Science and Technology Law Review

Monitoring, Oversight, and Learning in Medical AI

W. Nicholson Price — 2026-06-08

When medical AI errs, it often goes unnoticed. If there’s a specific patient injury, and the link to AI is obvious, that problem might be reported to the Food and Drug Administration (FDA), but not always. And many other types of problems, like worse performance on specific groups or ineffective integration into health system workflows, simply don’t fall within the contours of regularized reporting. Even if they are noticed by the health system—far from a given—there’s no obvious way to share that information more broadly. Against this backdrop, there are justified calls for better oversight and reporting. But there’s the opportunity to do more. If now is the time to build more robust surveillance systems and standards for sharing that information, it should also be the time to build systems to share information about positive performance and learning, so that AI can help enable the vision of a learning health system that not only fixes mistakes but also constantly improves.

The Role of Standards in Enabling the AI Stack

Christopher Yoo — 2026-06-08

The debate over AI standards is undergoing a significant reorientation. In major jurisdictions, such as the United States, European Union, United Kingdom, and Australia, policymakers are increasingly shifting from a predominantly harm-centered approach toward one that emphasizes productivity, innovation, and the beneficial uses of AI while still managing risk. At the same time, AI is increasingly being developed and deployed through a multi-party stack composed of data providers, foundation model developers, fine tuners, deployers, and other actors rather than by a single integrated firm. In this environment, standards are essential not merely as a safety tool but as enabling infrastructure: they clarify roles, define interfaces, support validation, facilitate interoperability, and create the basis for ex post evaluation of real-world performance. The Essay further contends that standards should generally emerge through flexible, multistakeholder processes and will likely vary across industry verticals rather than take the form of a single universal rule set. It explores five core components of effective AI standards: substantive performance requirements, meaningful data disclosures, transparent validation methods, protection against attacks, and articulation of acceptable levels of risk measured against real-world counterfactuals rather than zero-risk baselines. Properly designed, such standards can unlock AI’s economic and social value while also mitigating bias, error, and safety concerns.

Emulating Brussels but Recalibrating in Seoul: Challenges in South Korea's AI Framework Act

Sangchul Park — 2026-06-08

This paper examines the pathway of South Korea’s AI Framework Act (“AIFA”), which became the world’s first fully implemented horizontal AI regulatory regime on January 22, 2026. While the United States and other leading innovators have largely pursued sectoral or context-specific AI regulatory frameworks, the AIFA initially moved toward an EU-style horizontal framework. This was driven by the institutional incentives from the leading regulatory agency and the political clout of citizen activist organizations.

However, the rapid acceleration of AI adoption, a global shift toward “light-touch” governance, and Seoul’s “AI G3” agenda called for a strategic recalibration to mitigate regulatory overlap and unnecessary compliance burdens. Instead of amending the primary statute, the regulator used administrative rulemaking to alter the underlying logic of the regime. This pivot shifted the AIFA away from a safety-centric structure designed to protect “affected persons” from developers and deployers into a quasi-utility regime to shield “users” (including downstream deployers) from providers and distributors.

While this change narrowed the scope of application of the AIFA, it misaligned accountability across the AI value chain and pushed developers to protect sophisticated deployers who often wield substantial operational control. In response, the Presidential Council on National AI Strategy (“AISC”) has mandated legislative amendments in the National AI Action Plan to rectify these structural flaws.

By examining unresolved tensions in the AIFA and the rationale behind the AISC mandate, this paper suggests an alternative framework that balances context-specific and horizontal approaches. Furthermore, it offers three lessons for jurisdictions that are modeling legislation after the EU’s horizontal approach with less stringency: (i) even within streamlined rules, the differences between actors in the AI value chain must remain legally distinct; (ii) scalability is best achieved through taking up international industry standards rather than ambiguating legal duties or covered entities; (iii) the regulatory structure must decouple the rules for the underlying AI systems and regulations for AI-enabled products and services.

AI Suppression: E-Discovery Software And Brady

Jason Hartline — 2026-06-08

Prosecutors regularly rely on AI e-discovery software, known as technology assisted review (TAR) tools, to sort and prioritize digital evidence. These tools implicate constitutional concerns: they can either risk suppressing or help to surface exculpatory and impeachment evidence that prosecutors must disclose under the Brady due process rule. Yet doctrine, agency guidance, and scholarship offer virtually no direction on their use.

This Article examines how TAR affects Brady compliance. Using computer science simulations on synthetic data sets, we show that TAR can either hide or help to expose Brady evidence, depending on how it is configured and the configurations of evidence to which it is applied. From these results we derive three TAR workflow recommendations: prosecutors should run TAR separately for inculpatory and Brady evidence; TAR coding of Brady material should be permitted even when active searching is constitutionally contested; and procurement guidelines should favor flexible classifiers.

Our examination of TAR also highlights unresolved tensions in Brady doctrine: whether liability attaches when the prosecution possesses but does not know about Brady evidence; how Brady interacts with Fourth Amendment privacy protections; and whether Brady should be limited to preventing suppression, as current doctrine states, or expanded into a full duty to assist defense investigations. We argue that Brady liability should apply strictly to all Brady evidence in the control of the prosecution team, regardless of whether anyone on the team knows or has reason to suspect that it exists.

An Initial Assessment of Standards in Technology Tort Litigation

Justin Hurwitz — 2026-06-08

Standards are a central preoccupation of AI governance, yet the dominant policy conversation—focused on regulatory compliance and coordination—does not capture how standards actually function in the legal domain most likely to govern AI-related harms: tort law. This Article offers an empirical and doctrinal assessment of standards in technology tort litigation, surveying reported decisions to map how courts use standards across negligence and products liability contexts.

The central finding is that tort law consistently resists treating standards as dispositive. Compliance is almost never conclusive; noncompliance is rarely sufficient by itself. What matters is how a standard is used: what it is offered to prove, whether it applies to the defendant and the risk at issue, whether it reflects contemporaneous knowledge, and whether it is presented through reliable expert methodology. This pattern recurs across technological domains—from railroads and automobiles to medical devices and digital platforms—because tort law is structurally committed to contextual, fact-intensive evaluation rather than categorical rule-following.

These findings carry direct implications for AI. Because AI standards are emerging unusually early in the technological lifecycle—before stable engineering norms or accumulated accident experience exist—courts are unlikely to treat them as duty-defining baselines or safe harbors. Instead, they will enter tort litigation primarily as evidentiary tools: anchors for expert testimony, benchmarks for feasibility arguments, and reference points for evaluating warnings and governance practices.