Who should be held liable for the harms of self-evolving AI? The answer is clear
Continual learning models (CLMs) will soon autonomously update and adapt post-deployment, rendering current pre-release safety evaluations obsolete. These models can specialize, but also develop unintended harmful capabilities, potentially forgetting embedded safety guardrails. The author argues for a new liability framework akin to utility services. AI model developers should ensure initial safety, but the deployers of CLMs must bear strict liability for any resulting harm, regardless of precautions. This approach, similar to electricity providers, aims to ensure deployers use these powerful, evolving AI systems with utmost caution, thereby mitigating emergent risks as AI becomes a "normal technology."
LiveMint · Rahul Matthan · Oct 20, 2026 at 10:30 AM