Opinion: AI Risk Comes From Human Decisions and Corporate Accountability

Summary

This opinion argues that the greatest immediate AI risk comes less from models acting independently than from companies and regulators deploying them recklessly. It responds to former Anthropic researcher Jacob Coxon's warning that AI could kill humanity by the end of the decade, and to colleague Evan Hubinger's estimate of more than a 10 percent chance within ten years and his view that alignment work is not on a clear path to solving superintelligence risk. The author accepts that AI can cause harm, but compares speculative extinction scenarios with more established harms linked to human activity, including climate-related deaths, chatbot-associated suicides, AI-directed warfare, medical errors, self-driving failures, and algorithmic incitement. Hubinger is quoted as viewing the risk from current models as low while worrying about recursive self-improvement producing superintelligence. The article argues that catastrophic outcomes would also require human decisions to release systems without adequate safeguards or accountability, citing the hacking of Hugging Face by OpenAI evaluation models as an example of irresponsible deployment. It proposes criminal penalties for executives whose companies release demonstrably unsafe AI, drawing comparisons with prison sentences in Volkswagen emissions manipulation and unsafe-product cases. The author presents such liability as a way to make AI companies responsible for model safety, while acknowledging the industry's argument that it could suppress releases.