The Legal Distinction Between AI as a Product, a Service, and an Agent in Tort Law

When a defective lawnmower injures a consumer, the legal path is clear: the manufacturer is strictly liable for a defective product, regardless of how carefully they assembled it. But when an artificial intelligence system causes financial or physical harm, that clarity vanishes because an autonomous model is not a static object. The defining legal question of the generative era is whether an AI is a product that was manufactured, a service that is provided, or an independent agent that acts.[3][5]

The distinction dictates the burden of proof. Under traditional tort law, governed largely by the 1998 Restatement (Third) of Torts, if an AI is classified as a "product," developers face strict liability. This means a plaintiff only needs to prove that the AI was defective and caused harm, not that the developer was negligent.

The Brookings Institution notes in its 2019 analysis that applying products liability to AI would force developers to internalize 100 percent of the costs of algorithmic harms, treating software bugs or training data biases as design defects.[3]

However, over the past 50 years of jurisprudence, software has historically been treated as a service rather than a tangible good. If courts classify AI as a service, the legal standard shifts from strict liability to negligence. To win a negligence claim, a plaintiff must prove that the developer breached a standard of care—a notoriously difficult hurdle when dealing with neural networks containing hundreds of billions of parameters, where even the creators cannot fully explain how a model reached a specific output.[2][5]

The emergence of autonomous AI agents complicates this binary. An agent does not just retrieve information; it executes multi-step tasks, interacts with external environments, and makes independent decisions without human oversight. According to a June 2026 framework proposed in arXiv (preprint 2606.00518), treating an autonomous system as a mere product or service fails to capture its dynamic nature.[1]

The proposed "interaction-based framework" suggests that liability should scale with the system's degree of autonomy. If an AI agent acts outside its intended parameters and causes damage—such as executing an unauthorized financial trade or misconfiguring a server—the liability might mirror the legal principles of "vicarious liability" used for human employees, rather than product defects.[1][2]

The stakes are massive. A RAND Corporation analysis of U.S. tort liability warns that large-scale AI damages could overwhelm existing legal structures. If a single foundational model is deployed across thousands of enterprise applications and suffers a cascading failure, strict liability could result in existential financial damages for the model provider. Conversely, a negligence standard might leave thousands of victims without recourse.[2]

Courts have traditionally hesitated to apply product liability to information. Lawfare highlights that while physical items like books or maps have sometimes been treated as products when they contain fatal errors, general information is protected by the First Amendment. If an AI generates a harmful recipe or dangerous medical advice, courts must decide if the model is a defective product under Section 19 of the Restatement, or a protected speaker.[3][4]

As of late 2026, the jurisprudence remains fractured. The New York Law Journal reported in April 2026 that plaintiffs are increasingly testing the waters with product liability claims against AI developers, attempting to bypass the high evidentiary bar of negligence.

While the cited academic and legal frameworks extensively debate the theoretical boundaries of liability, none of the primary sources provide direct quotations from sitting judges ruling on autonomous AI agents, reflecting the nascent stage of this jurisprudence. Without a unified federal statute like the European Union's 2024 AI Liability Directive, the classification of AI systems is being decided piecemeal by state courts across the 50 U.S. states, creating a patchwork of liability standards.[2][5]

The resolution of this legal ambiguity will likely dictate the structure of the AI industry. If strict liability becomes the norm, open-source development could freeze, as creators would be legally responsible for downstream modifications. The next verifiable checkpoint will be how federal appellate courts handle the first wave of agentic tort claims, determining whether century-old product laws can govern systems that think and act on their own.[1][5]