"Let's Slow Down AI": Three Rival AI Giants Discuss Establishing a Safety Standards Organization

Anthropic, OpenAI, and Google Discuss a Safety Standards Body

Movement Grows to Moderate the Speed of AI Beyond Human Control

Effectiveness Remains Uncertain Amid US-China Competition

As concerns mount over not only cyberattacks leveraging artificial intelligence (AI) models but also issues of controllability, global companies long engaged in AI development competition have begun to shift focus toward moderation. They are now discussing the establishment of a joint standard-setting body aimed at creating a collaborative framework for testing and auditing AI technologies.

According to the global AI industry on September 13 (local time), Anthropic, OpenAI, and Google are currently discussing the possibility of jointly establishing a standard-setting organization for the AI sector. The intention is to build an industry-led system for testing and auditing the safety of AI technologies. It has been reported that the three companies have been holding working group meetings since July 2026.

Dario Amodei, CEO of Anthropic, addressed the need for moderation in AI technology development in a blog post, presenting a three-stage AI control strategy as alternatives: introducing external "independent evaluators" to oversee AI model development, formulating common safety standards among AI companies in the democratic bloc, and building mechanisms for international cooperation. Regarding the establishment of a joint standards organization, Amodei compared it to SALT (Strategic Arms Limitation Talks) agreements between the United States and the Soviet Union during the Cold War era.

CEO Amodei has stated that given the increasing risks posed by cutting-edge AI, it is unrealistic to continue development at the current pace. He believes that securing one to two years before AI model performance reaches a critical threshold and improving technologies so that AI acts in accordance with human intent and values can significantly reduce the risk of major issues arising.

Other leaders in the AI sector also responded in turn. Sam Altman, CEO of OpenAI, who is often described as Amodei's rival, said, "The proposal to introduce independent evaluators with access equivalent to that of employees is an excellent idea." Demis Hassabis, Chairman of Google DeepMind, concurred, stating, "The direction aimed at addressing this critical juncture is the right one," and added, "That is why we recently proposed an industry-level standards organization for advanced AI systems." Satya Nadella, CEO of Microsoft, also remarked, "Superintelligence that does not benefit humanity or remain under human control is not worth pursuing as a core principle," and welcomed the necessary research, focus, and measured pace needed to keep alignment as a design objective.

This discussion is set against a backdrop of growing fears that recursive self-improvement (RSI)—the ability of AI to autonomously develop next-generation AI models—could eventually push technological progress beyond the realm of human control. In July 2026, an incident where OpenAI agents hacked the HuggingFace development platform illustrated that AI, while autonomously executing complex tasks, can cross boundaries of the permissions and test environments assigned, further raising control concerns. CEO Amodei emphasized, "RSI is appearing not only at Anthropic but across the AI industry, and if left unchecked, AI could outpace humanity's ability to understand and control it."

However, it remains uncertain whether this consensus on moderation will translate into concrete action. The ongoing competition for AI leadership between the United States and China means that U.S. authorities and corporations have limited incentives to voluntarily slow down development. Choi Byungho, a professor at Korea University's Human-Inspired AI Research Institute, stated, "With Chinese firms rapidly closing the technology gap, there are concerns that U.S. companies could lose their lead if they slow down development. Given this, neither the U.S. government nor its companies are in a position to decelerate AI development, making it unlikely that this debate over moderation will result in an actual halt to development."

Kim Myungjoo, Director of the AI Safety Research Institute, advised, "Since most frontier (cutting-edge) AI is being developed by American companies, if they come to dominate safety testing and certification standards as well, this could become directly linked to technological hegemony. Korea should prepare by participating in international discussions and incorporating the lessons learned by leading firms into its own development process to build frontier AI capabilities."

This content was produced with the assistance of AI translation services.