The Responsible Party is… Not?

I. Introduction

In February 2024, a 14-year-old named Sewell Setzer III took his own life after spending months in conversation with a Character.AI chatbot. His mother sued. The company's first move was to argue that chatbot outputs are protected speech under the First Amendment, that holding an AI developer liable for what its system says is no different from holding a publisher liable for what it prints.¹

That argument did not work. But the fact that it was tried at all says something about where AI liability law actually stands, that developers are still figuring out what theory, if any, can make them responsible for the harm their systems cause. So far, no one has a clean answer.

II. Frameworks in Conflict

The problem is not that tort law ignores AI. It is that the frameworks we have were built for a different kind of harm. Negligence requires proving a defendant breached a duty of care, but there is no established legal standard for what "reasonable care" means when you are building a large language model.² Strict product liability holds manufacturers responsible for defective products regardless of fault, but courts have historically treated software as a service rather than a product, which cuts off that theory before it starts.³ Both frameworks assume you can trace harm back to a specific decision by a specific actor. AI complicates that because LLMs are stochastic by design. Businesses deploy systems in contexts that developers never anticipated. Users interact with them in ways nobody planned for. It is as if Achilles had many heels but no single archer, and when the arrow lands, the wound is real, but nobody can agree whose bow fired it.

Garcia v. Character Technologies is the first case to seriously test these questions.⁴ Judge Anne Conway's May 2025 ruling allowed strict product liability, negligence, and wrongful death claims to proceed, and rejected the First Amendment defense outright. It is a significant win for plaintiffs, and it established something courts had not said before. An AI chatbot can be treated as a product rather than a speaker, but a motion to dismiss ruling only asks whether the claims are plausible. It does not assign liability, and it does not resolve whether AI broadly meets the legal definition of a "product" across jurisdictions.⁵

Meanwhile, the EU answered one of those questions. In 2024, it extended its Product Liability Directive to explicitly cover software, including AI, regardless of whether it is embedded in physical hardware.⁶ No equivalent action has happened in the United States. That gap is not a minor detail. It means that a company can build an AI system, deploy it to millions of users, watch it cause documented harm, and face no strict liability exposure, because American courts have not decided whether the thing they built counts as a product.

III. The Black Box Problem

Large language models process inputs through billions of numerical parameters, producing outputs through statistical transformations that resist straightforward interpretation. Unlike traditional software with explicit logical branches, these systems have no clear decision path to trace from input to output. When a model generates a harmful response, developers cannot audit a sequence of steps to identify where the system went wrong. The "black box" problem makes this worse.⁷ Even in cases where someone was clearly harmed and an AI system was clearly involved, establishing that a specific design choice caused that specific harm is genuinely difficult. The internal logic of large language models is opaque. Developers cannot always explain why their systems produce a given output. Courts cannot easily apply traditional causation analysis to something that works that way. Plaintiffs end up needing to prove things that may be unprovable.

IV. Legislative and Judicial Responses

Some legislators are trying to build something better. Colorado's SB24-205 places joint responsibility on both AI developers and the businesses deploying AI in high-stakes contexts, employment, lending, and similar decisions, requiring both parties to disclose how the AI is used and to prevent discriminatory outcomes.⁸ Senator Blackburn's proposed federal bill would impose a statutory duty of care on developers to address foreseeable risks, while preserving states' ability to enforce their own rules.⁹ These are by far not perfect solutions, but they at least address the right problem: assigning responsibility at the points where decisions are actually made, rather than waiting for a victim to file suit and figure out who to name.

The Trump administration's National AI Legislative Framework, released in March 2026, would preempt state-level AI regulation and consolidate rulemaking at the federal level.¹⁰ The stated goal is consistency and reduced friction for developers. The practical effect, if Congress follows the administration's lead without enacting its own substantive protections, could be stripping states of their ability to experiment while replacing that authority with nothing enforceable. That is not a hypothetical concern. It is what happens when preemption is not paired with a federal floor.

V. Conclusion

There is an argument that courts should develop an AI liability doctrine the way tort law has always developed, incrementally, through specific cases, building principles out of hard facts over time. Garcia may eventually do that. But "eventually" is doing a lot of work in that sentence. The litigation timeline for a single case runs for years. The pace of AI deployment runs for months. New systems are reaching millions of users faster than any court can process the claims from the last ones. Waiting for case law to catch up is a choice, and it is a choice that costs real people in the meantime.

The current frameworks are not adequate. Negligence lacks a standard. Product liability lacks a threshold. Federal legislation is either stalled or tilted toward industry. What exists is a patchwork of state experiments that may soon be preempted, a handful of cases in early discovery, and a set of genuine doctrinal questions that nobody has resolved. That is not a legal system equipped to handle what is already happening, let alone what comes next.