A self-driving vehicle runs a red light and kills a pedestrian. The plaintiff’s lawyer demands that the manufacturer explain why the car’s algorithm chose to accelerate through the intersection rather than stop. The manufacturer’s engineers pull the logs, review the sensor data, and reconstruct the decision sequence. They can tell you what the system did, but they cannot tell you why. The neural network that processes sensor inputs and selects the driving action operates through millions of weighted connections trained on billions of data points. No engineer designed those weights by hand, and no programmer wrote an if-then rule that said run the red light. The system taught itself, and the reasoning behind its decision is embedded in a mathematical structure that no human can fully interpret. That is the black box problem, and it is about to become the central challenge in product liability litigation.
Every consumer product that relies on machine learning to make decisions shares this problem: AI-powered medical diagnostic tools, smart home devices that control gas and electricity, industrial robots that work alongside human employees, and autonomous vehicles that share the road with pedestrians all depend on algorithms that learn from data rather than follow explicit instructions. When those products cause harm, the plaintiff’s first question will be the same question it has always been in product liability cases: What went wrong? The difference is that, with AI-embedded products, the manufacturer may not be able to answer that question in terms a jury can understand, and that inability to explain is becoming the basis of the claim.
Why Traditional Product Liability Doctrine Struggles with AI
Product liability law developed around products that could be understood: A pressure cooker has a locking mechanism, an automobile has brakes, and a pharmaceutical has a chemical formula. When those products fail, engineers and experts can trace the failure to a specific component, decision, or manufacturing error. The design-defect analysis asks whether a reasonable alternative design would have reduced the risk of harm. The failure-to-warn analysis asks whether the manufacturer adequately communicated the risk. Both analyses assume that someone, somewhere, can explain how the product works and why it failed.
AI-embedded products break that assumption. A machine-learning model does not have a locking mechanism that failed or a brake line that ruptured. It has a trained model with millions of parameters that processes an input and produces an output. The path from input to output is mathematically traceable but practically incomprehensible. An expert can tell a jury that the system assigned a particular weight to a particular data feature. Still, that explanation does not answer the question the jury actually cares about: Should the car have stopped? Should the diagnostic tool have flagged the tumor? Should the robot have recognized the worker standing in its path?
The black-box problem gives plaintiffs’ lawyers a powerful structural advantage: They do not need to explain why the algorithm failed; they only need to point to the harm and argue that the manufacturer deployed a system it could not control, predict, or explain. The burden then shifts, practically if not legally, to the defense to demonstrate that the system was reasonably safe despite the manufacturer’s inability to trace every decision the system makes. That is a hard argument to win in front of a jury that just saw photographs of the plaintiff’s injuries.
The Explainability Gap
Plaintiffs are beginning to argue that the inability to explain an AI system’s decision-making process is itself a design defect. The theory runs like this: A manufacturer that deploys a product capable of autonomous decision-making but incapable of explaining those decisions has designed a product that is unreasonably dangerous because it cannot be monitored, audited, or corrected in real time. A reasonable alternative design would incorporate explainability features that allow the manufacturer, the user, and regulators to understand why the system made the decisions it made.
This theory has real teeth. The Restatement (Third) asks whether omitting a reasonable alternative design renders the product unsafe. If the plaintiff can present expert testimony that the manufacturer could have used an interpretable model instead of a black-box model, or could have implemented a logging system that recorded the factors driving each decision, or could have built a human override that activated whenever the system’s confidence level dropped below a threshold, the jury has a basis for finding that the design was defective. The plaintiff does not need to prove that the specific decision that caused harm was wrong. The plaintiff only needs to prove that the system’s inability to explain its decisions rendered it unreasonably dangerous as a whole.
Defense counsel must take this theory head-on. The response is not that black-box models are safe. The response is that interpretable models are not always available, and when they are, they often sacrifice the accuracy and performance that make the product effective in the first place. A simpler model that can explain its reasoning but misses a tumor on a scan is not a reasonable alternative to a complex model that catches the tumor but cannot explain how. The tradeoff between explainability and performance is real, and defense counsel must force the jury to confront it rather than pretend it does not exist.
The Defect That Did Not Exist at the Time of Sale
Traditional product liability law evaluates the product as it existed when it left the manufacturer’s control. A product that was safe at the time of sale does not become defective simply because it later causes harm. AI-embedded products challenge that principle because they change after deployment. Machine learning models that continue to train on new data after the product ships can develop behaviors that did not exist when the product was sold. A chatbot that was safe at launch can become dangerous after ingesting months of user interactions, leading to shifts in its response patterns. An autonomous vehicle’s driving model can degrade when it encounters edge cases not represented in its training data.
This creates a timing problem for both sides. Plaintiffs will argue that the manufacturer had a duty to monitor the system’s post-deployment behavior and intervene when the system drifted from its intended performance envelope. Defense counsel will argue that the product was not defective at the time of sale and that subsequent changes in the model’s behavior were caused by inputs and conditions outside the manufacturer’s control. Both arguments have merit, and courts have not yet established a clear framework for resolving this tension.
The strongest defense is to demonstrate that the manufacturer implemented robust post-deployment monitoring systems, established performance benchmarks that triggered reviews when the model’s behavior changed, and maintained a rapid-response protocol for pushing corrective updates when the system drifted. A manufacturer that can show it monitored the system, measured its performance, and intervened when necessary has a powerful answer to the argument that it deployed a product and walked away. A manufacturer that shipped the product and never looked at it again has no answer at all.
Before the Next Product Ships
Defense counsel representing companies that develop or deploy AI-embedded consumer products should deliver specific guidance right now, while the doctrine is still forming and before the first wave of verdicts establishes the rules that will govern this space for the next generation.
First, clients should document every design decision involving the choice between model complexity and interpretability. If the engineering team chose a deep neural network over a decision tree, record why. If explainability features were considered and rejected, record the tradeoff analysis that drove the decision. If the team implemented partial explainability tools, such as feature-importance rankings or confidence scores, document what those tools can and cannot tell the user. Every one of these decisions will be second-guessed in litigation. The manufacturer that documented them will control the narrative. The manufacturer that did not comply will face the narrative imposed by the plaintiff’s expert.
Second, clients should build post-deployment monitoring into the product architecture from day one. The system should log every decision it makes, the inputs that drove the decision, and the confidence level associated with the output. Those logs should be reviewed regularly by qualified engineers, and any deviation from expected performance should trigger an investigation and, if necessary, a corrective update. The monitoring system is both an engineering safeguard and a litigation asset. A manufacturer that can show the jury two years of monitoring data demonstrating consistent, safe performance has a far stronger defense than a manufacturer that has nothing to show at all.
Third, clients should implement meaningful human oversight at the points where the AI system’s decisions carry the greatest risk. In a medical diagnostic tool, that means requiring physician review of every AI-generated recommendation before it reaches the patient. In an autonomous vehicle, that means maintaining driver engagement requirements and intervention capabilities. In a consumer product, that means giving the user clear, prominent notice that the product uses AI to make decisions and that those decisions may not always be correct. Human oversight does not eliminate liability; it distributes it, and it gives the defense the argument that the system was designed to operate with a human in the loop who had the ability and responsibility to catch errors.
Fourth, clients should prepare for the discovery fight of a lifetime. Plaintiffs in AI product liability cases will demand access to training data, model architecture, testing protocols, internal communications about known limitations, and every document that touches the system’s safety evaluation. The volume of discoverable material in an AI product liability case dwarfs a traditional product case. Prepare the privilege log now. Identify the documents that are protected and the documents that are not. Organize the technical documentation in a format that defense counsel and defense experts can navigate efficiently. The manufacturer that is organized wins the discovery war. The manufacturer that is not organized loses the case before trial.
AI product liability law is being written right now in courtrooms; not in law-review articles. The first generation of verdicts will establish the principles that govern this field for decades. Defense counsel who understand the technology, can explain black-box decision-making in terms a jury can grasp, and can present the engineering tradeoffs honestly and persuasively, will win cases that less-prepared lawyers will lose. The algorithm cannot explain itself. The defense lawyer must.
About the Author:
Frank Ramos is a partner at Goldberg Segalla in Miami. framos@goldbergsegalla.com