘清司法过载的本质，团队深入分析了司法体系的结构性运力。数据显示，阿塞拜疆目前每 10 万人仅拥有 **6.8 名法官**，而欧洲的法官数量中位数是每 10 万人 **22 名法官**。这意味着阿塞拜疆的人均法官配比比欧洲中位数低了约 3.2 倍。尽管国家持续扩大法官队伍，但仅靠财政增编无法在短期内填补庞大的运力鸿沟。 从案件类型分布来看： - 民事案件每年超过 **242,000 件**； - 行政违法案件每年达 **143,000 件**； - 其余为上诉案件与刑事诉讼。 这些高度标准化、数量庞大的民事和行政案件构成了司法生产力瓶颈的主要来源。法官在进入核心法律逻辑推理之前，被迫将大量时间消耗在信息检索、案卷翻阅和材料整理等琐碎事务上，而这正是 AI 能够发挥巨大价值的切入点。 ### [12:23](https://www.youtube.com/watch?v=DNz4wDeffNE&t=743s) e-Court AI 架构：重塑案卷预审与判决起草 在设计面向法官的辅助系统 e-Court AI 时，团队确立了一条不可逾越的红线：“我们绝不是在建造一个‘AI 法官’（AI Judge）。”司法独立不容妥协——做出判决的是法官，承担最终法律责任的也是法官，AI 的定位纯粹是帮助法官更高效地做出裁判。 e-Court AI 专注于案卷审理前的预处理阶段，其执行流程包含四个阶段： 1. **案卷事实提炼**：自动通读全案卷宗，提取关键实质性事实、诉讼请求及双方当事人立场。 2. **相似先例检索**：智能匹配并调取历史相关裁判先例。 3. **法律法规映射**：将案件事实精准映射至适用的法条与最新修正案。 4. **生成草案框架**：生成逻辑结构完整的判决书草案，供法官全权审查、修改或否决。 | 案件复杂度分类 | 传统人工预审准备耗时 | e-Court AI 辅助后耗时 | 效率提升幅度 | |:--- |:--- |:--- |:--- | | **简单案件（Simple Cases）** | 15 - 30 分钟 | 约 5 分钟 | 3 - 6 倍 | | **普通案件（Average Cases）** | 约 3 小时 | 7 - 10 分钟 | 约 18 - 25 倍 | | **复杂案件（Complex Matters）** | 4 - 5 天 | 10 - 15 分钟 | 数十倍 | 通过这种流程优化，模型测算预计每年可为司法系统节省超过 **300,000 个法官工时**。在不增设物理编制、完全保留法官自由裁量权的前提下，这相当于为全国司法系统直接增加了约 **167 名法官**的有效生产力，使阿塞拜疆的等效司法运力从每 10 万人 6.8 名法官跃升至约 **8.4 名法官**。 ### [15:31](https://www.youtube.com/watch?v=DNz4wDeffNE&t=931s) 模型闭环与未来规划：人机协同的长期演进 为司法系统构建 AI 是一项长期的系统工程。今年初启动以来，团队首先汇集了阿塞拜疆共和国历史上的全量司法判例数据，用于训练理解案件事实与法律逻辑的专用大模型。 在系统的实际运转中，形成了一套严格的人机交互闭环： - 模型阅读案卷并理解事实细节； - 检索历史同类案件的判决规律； - 匹配现国立法知识库； - 输出判决文书草稿交由法官裁决。 - **反馈微调机制**：法官可以采纳、修改或驳回该草案。一旦法官选择否决，系统强制要求法官填写否决理由，这些带有专业法学解释的反馈数据会被沉淀并重新输入训练集，用于模型持续迭代优化。 在未来的推进阶段中，中心计划将案件模拟功能向公众开放。公民若面临潜在法律纠纷，可直接在平台输入案情并上传证据材料，系统将基于现行法与海量判例，为公民生成该纠纷在诉讼程序中的预期走向与建议，从源头上减少非理性诉讼，推动整个法律生态迈向更透明、更可信赖的现代化治理体系。 ## Notable quotes > [00:04](https://www.youtube.com/watch?v=DNz4wDeffNE&t=4s) AI 真的值得信任吗？它能理解复杂的法律规则吗？ — Ulfan Aliyev > [00:04](https://www.youtube.com/watch?v=DNz4wDeffNE&t=4s) AI 在司法系统中应当扮演“支持者”与“赋能者”的角色，而绝非取而代之。 — Ulfan Aliyev > [01:21](https://www.youtube.com/watch?v=DNz4wDeffNE&t=81s) 是简单地采购和消费他国开发的技术，还是建立自主的本土能力？ — Ulfan Aliyev > [03:49](https://www.youtube.com/watch?v=DNz4wDeffNE&t=229s) 既然现行法律法规全部对公众公开，为什么普通人运用法律依然如此艰难？ — Ulfan Aliyev > [03:49](https://www.youtube.com/watch?v=DNz4wDeffNE&t=229s) 不再要求公民学习法律的语言，而是让法律开口说人民的语言。 — Ulfan Aliyev > [05:32](https://www.youtube.com/watch?v=DNz4wDeffNE&t=332s) 在法律领域，“显得很有信心”毫无意义，真正的公信力来自于“精准展示答案为何正确”。 — Ulfan Aliyev > [08:18](https://www.youtube.com/watch?v=DNz4wDeffNE&t=498s) 当公众法律意识被唤醒、参与门槛降低时，司法系统必须同步提升吞吐能力与效率，以承受快速增长的案件负荷。 — Ulfan Aliyev > [12:23](https://www.youtube.com/watch?v=DNz4wDeffNE&t=743s) 我们绝不是在建造一个‘AI 法官’（AI Judge）。 — Ulfan Aliyev ## Mentioned Ulfan Aliyev, National Artificial Intelligence Center of Republic of Azerbaijan, Ai4, AI Academy, Data Race, e-qanun AI, e-Court AI, NVIDIA, Sovereign Supercomputer Center ## Video description Ulfan Aliyev - Chief Commercial Officer, National Artificial Intelligence Center of Republic of Azerbaijan How can artificial intelligence bridge the gap between complex legislation and the public while simultaneously empowering the judiciary to handle immense workloads? To modernize our legal infrastructure, I will share how we deploy integrated AI solutions that transform both citizen access and courtroom operations. The talk will illustrate how our e-qanun AI platform empowers everyday people by translating dense statutory language into simple terms, while our specialized judicial AI acts as a comprehensive partner that reads incoming claims, retrieves precedents, and actively drafts court decisions. I will explain how these combined systems drastically diminish manual effort for judges by evaluating facts and processing cases with advanced precision. Join me to explore how deploying AI across public legal portals and directly accelerates operations and builds a more transparent legal ecosystem for all. [Source](https://www.youtube.com/watch?v=DNz4wDeffNE)

[0:04] Ulfan Aliyev: AI 真的值得信任吗？它能理解复杂的法律规则吗？ [0:04] Ulfan Aliyev: AI 在司法系统中应当扮演“支持者”与“赋能者”的角色，而绝非取而代之。 [1:21] Ulfan Aliyev: 是简单地采购和消费他国开发的技术，还是建立自主的本土能力？ [3:49] Ulfan Aliyev: 既然现行法律法规全部对公众公开，为什么普通人运用法律依然如此艰难？ [3:49] Ulfan Aliyev: 不再要求公民学习法律的语言，而是让法律开口说人民的语言。 [5:32] Ulfan Aliyev: 在法律领域，“显得很有信心”毫无意义，真正的公信力来自于“精准展示答案为何正确”。 [8:18] Ulfan Aliyev: 当公众法律意识被唤醒、参与门槛降低时，司法系统必须同步提升吞吐能力与效率，以承受快速增长的案件负荷。 [12:23] Ulfan Aliyev: 我们绝不是在建造一个‘AI 法官’（AI Judge）。

[0:00] Ulfan Aliyev: How can artificial intelligence bridge the gap between complex legislation and the public while simultaneously empowering the judiciary to handle immense workloads? [0:04] Ulfan Aliyev: Rather than relying on foreign-built AI solutions, Azerbaijan chose to build native capabilities aligned with its national AI strategy. [3:49] Ulfan Aliyev: Simple legal questions often require parsing multiple laws and dozens of articles. [3:49] Ulfan Aliyev: Traditional legal search tools rely on exact keywords or article numbers, whereas citizens ask questions based on everyday scenarios [12:23] Ulfan Aliyev: Reviews the case file to extract material facts, claims, and party positions. [12:23] Ulfan Aliyev: Generates a structured draft decision for the judge to review, modify, or reject. [12:23] Ulfan Aliyev: If a judge rejects an AI-generated draft, they provide a reason for the rejection, which feeds back into the system to retrain and refine future models.

[0:04] I'd like to start with a simple observation.  Around the world, artificial intelligence is [0:09] transforming almost every industry:  healthcare, finance, manufacturing, [0:14] and everybody's asking same question, "How  can AI make our work better?" But when it [0:19] comes to the justice, it's the conversation  immediately becomes much more sensitive. [0:23] People naturally ask, "Can AI really be trusted?  Can it understand complexity of the rule?" [0:30] These are very important questions for us, and  honestly, uh these were the same questions when we [0:35] began to do our journey. Over the past year, our  team in National Artificial Intelligence Center [0:39] has been working with judges, legal experts,  public institutions to explore one idea: Can AI [0:46] make justice more accessible and more efficient  without compromising the trust, transparency, [0:51] and judicial independence? That work led us to  build two AI platforms and one legal system. [0:58] Today, I would like to share why we built them,  what we learned along the way, and why we believe [1:04] AI have support why we believe that AI should be  supportive in the legal system, not replace it. [1:11] Before I talk, I would like to little bit  focus on Azerbaijan. Uh before 1 minute. Sorry. [1:21] Okay. Okay. Okay. Okay. A few years  ago, we asked ourselves same question: [1:26] If AI is going to transform government and  economies, should we simply consume its technology [1:31] technologies developed by in another country,  or should we build it our own capabilities? [1:37] We choose second path. Azerbaijan already adopted  its national AI strategy with a clear objectives. [1:45] Our national artificial intelligence strategy  have four main pillars: AI talent, where we are [1:51] focusing on human development, if we are  investing in human capital; AI governance, [1:56] supporting responsible adoption, transparency, and  stronger public services. Data and infrastructure, [2:02] building the data centers within Azerbaijan,  and creating innovation and investment ecosystem [2:09] within Azerbaijan Azerbaijan, which will allow  public and private sectors collaboration. [2:16] And also, we have our sovereign super center,  which supercomputer center, which has NVIDIA GPUs, [2:22] which allow us to work with sensitive information  citizen information to build AI models. [2:31] I would like to talk a little  bit about our product portfolio. [2:35] We have four main products. First is that AI  Academy. As I said, we are focusing on human [2:40] capital. We are teaching our students from  applied mathematics to the AI engineering. [2:46] Second is that data race platform, where we are  giving experience. We are letting these students [2:52] to work with real public sector challenges. Which  first machine learning based competition platform, [2:59] where government institution private companies  upload their problems there, and our students or [3:06] in the ecosystem players try to solve that  platforms. And we have two interesting other [3:11] products, which is Agoun AI. It is our citizen  facing legal intelligence platform, and E-court [3:16] AI planned to be piloted in this year. Which is  going to support judges for during the case and [3:23] reviewing case materials, identifying precedents,  and preparing structured draft decisions. [3:31] And these four platforms are connected to each  others. For example, the AI Academy develops the [3:37] talent. Data rates data rates platform creates  the practical experience for everybody. And our [3:43] sovereign infrastructure provides the technical  foundation to build Agoun AI and E-court AI. [3:49] When we started the project, we didn't begin with  artificial intelligence. We began to to try to [3:55] understand the main problem. We thought We sat  down with a very simple question. Why do people [4:01] still struggle to use the law even  when the law is publicly available? [4:05] We quickly realized that problem wasn't  about the access to the legislation. We kept [4:13] people read the law, but they doesn't understand  it. And we saw during this process four main [4:18] problems. First one is that people often  receive the document instead of the answers. [4:24] A simple legal question may require reading  several laws, dozens of articles. So, it is a [4:30] very challenging for citizens to understand their  rights. Second is that the language itself is a [4:35] difficult. For example, legislation is naturally  written for lawyers. Which is make it very hard [4:43] to understand for everyday citizens. And third is  that keeping track of amendment is very difficult [4:48] even for professionals because as you know, law  doesn't stay same always. It changes frequently. [4:54] And finally, traditional keyword search finds  keyword search to find the correct document. But [5:00] uh ordinary citizen doesn't know which keywords to  find to search correct to find the correct answer. [5:09] People usually doesn't think in article numbers.  They think in questions. Can my employer do this? [5:14] Do I have this right? What should I do next?  The observation completely changed our thinking. [5:20] Instead of asking citizens to learn the language  of the law, we decided to make the law speak [5:25] to the language of the people. That's this  idea becomes the foundation of Egov AI. [5:32] Uh from the beginning, our goal was never to  build another legal database. There are a lot of [5:37] plenty of databases. Our objective was much  more simpler to make legal information easier [5:42] to understand, navigate, and apply through  artificial intelligence. For the citizen, [5:47] the experience it is a straightforward. Instead  of searching the for article numbers, they [5:53] simply ask a legal question in plain Azerbaijan  language. Behind that simple experience, however, [5:58] platform performs four important tasks.  First is that, as you can see from here, [6:05] first it understands the question through natural  natural language search. Second, it performs [6:10] semantic legal retrieval to understand the  searches by meaning, not by the keywords. Third, [6:15] it generates clear explanations supported by legal  reasoning and applicable legislation. And finally, [6:21] every answer is fully verifiable. Every response  is linked to directly to official legal source. [6:28] This This is one of our main decision during  building this platform. In law, trust doesn't come [6:34] from sounding confidence. Trust comes from showing  exactly why answer is a correct. That's why [6:40] Aigun AI is not a simple search engine.  It's a legal intelligence platform. [6:46] Building that experience required much  more than connecting an AI model and [6:52] It is legal intelligence system.  We want Aigun AI to be measured [6:58] not by promises, but transparent indicators. The  first indicator is scale of the legal knowledge [7:04] that platform can understand. Today, it is  we indexed it more than 57 thousand legal [7:10] documents representing Azerbaijan's consolidated  legislative base. But perhaps more importantly, [7:16] it works at the level of the sentence, not the  documents. That means that our platform already [7:22] indexed it more than 12.5 million sentences.  Because legislation evolves continuously, the [7:29] knowledge base is also updated every month. With  more than 10,000 new or amended legal documents, [7:36] ensuring that responses reflect the latest version  of the law. Law. But scale alone isn't enough. [7:42] What matters is the quality. Today, platform  achieves 96.88% generative accuracy and 98% [7:52] reference accuracy. All a fully cited legal  analytics is produced in less than 1 minute and [7:59] our most uh proud of number is that our citation  hallucination rate is 0.07%. In a legal domain, [8:07] every incorrect citations matters. That's  why our objective has never been to generate [8:12] answers quickly. It's been to generate answers  that are accurate, transparent, and verifiable. [8:18] Okay. Once a Qanun AI was in place, we naturally  started thinking about the broader impact. If [8:26] people understand the law more easily, does that  change how they engage with the justice system? [8:31] It illustrated at the top slide uh  shows all what's happening in the [8:36] legal system in Azerbaijan. To better understand  it, I would like to give you some examples. [8:44] Uh better access better access to the  legislation to the laws leads to the better [8:51] encourage people to be part of the  legal participation system. As a result, [8:55] it leads to the higher number of  the court cases within the country. [9:01] These figures do not provide that AI is only  reason to see that increase. These figures do not [9:07] The court is only influenced by many factors,  including economic condition, legal reforms, [9:11] and broader social developments. But,  they do reinforce one important point. [9:16] As legal participation grows, the justice system  must also become more efficient and more scalable [9:21] to support this growth rate. As you can  see, we all we had more than 20% growth [9:26] within the last 3 years because citizens begin to  understand more easily. They know their rights and [9:33] it leads to the increased number of the  legal cases within the legal ecosystem. [9:39] To answer that questions we started  analyzing judicial workloads. [9:47] As you can see in this slide court cases increased  from 4,080 485,000 to the more than 600,000 [9:58] between 2022 and 2025. That is an increase  of 23.8% in one year in three years actually. [10:07] During the same period of the time the number of  judges also increased from 554 to the 672. That's [10:15] important investment in in the judicial capacity,  but it will represent growth of 21.3% which is [10:21] slightly slower than the increase in the caseload.  As a result each judge is now responsible for an [10:27] average of 865 cases every year. And that is  a huge number. The challenge is that demand is [10:35] growing faster than the capacity. We realize that  increasing number of judges will remain important, [10:41] but over the long term we also need the help of we  also need to increase efficiency of each judges. [10:49] But increasing caseload is only part of the one  picture. We also want to understand whether this [10:56] was simply a workload issue or broader structural  capacity capacity challenge. Okay. I will begin to [11:03] talk again. Okay, as you can see from this slide  we find we find out another interesting insights. [11:09] In Azerbaijan currently have 6.8 judges  per 100,000 people. European median is 22. [11:17] In other words, we have approximately three point  times lower fewer judges than European median. [11:23] At the same time judicial capacity continues to  expand. As shown in the chart number of judges has [11:29] increased steadily. This demonstrates a clear  national commitment to strengthen the judiciary, [11:35] but workforce expansion alone cannot immediately  close the gap. We also examined where judges [11:41] must spend most of their time. Every year  courts process more than 242,000 civil cases [11:49] and administrative offense account for another  100,040 143,000 cases followed by appeal and [11:58] criminal proceedings. Today this high volume  categories represents the greatest opportunity for [12:04] productivity improvement. That's exactly where we  believe AI could create the value, not replacing [12:10] the judge, but reducing the time they spend  searching, reading, organizing the information [12:15] before legal reasoning even begins. That became  the design principle behind the e-court AI. [12:23] How this process works is there's actually Let  me start with one principle that guided every [12:29] decisions we have made. We are not building  the building an AI judge. That was never been [12:35] our object objective. Judicial independence is  non-negotiable. Judge makes a decision decision, [12:41] judge carries the responsibility. AI simply helps  the judge to reach that decision more efficiently. [12:47] We found that greater opportunity in that  present in prepa- in the preparation process. [12:53] First what our AI engine does, first it reviews  the case file, identifies the material facts, [13:01] claims, and position of each party. Second,  it identifies judicial precedent. Third, [13:07] it maps the case to the applicable legislation,  and finally it creates structured draft decision [13:13] for the judge to review, modify, or reject  it. Every one of these steps reduce the time [13:17] spent searching, reading, organizing legal  information, but one principle never changes. [13:23] Every AI generated recommendation is reviewed  by the judge. This isn't a simple product [13:28] future is the principle on which we built  that platform. As you can see from this graph, [13:37] we have categorized legal cases to the three  categories mainly. And we kept asking ourself [13:44] one simple questions. If we reduce the time  spent by judges during the preparation case, [13:53] how is that going to affect overall justice  system? First thing we measured was preparation [13:58] time. Today preparing a simple case may require  15 to 30 minutes. With eCourt AI, that preparation [14:04] can be reduced to around 5 minutes. On average  case may currently require around 3 hours [14:11] and which is decreased to the 7 to 10 minutes.  And more complex matters preparation that may [14:15] take 4 to 5 days can be completed in 10 to 15  minutes before the judge begins the legal review. [14:22] It is important to emphasize that these figures  are related only to the preparation time, [14:27] do not replace judicial reasoning and  decision-making process. The judge continues [14:31] to review the improvements. Judge continues  to review the materials, apply legal judgment, [14:36] and make every final decision. Based on  these productivity numbers, we modeled that [14:41] our potential impact on the systematic level will  be more than 300,000 judicial hours. In practical [14:50] terms, that represents productivity equivalent  to approximately 167 judges while preventing full [14:59] preserving, sorry, full judicial oversight. We  also estimated that effective judicial capacity [15:05] could increase from 6.8 to approximately  8.4 judges per 100,000 people. These are [15:13] modeled estimates based on productivity gains, not  additional judicial appointments. Their purpose is [15:19] not to suggest that AI replace the judges, rather  than demonstrate how AI can help judges judges [15:24] decide more of their time to what matters more  most, legal reasoning and delivering the justice. [15:31] As you can see in outcome case, we plan  to add 167 new judges to our country and [15:39] building AI for justice system is not a one-time  project. It's a long-term journey. As you can see [15:46] in this road map slide, we spent we begin to the  process in the beginning of this year. First, we [15:54] analyze impact we made impact analysis of how we  what's the current problem. As we as I explained [15:59] previously a Kanun AI lead to the increased number  of the cases. First, we understand how many we [16:09] begin with the data collection part. We collected  all of the cases which has happened in Azerbaijan [16:15] Republic. Then, what we did is that we created AI  model which understand and analyze what's going [16:24] what's what should be the next appropriate  action to take. How model is working that [16:30] first, it read the text, understand what's  going on. In next step, it analyze similar cases [16:36] and according to similar cases it already  have solved that I mean what's going to [16:41] be the generative output for this model. Then,  it analyzes the current legislation base which [16:47] is going to be matched with previous decision  and in the final step, it creates sample draft [16:57] document which is going to be reviewed  by the judge. And after that process [17:02] what happens is that judge can accept it amend  it or reject it, but in if judge is not accepting [17:09] the final document, it have to explain why it's  not accepting that document and we're going to [17:14] use that data for retraining purposes. And as  you can see here what we are planning to do [17:20] is that in next phases is that for example if you  have a potential legal issue you can just we will [17:26] open it to the public. You can just write your  case you upload your documents and AI judge will [17:32] give you recommendations that what's going  to be likely outcome of that process. [17:38] Summary of the work done I already talked about  it. Okay and as a result already more than 35,000 [17:46] citizens is using our platform. Our house nation  rate is 0.07% and in good cases we are optimizing [17:55] time spent by judges more than 18 times and we are  adding 167 judges to the or judicial ecosystem. [18:06] Thank you very much. If you have any questions  I would be happy to answer it. [applause]