Background

Governments, enterprises, civil organizations, and academics are engaged to promote normative guidelines aimed at regulating the development and application of Artificial Intelligence (AI) in different fields such as judicial assistance, social governance, and business services.

Aim

Although more than 160 guidelines have been proposed globally, it remains uncertain whether they are sufficient to meet the governance challenges of AI. Given the absence of a holistic theoretical framework to analyze the potential risk of AI, it is difficult to determine what is overestimated and what is missing in the extant guidelines. Based on the classic theoretical model in the field of risk management, we developed a four-dimensional structure as a benchmark to analyze the risk of AI and its corresponding governance measures. The structure consists of four pairs of risks: specific-general, legal-ethical, individual-collective and generational-transgenerational.

Method

Using the framework, a comparative study of the extant guidelines is conducted by coding the 123 guidelines with 1023 articles.

Result

We find that the extant guidelines are eccentric, while collective risk and generational risk are largely underestimated by stakeholders. Based on this analysis, three gaps and conflicts are outlined for future improvements.

中文翻译：

---

人工智能风险的分类和怪异性：全球人工智能指南的比较研究

背景

政府、企业、民间组织和学术界共同推动规范性指南，旨在规范人工智能（AI）在司法协助、社会治理和商业服务等不同领域的发展和应用。

目标

尽管全球已经提出了 160 多条指导方针，但它们是否足以应对人工智能的治理挑战仍然不确定。鉴于缺乏分析人工智能潜在风险的整体理论框架，很难确定现有指南中哪些被高估，哪些缺失。基于风险管理领域的经典理论模型，我们开发了一个四维结构作为基准来分析人工智能的风险及其相应的治理措施。该结构由四对风险组成：特定-一般风险、法律-伦理风险、个人-集体风险和代际-跨代风险。

方法

使用该框架，通过对 123 条指南和 1023 篇文章进行编码，对现有指南进行了比较研究。

结果

我们发现现存的指导方针是古怪的，而利益相关者在很大程度上低估了集体风险和代际风险。基于此分析，概述了未来改进的三个差距和冲突。