Mark Geistfeld (New York University School of Law) has posted When Information Functions As A Product: Implied Representations And The Tort Divide Between Products And Services For AI Systems on SSRN. Here is the abstract:
With the rapid development and deployment of artificial intelligence (AI) systems and the ensuing onset of cases seeking recovery for AI-caused harms, courts must resolve the threshold tort question whether these systems constitute “products” subject to the defect-based rules of strict liability, or instead are “services” governed by negligence. Courts make this classification by determining whether the intangible informational outputs of AI systems function analogously to tangible products. The approach relies on established common-law methodology, but it lacks a foundation. Properly classifying AI outputs requires two accounts that neither the cases nor the commentary has developed—one explaining why tangibility bears on the choice between negligence and strict products liability, the other identifying which of an AI system’s informational outputs are relevantly analogous to products. Without them, an AI companion alleged to have drawn a teenager into a fatal emotional dependency is indistinguishable from an AI system that writes short stories, since both output only words. Either both fall outside strict products liability or both are subject to it.
This Article supplies the missing accounts. An analysis of paradigmatic cases involving tangible tools and intangible human labor identifies tangibility as a proxy for two substantive criteria: whether an item’s safety can be evaluated by reference to its condition or performance without examining the reasonableness of the provider’s conduct, and whether the item can be sequentially possessed and distributed through a stream of commerce. Commercial distribution of AI systems ordinarily satisfies the second criterion. Because every AI output is information, the first criterion requires a theory of information. That theory distinguishes judgment-based content, whose quality cannot be assessed apart from the provider’s exercise of discretion, from representational content, which mirrors existing fact and is either true or false. A commercial provider of representational content impliedly represents that it is true; a product seller, that the product will perform reasonably safely. Either representation, whether of existing fact or of expected safety performance, can be evaluated without examining the provider’s conduct.
An AI output accordingly functions analogously to a tangible product when its safety performance rests on an implied representation that can be evaluated independently of the provider’s conduct, and functions analogously to a service when it rests on no such representation. An AI companion marketed to adolescents carries the implied representation that its interactions with them will be reasonably safe. Whether its outputs, taken together, satisfy that representation can be determined by examining them rather than the developer’s conduct, so those outputs function like a product. The content of AI-written fiction carries no implied representation that it is true and functions like a service.
Beyond the choice of liability rule, the consequences long attached to the classification—charging the distributor with expert knowledge and extending liability through the chain of distribution—belong to the commercial enterprise and therefore apply across all outputs of a general-purpose AI model, however each is classified. So understood, the product/service divide does not shift the underlying basis of responsibility for commercial actors. It selects the form in which tort law enforces a single duty to place reasonably safe items into the stream of commerce.
Highly Recommended!
To receive new posts from Legal Theory Blog by email, get a free subscription to Legal Theory Stack.
Lawrence Solum