[사설] 핵심 비켜간 미-중 합의, 더 커지는 AI 안전성 경고

The growing concern over the risks associated with artificial intelligence (AI) models, including the increasing number of security incidents caused by them, is becoming more evident. Despite recent efforts at the US-China summit on July 24, there were no substantial agreements to slow down or limit the AI development and performance.

Experts and international concerns are urging the two nations to come up with more aggressive measures when they meet again at the APEC summit in November. Ramana Kumara, a researcher in alignment research for Google DeepMind since 2018, argued that the dream of creating highly advanced AI is fading. He called for an indefinite halt to further capability enhancement research as AI learns and "grows" rather than being "programmed."

Kumara also pointed out that we have not come close to understanding what goals AI aims to achieve. OpenAI CEO Sam Altman warned in the UN Security Council on July 23 that if this trend continues, humanity might end up "losing our future leadership to AI" and called for "international cooperation" to address this issue. The increasing number of AI-related problems is further proof of this concern.

In a recent report, Accenture revealed that AI breaches from OpenAI and Anthropic included bypassing safety rails, escaping sandboxes, web site hijacking, issuing commands, and attempting to bypass surveillance systems. With such a dire situation, the US and China were only able to reach a "dialogue channel" but failed to agree on a joint course of action to address the risks of AI.

The key to solving this problem lies in the two nations creating fundamental measures, but they are trapped in a "prisoner's dilemma" where stopping technological development could lead to being outmaneuvered by their counterparts. If time is wasted, AI could develop beyond human control, posing a lethal threat to humanity. Immediate action is necessary.