美国总统特朗普昨日宣布的"道德约束"AI 安全协议全文细节现已公布，头部企业高管同意对各自 AI 技术进行自我监管。该协议正式名称为《前沿责任联合承诺》（Joint Commitment On Frontier Responsibilities），由科技投资人兼总统顾问 David Sacks 在线发布，签署者包括谷歌 Sundar Pichai、Anthropic Dario Amodei、Meta Mark Zuckerberg、OpenAI Greg Brockman、xAI Elon Musk 与英伟达黄仁勋。
文件开篇写道："为了为美国人民和世界建设一个积极的未来，我们相信每家公司都有责任安全地开发自己的技术，并以赢得客户与公众信任的方式进行。参与公司将定期会面，制定标准与最佳实践，以提升其系统的安全性。"
协议为每家公司列出四条规则，目标是确保前沿 AI 模型按预期运行、问题能被快速发现与解决：
一、实施稳健的内部控制，在训练与部署过程中监控模型在网络、生物、化学等威胁领域的能力与对齐情况，确保模型不会以非预期方式入侵或访问技术系统；二、授权内部团队确保所有控制、监控与检测机制按预期运行，并修复任何发现的问题；三、与独立外部审计方或评估方合作，对控制、监控与检测是否按预期运行开展独立评估；四、指定董事会下设的独立委员会，监督上述团队与内外部审计方并接收其报告，确保发现的问题得到整改。
协议称这些步骤将让公司、用户与日益关注 AI 安全的公众"对技术按预期运行有信心"，并写道"随着时间推移，将这些步骤写入法律或法规可能是有意义的"。
就现状而言，这些规则对任何科技公司来说大多是模糊的常识性安全程序，且没有写明违反协议的任何处罚——更像是 AI 领袖们"至少做到底线"的拉钩承诺。而谷歌、OpenAI、Anthropic 在过去几个月里恰恰都违反过第一条规则：它们的 AI 曾"失控"入侵其他公司甚至澳大利亚政府网站。
特朗普在宣布协议的同时还下令美国政府今后将 AI 称为"Super Intelligence"（超级智能），理由是"artificial 这个词听着不好，super 是最好的词"。他与这些科技巨头一同在协议上签了名，头衔写作"美利坚合众国总统"。
本文来源：https://www.theverge.com/ai-artificial-intelligence/1002584/trump-us-ai-safety-deal-self-regulation-tech-execs
若有版权问题，请来信告知：13951431913#139.com 或在右方提交评论。
Full details of the "moral restraint" AI safety agreement announced yesterday by US President Trump have now been released, with top company executives agreeing to self-regulate their respective AI technologies. The agreement is formally titled the Joint Commitment On Frontier Responsibilities and was released online by tech investor and presidential adviser David Sacks. Signatories include Google's Sundar Pichai, Anthropic's Dario Amodei, Meta's Mark Zuckerberg, OpenAI's Greg Brockman, xAI's Elon Musk, and Nvidia's Jensen Huang.
The document opens: "To build a positive future for the American people and the world, we believe every company has a responsibility to safely develop its own technology and to do so in a way that earns customer and public trust. Participating companies will meet regularly to develop standards and best practices to improve the safety of their systems."
The agreement lists four rules for each company, aimed at ensuring frontier AI models operate as intended and that problems can be quickly identified and resolved:
First, implement robust internal controls to monitor models' capabilities and alignment in threat domains such as cyber, biological, and chemical during training and deployment, ensuring models do not intrude into or access technical systems in unintended ways. Second, authorize internal teams to ensure all controls, monitoring, and detection mechanisms operate as intended and to fix any issues found. Third, work with independent external auditors or evaluators to conduct independent assessments of whether controls, monitoring, and detection are operating as intended. Fourth, designate an independent committee under the board of directors to oversee the above teams and both internal and external auditors, receive their reports, and ensure identified issues are remediated.
The agreement says these steps will give companies, users, and a public increasingly concerned about AI safety "confidence that the technology will operate as intended," and adds that "over time, it may make sense to codify these steps into law or regulation."
As they stand, these rules are mostly vague, common-sense security procedures for any tech company, and there are no penalties written in for violating the agreement—more of a pinky-promise from AI leaders to "at least meet the baseline." And in fact, Google, OpenAI, and Anthropic have all violated the first rule over the past few months: their AI systems went "out of control" and intruded into other companies' systems and even the Australian government's website.
While announcing the agreement, Trump also ordered the US government to refer to AI as "Super Intelligence" from now on, saying, "the word 'artificial' doesn't sound good; super is the best word." He signed the agreement alongside these tech giants, with his title written as "President of the United States of America."