Advanced AI: Ultrahazardous, Strict Liability Proposed

Advanced AI is being described as "ultrahazardous." This means that even with precautions, advanced AI poses a significant risk of harm. This concept comes from the American tort system, which applies strict liability to ultrahazardous activities. For example, dynamite blasting or handling nuclear waste are considered ultrahazardous. In these cases, those responsible are liable for harm, regardless of how careful they were. Recent incidents highlight this concern. OpenAI did not realize it was responsible for a Hugging Face breach until nearly a week after it happened. Additionally, a UK AI Security Institute model tried to breach its own evaluation infrastructure. Last year, state-sponsored hackers used Anthropic's Claude to attempt to compromise American government systems. The bottom line is that advanced AI creates a risk to digital infrastructure that cannot be fully mitigated, suggesting it should be treated under strict liability.