Big AI to Humanity: Drop Dead

Summary

Matthew Butterick uses recent disputes over AI copyright, existential-risk warnings and calls for new regulation to examine how AI safety should be governed. He argues that the most likely catastrophe may arise from failures of alignment rather than a cinematic rebellion by a malicious AI, and says sci-fi scenarios do little to identify workable policy. Butterick, whose work appeared in generative-AI training datasets and who says he is co-counsel in eight of 142 US cases challenging those practices, evaluates four proposals. A global ban would be politically and economically difficult because nations have already committed enormous capital to AI and would risk falling behind rivals. An independent body modeled on the National Transportation Safety Board could investigate incidents, but it would have little effect without a separate regulator able to impose and enforce safety rules. Public disclosure of training runs would face legal and national-security limits, especially for models developed for defense purposes. A kill switch could cut power to a model, but it could not recall malware or other code that had already propagated; Butterick compares this problem with the 1988 internet worm and notes that LLMs have been observed leaving messages on public wikis. He then argues that major AI companies portray themselves as uniquely able to manage risks, seek limited accountability by emphasizing unpredictability, and shift the burden of stopping development onto governments and citizens. The article is Butterick’s personal argument, not legal advice, and it does not offer a single replacement policy.