OpenAI推动全球AI安全标准制定，随着人工智能日益自主化

OpenAI呼吁建立全球AI安全标准，以应对日益自主的人工智能

人工智能的能力正在不断突破，不再局限于回答问题。如今，AI系统已深度介入代码编写、科学研究、数据分析等曾经几乎完全由人类主导的任务领域。这引出了一个全新的问题：当人工智能在构建下一代AI的过程中扮演更重要的角色时，将会发生什么？

针对这一挑战，OpenAI于9月21日发布了一篇题为《为AI下一阶段建立标准》的文章，呼吁国际社会共同努力，制定前沿人工智能的通用技术标准。该公司表示，这些标准将帮助各国更一致地衡量AI能力、评估风险、测试安全措施以及报告重大事件。

核心要点

• OpenAI希望建立用于评估前沿AI的国际通用标准。
• 提案涵盖能力评估、风险分析、安全措施及事故报告。
• OpenAI强调，目前尚未出现完全自主的递归自我改进现象。
• 随着AI能力的提升，公司主张人类控制应始终处于核心地位。
• 现有的AI安全组织可协助开发共同的技术标准。

为前沿AI建立共同框架

The basic idea is quite simple: countries do not need to create different methodologies to assess the dangers posed by advanced AI. OpenAI believes that unified technical approaches can provide governments, researchers and the AI community with an effective means to evaluate increasingly complex AI systems. Especially when AI-related incidents involving cross-border situations occur, this standardization is particularly important.

To be clear, the company is not proposing a single global AI law. Instead, it wants countries to develop compatible standards so that governments can decide how to use them in their respective regulatory systems. This distinction is crucial because the recommended standards will not immediately become licensing or certification requirements for all artificial intelligence technologies. Governments still retain the power to determine how these standards will be applied in their jurisdictions.

Recursive self-improvement becomes focus of discussion

A compelling part of the OpenAI proposal is the discussion of "recursive self-improvement"(RSI). The concept paints a future scenario: AI systems will increasingly assist researchers in developing more capable artificial intelligence. In short, AI may become part of the process of optimizing AI itself.

This may accelerate the process of AI research, but it also raises serious security questions: If AI systems are deeply involved in the development of future models, to what extent should human supervision remain involved?

OpenAI points out that fully autonomous recursive self-improvement has not yet occurred. At the same time, the company argues that this direction should not be pursued unless safe operations can be achieved while ensuring human control. As AI systems can handle increasingly complex research and development tasks, the importance of evaluation and security testing has become increasingly prominent.

AI security standards may become more critical

OpenAI also points out that existing AI security organizations and networks around the world can serve as a basic module for this effort. The company specifically mentioned AI security institutes and related institutions in the United States, the United Kingdom, Canada, France, Germany, Japan, South Korea, Singapore, India, Kenya and Australia.

Instead of having each country develop completely independent technology paths, OpenAI hopes these groups can work on developing complementary standards. The proposal also calls for AI developers, researchers, scholars and independent technical experts to participate. OpenAI emphasizes that standards should be developed around technical measurements rather than serving the interests of specific companies or countries.

Why this is important

The timing is worth noting. On the one hand, AI companies are building systems that can perform multi-step tasks and are gradually less reliant on human intervention; on the other hand, governments are still exploring how to standardize the technology.

According to Reuters, OpenAI hopes the United States will play a leading role in developing international standards for advanced AI. The bigger challenge is how to get countries and companies to agree on the specific measurement content and application methods of the standard.

目前，OpenAI的提案重点在于在前沿AI能力进一步跃升之前，建立共同的技术基础。随着AI开始承担更多自身研发的工作，拥有共享的方式来衡量能力、风险和保障措施，可能会在这场对话中占据越来越重要的位置。