Trent Kannegieter (Yale U Law) has posted “Nondeterministic Torts: A Mechanistic Approach to Large Language Model Tort Liability” on SSRN. Here is the abstract:
Our laws were built for deterministic machines. When given the same inputs, we expect a system to consistently produce the same outputs. But modern artificial intelligence (AI) systems, specifically large language models (LLMs) and other types of “generative AI” (GenAI), challenge this assumption. These systems are nondeterministic, meaning that they produce varied outputs even when given identical inputs. By their nature, nondeterministic LLM-based predictions carry an arbitrary randomness that is an inherent feature, not a bug, of the product.
Even as AI applications spread rapidly, liability for AI systems going wrong remains an open question. But despite the rise of attention and scholarship around artificial intelligence since the “GenAI explosion” sparked by ChatGPT’s release in fall 2022, researchers have so far overlooked nondeterminism’s profound consequences for the law of AI.
This article argues that nondeterminism is the key link to successfully arguing a host of tort claims against the creators and deployers of AI products when these products cause injury. The inherent dangers of deploying a nondeterministic, LLM-based system give rise to multiple potential tort claims, especially when AI systems are deployed in contexts that aresafety-critical (where the cost of individual errors is comparatively high) oragentic (when AI agents have comparatively few checks from humans or deterministic software). Even if the average legal reader isn’t aware of nondeterminism, the average AI engineer is. By deploying a system that is known to be nondeterministic and unpredictable, developers might be accepting responsibility for the harms that emerge from a model’s behavior.
Nondeterminism is the critical hook into multiple popular tort doctrines, especially negligence and product liability (design defect) claims. In claims of negligence, developer knowledge of nondeterminism helps establish a duty and makes even nominally unexpected harms foreseeable, as they are downstream from an unpredictable, nondeterministic system. In product liability, nondeterministic systems deployed in safety-critical or agentic contexts might be a defective design for their use case. Quality assurance (QA) procedures, critical to ensuring the safety of mission-critical systems, are inherently incomplete for nondeterministic systems. Nondeterminism might even make deploying an LLM-based system in such contexts an abnormally dangerous activity, further strengthening the case for a strict liability regime.
This piece arrives at a critical moment in the development of common law around AI. AI might be eating the world, but regulation has not kept pace. Little federal legislative movement is expected, leaving soft law (industry self-regulation) as a stopgap. But the rewards for AI companies deploying quickly in an era of skyrocketing valuations will likely prove too large to ignore. Without a clear liability regime, firms see little costs to counter the lucrative benefits of capturing the AI market. Absent additional regulation, the common law as it stands today is the best tool in today’s toolkit to guard against AI harms.
The article offers a novel, mechanistic approach to assessing AI liability in high-stakes domains: beginning with technical and use case specificity. By looking at the technical features of products on the market today, we can discover ways to use existing doctrine to regulate new technologies. Inquiries into the law of AI must begin with how AI systems work. This framework could drive meaningful reforms, helping strengthen the effort of using tort law to guard against AI harms without having to rely on speculative future harms. If utilized, the arguments within this piece might incentivize AI product developers—the least cost avoiders—to only deploy agentic GenAI systems when the benefits exceed the costs.