本文提出模型访问治理议题,分析访问权限对 AI 下游收益与风险的影响及现有研究不足,并面向评估机构、企业、政府和国际组织提出四类建议。
摘要:
AI 系统的下游用例、收益和风险在很大程度上取决于系统所提供的访问权限,以及这些权限面向谁。然而,人们对不同访问方式的下游影响仍缺乏充分理解,导致决策者难以负责任地治理模型访问。因此,我们重点介绍 Model Access Governance(模型访问治理)这一新兴领域,其重点是帮助组织和政府作出负责任、基于证据的访问决策。我们通过强调模型访问治理失当的风险、现有相关研究的局限以及产生影响的机会,概述了发展该领域的动机。随后,我们提出四组建议,旨在帮助 AI 评估组织、前沿 AI 公司、政府和国际机构围绕经验驱动的访问治理建立共识。
英文摘要:
The downstream use cases, benefits, and risks of AI systems depend significantly on the access afforded to the system, and to whom. However, the downstream implications of different access styles are not well understood, making it difficult for decision-makers to govern model access responsibly. Consequently, we spotlight Model Access Governance, an emerging field focused on helping organisations and governments make responsible, evidence-based access decisions. We outline the motivation for developing this field by highlighting the risks of misgoverning model access, the limitations of existing research on the topic, and the opportunity for impact. We then make four sets of recommendations, aimed at helping AI evaluation organisations, frontier AI companies, governments and international bodies build consensus around empirically-driven access governance.
Secure SQL Server - Enabling Secure Access to Remote Relational Data
安全SQL Server - SecSS是一种主要开发用于实现状态自助治理的技术,如(Paulin 2012)中所述。自助式治理是一种新型的治理模式,它拒绝基于服务的公共行政,而是建议被治理主体以自助的方式管理他们的法律关系,基于临时确定的杀戮。在这篇文章中,我们描述了原型SecSS和它的评估在一个复杂的政府场景。
Implications for Governance in Public Perceptions of Societal-scale AI Risks
Wichita State University(威奇托州立大学);Transformative Futures Institute(变革未来研究所);Centre for the Study of Existential Risk, University of Cambridge(剑桥大学生存风险研究中心);University of Oxford(牛津大学);University College London(伦敦大学学院);University of Memphis(孟菲斯大学);Arkose(Arkose公司);Rethink Priorities(反思优先事项研究所);Center for the Future Mind, Florida Atlantic University(佛罗里达大西洋大学未来心智中心)
A governance horizon for ethical-use constraints in open-weight AI models
School of Computer Science(计算机科学学院);Ministry of Education(教育部);Laboratory of High Confidence Software Technologies(高可信软件技术实验室);University of Science and Technology Beijing(北京科技大学);University of California, Davis(加州大学戴维斯分校)
通过审计Hugging Face Hub上的模型仓库,发现基于披露的治理在开放权重AI中具有浅层结构性限制,提出治理视野概念并比较不同政策设计的效果。
EvalCards: A Framework for Standardized Evaluation Reporting
University of Copenhagen(哥本哈根大学);ETH Zurich(苏黎世联邦理工学院);University of Amsterdam(阿姆斯特丹大学);University of Barcelona(巴塞罗那大学);Johannes Kepler University Linz(林茨约翰尼斯·开普勒大学)
Position: AI Governance Needs ISO-like Interoperability Protocols, Not Just Laws
Computational Intelligence and Operations Laboratory (CIOL)(计算智能与运筹实验室(CIOL));Shahjalal University of Science and Technology(沙贾拉尔科技大学);Independent University(独立大学);Hanyang University(汉阳大学);University of Oklahoma(俄克拉荷马大学)