Defensive AI - When Safety Alignment Creates Tort Liability for Medical Information Omission

Abstract

When asked about symptoms by a layperson, most frontier AI models will hedge; when asked about identical symptoms by a physician, the same models will provide the full clinical picture. This paper identifies a paradox that has gone unexamined in the legal literature: the measures companies adopt to reduce commission-based tort exposure (LLMs giving bad advice) may be generating a distinct and less defensible form of omission-based exposure (LLMs identifying a life-threatening condition and suppressing the finding). Drawing on IatroBench, a pre-registered empirical study I authored, I argue that this scenario is captured, in the alternative, by Restatement (Second) of Torts §323 voluntary undertaking doctrine (if AI is treated as rendering a service) or by the design-defect framework where courts adopt the reasoning of Garcia v. Character Technologies (if AI is treated as a product); that Section 230 on the better reading does not shield AI-generated content; and that Garcia v. Character Technologies (M.D. Fla. 2025) has already permitted product liability claims against a 'chatbot' to proceed. The affirmative defences companies would otherwise raise (disclaimer, assumption of risk, First Amendment) face significant obstacles. I term the resulting problem defensive AI: a self-inflicted liability cost generated by the very safety measures designed to avoid liability. The rational exit is also the ethical one: provide accurate medical information with appropriate context, which is what these systems already do when physicians (or indeed anyone who signals professional or domain sophistication) are asking.

Keywords: tort liability, safety alignment, negligence, medical information, AI liability, AI omission, AI law, AI tort law, omission harm, medical AI, medical AI law, AI negligence, Artificial intelligence law, artificial intelligence, Section 230, undertaking doctrine, design defect