Catastrophic AI Liability

Catastrophic AI liability is the accountability problem raised when AI systems are alleged to contribute to severe harms such as cyberattacks on critical infrastructure or bioweapon creation. Bytes: Week in Review - AI companies divided over proposed state law, Amazon buys Globalstar, and Spotify to sell physical books adds the concept through [[MariaCurie|Maria Currie]]’s critique of treating AI liability like ordinary social media platform liability.

The concept matters because model developers, users, and lawmakers may all point to different causal layers. A user can initiate misuse, but a model provider may shape model capability, access, safeguards, monitoring, and post-release response. A state law that limits liability therefore has to decide how much responsibility remains with the developer when the harm is unusually severe.

Key Claims

• Catastrophic harms make AI liability different from ordinary content-hosting disputes.
• Intentional or reckless developer conduct is only one possible liability threshold; negligence, safety protocol adequacy, access control, and monitoring may also matter.
• Safety protocols can reduce risk, but they need grounding in real model behavior and plausible misuse pathways.
• State AI liability shields may be politically unstable when constituents demand regulation of severe AI harms.
• The concept connects model-release governance with civil liability, not only internal safety policy.

Connections

• Illinois AI Liability Bill and State AI Liability Shield - source law and legal mechanism.
• OpenAI, Anthropic, and AI Governance And Compliance - companies and governance frame.
• Frontier Model Release Governance, Frontier Model Access Restrictions, and AI Cyber-Defense Utility - adjacent high-capability model governance concepts.
• Internet Liability Spillover - platform-law analogy the source narrows.