There Is a Solution to AI’s Existential Risk Problem
Image credit: Getty Images
The existential risk posed by artificial intelligence (AI) is a pressing concern that requires a novel solution. The current approach of relying on AI alignment, which aims to make superhuman AI act according to humanity's values, is problematic and may not be effective. This perceived difficulty is attributed to the complexity of the problem, the slow scientific progress in alignment, and the philosophical question of which values to align a superintelligence to.
In contrast, a proposed Conditional AI Safety Treaty, where signatory countries agree to halt potentially unsafe AI training once they reach a certain threshold of capabilities, could be enforced by AI Safety Institutes, which would evaluate AI models and determine when they pose a risk of loss of control. The proposed solution would not obstruct current AI development, as most current AI systems are far below the loss of control threshold, the treaty would allow for the development of narrow AIs and general AIs that are deemed safe, provided that they meet specific safety criteria.
Although the Safety Treaty would require international cooperation and enforcement, which would be a major geopolitical challenge. However, this solution is necessary to mitigate the existential risk of AI and prevent potential risk to humankind. The proposed solution has been endorsed by prominent experts in the field, including Nobel Laureate Geoffrey Hinton and Turing Award winner Yoshua Bengio.
