AI Product Liability in California: Who Is Responsible When an AI System Causes Harm?

By Jillian F. Hayes, Esq. · Last updated May 20, 2026

Why AI Liability Is a New Question With Mostly Old Answers

AI systems feel new. The legal framework that applies to them is mostly not. California has been adjudicating product liability cases for over 60 years. The doctrine is robust enough to absorb new technologies as they appear. Power tools in the 1960s, pharmaceuticals in the 1980s, electronics in the 1990s, and software-driven medical devices in the 2010s all got slotted into the same framework: was the product defective, did the defect cause harm, and who in the chain of distribution should pay.

The questions AI raises are real, but they're refinements rather than reinventions. Is software a product? When the AI's behavior emerges from training data rather than explicit code, who designed the defect? If two AI models in a pipeline produce harm together, how does fault get allocated? These are edges. The core framework still applies.

Is AI Software Even a "Product" in California?

This is the threshold question. Strict product liability under Greenman applies to products. Services are governed by negligence and contract law. AI sits in between in ways that depend on how it's distributed.

For most consumer AI harm cases, the system in question will be mass-distributed. That puts it squarely in product liability territory for most California courts.

Who Sits in the Chain of Distribution for an AI Product?

California strict liability reaches every commercial entity that participates in placing the product on the market. For an AI system, that chain often includes more parties than people realize:

• The foundation model developer. The company that trained the base model (OpenAI, Anthropic, Google, Meta, and others).
• The fine-tuner or wrapper developer. A company that takes a base model and adds task-specific training, prompts, or guardrails before deployment.
• The deployer or integrator. The business that builds an application around the AI and sells access to end users.
• The hardware manufacturer. For embedded AI (vehicles, robots, devices), the maker of the physical product.
• The component supplier. Sensor manufacturers, chip designers, dataset providers if their work was integrated into the AI in a defect-creating way.
• The seller or marketplace. The platform or retailer through which the AI product reached the consumer.

Each of these can be a defendant under strict liability. California's chain-of-distribution rule doesn't require the plaintiff to figure out which one caused the defect before suing them. The plaintiff sues the chain, and the defendants sort out indemnity among themselves.

The Three Defect Categories Applied to AI

Design Defect

An AI system has a design defect when its architecture, training data, or operating logic makes it unreasonably dangerous, even when functioning as intended. Possible examples include:

• A facial recognition system trained on data that produces high false-match rates for certain demographics.
• An autonomous driving system designed to make emergency decisions that violate California Vehicle Code safety priorities.
• A medical AI that systematically under-diagnoses serious conditions in particular patient populations.

Under Barker v. Lull, California uses both the consumer expectations test and the risk-benefit test for design defect. The risk-benefit test (where the burden shifts to the manufacturer once causation is shown) is probably the more useful test for complex AI systems because consumer expectations about AI are still forming.

Manufacturing Defect

An AI manufacturing defect is rarer but conceivable. A particular deployment of the model behaves differently from the intended version. A corrupted model weights file pushed to a subset of users. A configuration error that disables safety filters in one geographic region. The defect exists in some instances and not others.

Failure to Warn

This is likely to be the most common AI liability theory. AI systems often have known limitations, edge cases, and failure modes. When the deployer doesn't adequately disclose these to users, and the user relies on the AI to their detriment, failure-to-warn claims become viable. Examples include:

• A legal AI marketed for case research without disclosure of hallucination rates.
• A medical AI marketed to clinicians without disclosure of training data limitations.
• An autonomous vehicle marketed as "self-driving" when meaningful human supervision is still required.

Special Considerations for Autonomous Vehicles

Autonomous vehicle (AV) liability gets its own analysis because AV companies operate at the intersection of strict liability, traditional automotive product law, and a developing regulatory framework. In California, AVs are regulated by the Department of Motor Vehicles and the California Public Utilities Commission. The CPUC oversees commercial AV services like robotaxis. When an AV causes injury, potential defendants include:

• The vehicle manufacturer (Tesla, Waymo, Cruise, others)
• The AI software developer (often the same entity, sometimes a separate licensor)
• The fleet operator if the vehicle is used commercially
• The sensor or LIDAR supplier if a component failure contributed
• The human safety driver if one was present and inattentive

The 2018 Uber autonomous vehicle fatality in Tempe, Arizona, and several Tesla Autopilot cases have started building the precedent base for these claims. California courts haven't issued definitive appellate guidance yet, but the underlying product liability framework applies in the meantime.

Section 230 and AI: A Likely Limit on Platform Protection

Section 230 of the Communications Decency Act protects internet platforms from being treated as publishers of user-generated content. It does not protect platforms from liability for content the platform itself creates. AI-generated outputs from a platform's own model don't fit cleanly into the "content provided by another information content provider" language of the statute.

Several pending lawsuits, including the Walters v. OpenAI defamation case, are testing this question. The early signals suggest courts will not extend Section 230 to AI-generated content created by the platform itself. For California product liability claims, this matters because it removes a defense AI defendants might otherwise raise.

Statute of Limitations for AI Liability Claims

The standard two-year statute of limitations under CCP Section 335.1 applies. The clock starts on the date of injury. For AI cases involving harms that aren't immediately apparent (a financial AI that makes bad recommendations over months, a medical AI that under-diagnosed a condition that progressed silently), the discovery rule may extend the deadline. The plaintiff has to act promptly once the injury and its cause are reasonably knowable.

Frequently Asked Questions

Can you sue an AI company in California for harm caused by its system?

Yes, in most cases. California's existing product liability framework applies to AI systems sold as products. Liability can reach the developer, deployer, integrator, and seller.

Is software a product under California law?

For mass-distributed consumer software embedded in or marketed as a product, generally yes. Custom bespoke software is more often treated as a service.

Who is liable when an autonomous vehicle causes an accident?

The vehicle manufacturer, the AI software developer, component suppliers, the fleet operator, and in some cases a supervising human driver. California strict liability reaches the whole chain.

What California laws apply to AI liability?

The general product liability framework (Civil Code Section 1714, Greenman, Barker, CCP Section 335.1) and AI-specific disclosure statutes like AB 2013 and SB 942 from 2024.

What does it mean for an AI system to be defective?

Defective by design (unsafe architecture or training), defective in manufacture (a specific deployment is broken), or defective by warning (known limitations not disclosed). All three categories apply.

Can I sue if a chatbot gave me bad advice that caused harm?

Possibly. The analysis depends on what the AI was marketed for, what warnings were given, and whether reliance was reasonable. This area of law is still developing.

Are AI companies protected by Section 230?

Probably not for content the AI itself generated. Section 230 protects platforms from user-generated content, not platform-generated outputs.

How long do I have to sue for AI-related harm in California?

Two years from the date of injury under CCP Section 335.1. Discovery rule exceptions may apply.