Courts are now being asked to decide how we should govern artificial intelligence. In Florida, a judge recently let a tort suit proceed under strict liability after a mother alleged that a chatbot built by Character AI encouraged her fourteen-year-old son to take his life. The sorrow is undeniable. The legal stakes are broad because strict liability holds a party liable for harm caused by a defective product or an abnormally dangerous activity, regardless of fault—that is, without requiring proof of negligence or intent.
If this decision is affirmed and becomes the first in a wave of strict liability claims against AI developers, it could reshape our emerging AI ecosystem, heavily favoring the seen harms and against the unseen benefits of artificial intelligence. First, it underscores the caution with which courts should approach emerging issues in AI because of their ramifications for the development of AI, an essential public good, including for our national security. Second, it highlights a foundational defect in the modern expansion of tort law: the tendency to focus on the visible victim while ignoring the silent mass of beneficiaries.
No one understood the mistaken impulse behind strict liability better than one of the greatest friends of liberty France has ever produced—the economist Frédéric Bastiat. He famously argued that it was necessary to distinguish between the seen and the unseen to achieve sound public policy. The broken window may generate work for the glazier, he observed, but it destroys the wealth the shopkeeper might otherwise have used elsewhere. Likewise, a chatbot’s harmful output is seen, while many potential benefits will go unseen. These could include easing the burden of everyday tasks, providing life-improving services, and—insofar as they often refer users to prevention resources—perhaps even averted suicides. For instance, studies of old age homes have shown that such interaction with chatbots reduced loneliness and caused more engagement with other residents. To ignore these benefits when assessing legal liability is both economically myopic and morally obtuse.
Strict liability for AI might appear attractive because it promises compensation for visible harm. But in practice, it reflects an epistemological arrogance—a presumption that courts can know all they must to weigh the full consequences of imposing faultless responsibility. As a result of this presumption, strict liability could make society worse off by discouraging the deployment of AI systems that are, on net, vastly beneficial due to their behind-the-scenes impact in specific cases.
This chatbot is designed not to give advice but to emulate human conversation—a kind of artificial companion. In a few cases, such as this one, the chatbot may have provided damaging content. But how many struggling users have found solace, distraction, or even life-saving comfort from interacting with such systems? We do not know. And we are unlikely ever to know. Courts deal in the seen, factual narratives. They cannot reliably quantify the unseen benefits—including the number of suicides that may have been prevented by interaction with the technology.
Moreover, the risk of strict liability will deter others from making useful AI products. If even after taking reasonable care, one can still be held liable for one’s product, there is less reason to innovate in product design. The unseen victims become those who would have benefited from innovations that never launch or arrive years late: patients denied therapies, commuters without safer transport, consumers missing cheaper energy, all of which will come from an AI start-up. Sound policy must price residual risk without suppressing experimentation. The fault standard does a better job of this because it holds a company liable only if they could have reasonably made the product safer.
Proponents of strict liability argue that such risks can be managed: perhaps AI companies can defend themselves by showing that their systems generate net benefits. But this analysis would turn liability into a kind of calculus that courts are ill-suited to perform. Can a judge or jury meaningfully weigh the emotional toll of a single visible suicide against the diffuse, statistical uplift experienced by millions of unseen users? In a courtroom, what is concrete will almost always dominate what is abstract.
To put this more formally, proceeding case by case and bound by rules of evidence, judges and juries face a serious knowledge problem. The information needed to measure those gains is dispersed across time and users. It is thus better revealed by iterative engineering, markets, or agencies capable of sustained cost–benefit analysis than by a single verdict. Comparative institutional competence counsels epistemic humility. Courts should punish proven negligence, but resist strict‑liability schemes that require them to feign omniscience about the unseen and, in doing so, throttle the decentralized discovery process that makes society safer and more prosperous.
Autonomous vehicles (AVs) raise the same issue. By one serious estimate, self-driving cars are already safer within their scope of operation. Human drivers are typically held to a negligence standard. Why should their successors face a more punitive rule?
Strict liability places a regulatory moat protecting the concentration of AI in a few, big tech actors who can self-insure.
One answer from liability proponents is that AI products are opaque and unpredictable—black boxes whose decisions we cannot fully understand or control. This might make it difficult for plaintiffs to prove negligence, so strict liability becomes necessary to ensure compensation. But this seems an exaggeration. If there is something that AV programming could do systematically, better software architects should be able to suggest it in expert testimony. If we want more safety on the roads, we should not deter innovation that self-driving vehicles represent. Moreover, AV vehicles not only deliver safety but also improve the quality of life for many, as those who once drove can work and play instead.
There is another reason not to impose strict liability on AI systems now. AI should be understood as a kind of public good—even the essential public good of modern times. The premise of President Trump’s recent executive order on AI is that the government should facilitate the development of AI for the long-term flourishing of its citizens. First, the nation that wins the AI race will have an enormous national security advantage. The strongest competitor in this race, moreover, is Communist China, a totalitarian regime that is now our chief ideological and geopolitical rival. Second, AI, being a general-purpose technology, will accelerate research for other public goods. Better AI will likely lead to a wide range of scientific breakthroughs. For instance, it will help address the problems of climate change both by accelerating new forms of clean energy and by making more efficient use of energy.
Progress in AI is incentivized by the profitable innovations it generates. Since strict liability will impose costs on those innovations, it will depress the production of this essential public good. Judges should also hesitate before imposing such costs, particularly when a democratically elected president has declared AI a national priority. In short, America’s geopolitical competitors are not holding back their AI deployments for fear of liability exposure. We should not hobble our own nation by overextending tort.
And consider that we already limit tort liability in domains essential to other public goods. For instance, the First Amendment protects much harmful speech, even if it is false and offensive, because the law recognizes that the broader good of free expression outweighs its dangers. In fact, we impose not strict liability but the most forgiving standards, even more lenient than negligence, for public figures. Unless driven by malice, even lies create no liability. The public good of intense democratic debate influences the restrained nature of the liability for libel under these circumstances.
Strict liability invites a troubling overreach in legal design. If we are to internalize all risks of AI, even those not due to fault, what principle will contain the scope of liability? Any AI system embedded in public services, education, or consumer products may generate unintended harm. The more AI is used, the more likely some incident will occur. Yet if the net effect is to reduce harm overall, strict liability becomes an engine of risk-aversion.
It will also have a particularly deleterious effect on startups, where much AI innovation comes. They cannot self-insure against the risks that strict liability imposes. Strict liability places a regulatory moat protecting the concentration of AI in a few, big tech actors who can self-insure. That kind of legal regime is not in the interest of either innovation or lower prices for consumers.
The better path is to apply fault-based standards. Thus, developers should be held accountable for genuine negligence. If a chatbot was deployed with known flaws, if guardrails were ignored or disabled, then fault-based liability is entirely appropriate. But we must not blur the line between failing to prevent harm and being responsible for it by default.
And when AI systems serve public functions with uncertain risks and dispersed benefits, we might consider alternative compensation schemes—modeled on vaccine injury funds or no-fault insurance. These instruments separate the question of fault from the goal of victim compensation, without chilling technological progress.
Law cannot heal grief, but it can refuse to let grief write the rules. The loss at the heart of this case is tragic, but strict liability would turn anguish into a tax on technologies whose benefits are yet largely unseen. Judges should punish negligence, not faultless innovation. New rules governing AI should reflect epistemic humility about the judicial capacity to evaluate its benefits, particularly when those benefits include advancing a national public good.