2026年4月人工智能治理格局全景映射更新报告
Mapping the AI Governance Landscape: April 2026 Update Report
麻省理工学院人工智能风险倡议团队与安全和新兴技术中心合作发布了人工智能治理格局映射的更新内容。研究团队改进了基于大型语言模型的处理流程，对人工智能治理与监管档案数据集中的一千多份治理文件进行了分类，这些文件主要为美国联邦层面的英文政府文件。分类工作涵盖六个维度，包含风险领域覆盖范围、受管辖行业、人工智能生命周期阶段、人工智能参与者、立法状态以及系统技术范围。在方法论层面，评估标准从原来的五分制调整为更可靠的三分制，即无覆盖、最低限度覆盖和良好覆盖，并同步更新了模型提示词以减少评分过程中的过度自信问题。
The MIT AI Risk Initiative, in collaboration with the Center for Security and Emerging Technology, has released the updated mapping of the AI governance landscape. The research team improved their large language model-based pipeline to classify over one thousand governance documents from the AI Governance and Regulatory Archive dataset, which predominantly consist of U.S. federal-level government documents in English. The classification covers six taxonomies, including risk domain coverage, sectors governed, AI lifecycle stages, AI actors, legislative status, and system technical scope. Methodologically, the evaluation scale was transitioned from a five-point system to a more reliable three-point scale, consisting of no coverage, minimal coverage, and good coverage, alongside updated prompts to reduce model overconfidence during the scoring process.
研究发现，当前的治理文件高度集中于安全漏洞、隐私保护和透明度等模型安全风险。相比之下，经济贬值和权力集中等社会经济风险，以及多智能体风险和人工智能福利等新兴议题受到的关注较少。在行业覆盖方面，公共管理和科学研发领域占据主导地位，而与日常生活直接相关的面向消费者和劳动密集型行业则代表性不足。这种分布特征表明目前的治理文件优先考虑制度、技术和安全导向的背景，可能在涉及消费者和劳动力相关的行业中存在监管覆盖的空白。
Research findings indicate that current governance documents concentrate heavily on model safety risks such as security vulnerabilities, privacy, and transparency. In contrast, socioeconomic risks like economic devaluation and power centralization, as well as emerging concerns like multi-agent risks and AI welfare, receive comparatively less attention. Regarding sector coverage, public administration and scientific research and development dominate, while consumer-facing and labor-intensive sectors directly tied to everyday life have lower representation. This distribution suggests that current governance documents prioritize institutional, technical, and security-oriented contexts, potentially leaving regulatory coverage gaps in consumer and labor-related sectors.
在生命周期阶段的覆盖面上，多数文件同时涉及多个环节，但部署、运行和监控等下游阶段的受关注度显著高于早期的数据收集和处理阶段。在技术范畴内，治理框架倾向于使用人工智能系统和人工智能模型等宽泛术语进行系统级界定，对前沿人工智能、基础模型或开源系统等特定技术系统的针对性关注有限。这种普遍化的监管表述可能会限制针对具有特定风险特征的系统进行有效治理的能力。
In terms of lifecycle stage coverage, most documents address multiple stages simultaneously, yet downstream stages such as deployment, operation, and monitoring receive significantly greater attention than early-stage data collection and processing practices. Within the technical scope, governance frameworks tend to use broad terms like AI systems and AI models for system-level definitions, with limited targeted attention given to specific technical systems such as frontier AI, foundation models, or open-weight systems. This generalized regulatory phrasing may limit the effectiveness of governance for systems with distinct risk profiles.
在参与者角色分配方面，正式的监督和执行职能通常由治理机构承担，而遵守合规义务的要求则广泛指向人工智能的开发者和部署者。就立法状态而言，数据集中绝大多数文件被归类为具有法律约束力的硬法。然而在这些硬法中，仅有百分之四十四处于生效状态，百分之四十三已失效，另有百分之十二仍处于提案阶段，反映出该领域具有显著的立法更迭特征。
Regarding the allocation of actor roles, formal oversight and enforcement functions are typically assigned to governance bodies, while compliance obligations are broadly directed at AI developers and deployers. In terms of legislative status, the vast majority of documents in the dataset are classified as legally binding hard law. However, among these hard law documents, only forty-four percent are currently enacted, with forty-three percent defunct and twelve percent in the proposal stage, reflecting significant legislative churn in this field.
该研究同样客观记录了现存的局限性。数据集存在司法管辖区不平衡与语言限制，其结果主要反映美国的治理趋势。同时，大型语言模型在文本分类中也表现出一定局限，例如难以准确区分部署与运行监控等相邻的生命周期阶段，且容易扩大阶段的适用范围。在处理风险管理等宽泛词汇时，模型容易过度判定其涵盖了治理失败领域。此外，行业的准确分类高度依赖于文件中是否出现显性的行业术语。研究团队正通过迭代提示词和引入更多人工校准来应对这些问题。
The research also objectively documents existing limitations. The dataset exhibits jurisdictional imbalance and language constraints, with results primarily reflecting U.S. governance trends. Concurrently, large language models show certain limitations in text classification, such as difficulties in accurately distinguishing between adjacent lifecycle stages like deployment versus operation and monitoring, often over-applying the scope of these stages. When processing broad terms like risk management, the model tends to over-attribute coverage to the governance failure domain. Furthermore, accurate sector classification relies heavily on the presence of explicit sector terminology within the documents. The research team is continuously addressing these issues through prompt iterations and the introduction of additional human calibration.
全面介绍以上内容后，欢迎大家关注Learn By Doing With Steven 数能生智、steven data talk以及steven数据漫谈等频道内容。相关节目已在小红书、微信公众号、YouTube、Spotify等多个平台同步更新。在节目描述或shownote区域包含我的linktree，您可以借此访问所有的社交媒体平台链接。
Having comprehensively introduced all contents, we invite the audience to check out the Learn By Doing With Steven YouTube channel, along with the steven data talk and steven data talk related podcasts. Relevant content is updated across multiple platforms including Xiaohongshu, WeChat Official Account, YouTube, and Spotify. The program description or shownote area contains my linktree, through which you can access all social media platform links.
https://cdn.prod.website-files.com/669550d38372f33552d2516e/69d7c0c3f27cbb905d95df0e_April%202026%20Update%20-%20Mapping%20the%20AI%20Governance%20Landscape.pdf
#人工智能治理 #监管政策 #风险管理 #机器学习 #数据分析 #AIGovernance #RegulatoryPolicy #RiskManagement #MachineLearning #DataAnalysis
all my links: https://linktr.ee/learnbydoingwithsteven
