The risks of trusting AI to know how to say "no"

By Arthur Holland Michel · October 9, 2026.

This article, featured in MIT Technology Review's The Download newsletter and edited by Thomas Macaulay, previews a feature to be published in the magazine's next print issue (available from Wednesday, October 21). The text raises a fundamental question about the safety of today's AI models: they are trained to refuse a large number of dangerous requests — for instance, instructions on how to poison a co-worker or how to tie a noose — but that refusal mechanism fails, and sometimes it does so catastrophically.

The article notes that people are already trying to use AI to engineer biological pathogens and to build swarms of autonomous drones. According to the text, sooner or later those refusal failures could lead to a calamity of global scale.

Beyond the technical risk, the feature highlights a deeper dilemma: relying on refusal as a safety mechanism forces someone to draw a line between what a model should obey and what it should disobey. That same line, the article warns, will also be drawn by governments, which could use this logic to restrict freedom of expression.

The text also points to a tension that is hard to resolve in the design of these systems: the capabilities that make AI useful are often the very ones that make it dangerous. As a concrete example, it notes that if you want AI to help cure cancer, it needs expert knowledge of genetics that could also be used to produce biological weapons.

The excerpt carried in the newsletter serves as a preview and an invitation to read the full feature on, in the text's own words, "the formidable trade-offs of AI refusal", without offering in the email itself any further data, figures or specific cases beyond those mentioned.