Visible to the Court: How AI Is (and Isn't) Litigated in U.S. Federal Court Opinions

Does the Law Actually See AI?

The legal system is currently managing the rise of artificial intelligence through a form of "piecemeal" governance. Rather than creating new, specialized AI laws, federal courts are forcing modern technological problems into aging legal boxes.

Researchers from the University of Washington investigated this phenomenon. They wanted to know if the federal court system is effectively capturing the actual risks posed by AI. Their study reveals a significant gap between the harms documented in the real world and the issues actually being resolved in court.

Mapping the AI legal battlefield

The researchers sought to map the current landscape of AI litigation in U.S. federal courts. They focused on "cognizability" (whether a harm is recognized by the law as something a person can actually sue over).

Even if an AI system causes a massive societal problem, a plaintiff cannot win a case unless that problem fits an existing legal category. This might include a breach of contract or a violation of a specific privacy statute. The authors aimed to see if the legal system is keeping pace with technology or merely reacting to whatever fits its current toolkit.

The myth of specialized AI law

One might assume that as AI becomes prominent, courts would develop a specialized body of "AI law." There was an expectation that the unique technical properties of machine learning would necessitate new legal doctrines.

However, the data suggests a different trend. While the volume of AI-related cases is rising—doubling since 2023 —the legal reasoning remains rooted in the past.

The study finds that courts are not creating new AI-specific rules. Instead, they are performing "piecemeal" governance. They take brand-new technological problems and force them into old categories like intellectual property or consumer protection law. This reliance on legacy statutes creates a bottleneck. The legal system's ability to provide recourse is limited by the flexibility of existing laws rather than the severity of the AI harm.

Decoding 559 federal opinions

To move beyond anecdotes, the authors conducted a systematic review of 559 U.S. federal court opinions. These were cases where AI played a central role. They retrieved these documents from GovInfo.gov, an official repository of government information.

The methodology involved a rigorous, multi-stage process. First, the team developed a "codebook" (a standardized set of definitions used to categorize cases). They refined this through five iterations of manual coding. They then used GPT-5 to annotate the remaining cases. They validated this by manually reviewing a random subset to ensure high agreement between the human and AI coders.

The researchers categorized the data across three dimensions: dispute topics, specific technologies, and types of litigants. By comparing their findings to the AI Incident Database (AIID)—a repository that tracks real-world AI failures—they measured the "disconnect" between real-world events and courtroom resolutions.

A mismatch between risk and recourse

The results reveal a striking divergence between documented AI risks and legal reality. The authors report that the most common dispute topics are "AI in Legal Proceedings" (38.82%) and Intellectual Property (24.15%) [Table 1].

It is important to distinguish the "Legal Proceedings" category. This refers to cases where AI is used as a tool within the court process itself, such as for legal case management. This is distinct from cases where the AI technology is the actual subject of the legal fight.

The most significant finding is the "incident and opinion disconnect." When the researchers compared their findings to the AIID, they found that many major AI risks do not appear in court. For example, while the AIID contains significant entries for "system safety" and "disinformation," the authors found no direct parallels for these in federal court opinions.

Furthermore, the study finds that the prevalence of certain harms is heavily skewed. While "unfair discrimination" is a frequent topic in AI incident reports, it represents only 4.69% of the litigated topics in this study. The authors suggest this is because proving discrimination in an opaque algorithmic system is a massive evidentiary burden. Plaintiffs often must prove "intent" rather than just "disparate impact" (the statistical reality that a system affects one group more than another). Consequently, many AI harms remain unresolved because they do not fit easily into current legal doctrines.

The consequences of piecemeal governance

The implications of this research are twofold. First, AI accountability is currently being driven by "upstream" leverage points rather than "downstream" outcomes. Because it is difficult to sue for the abstract harm of a biased algorithm, plaintiffs focus on how data was collected or how a product was marketed. This incentivizes transparency and disclosure over actual responsibility for the harm caused by an AI's decision.

Second, this "piecemeal" governance creates a fragmented landscape. If regulation only emerges case-by-case, there is no unified standard for developers to follow. This leads to an unpredictable legal environment. The paper concludes that until specific legislation is introduced, the legal system will remain reactive. It will only address the harms that happen to fit into the narrow windows provided by existing statutes.

Figures from the paper