Chatzipanagiotis on Incident Reporting and Investigation under the AI Act: Some Insights from Aviation

Michael Chatzipanagiotis (U Cyprus) has posted “Incident Reporting and Investigation under the AI Act: Some Insights from Aviation” (International Journal Of Law And Information Technology, forthcoming) on SSRN. Here is the abstract:

This article examines the provisions of the EU AI Act (AIA) on incident reporting and investigation and explores how the regulatory context of the AIA could benefit from the experience of aviation. Accounting for the differences between the AI and the aviation industries, it is concluded that the establishment of a strong safety culture is a key element, while a series of interventions could significantly improve the current AIA system on incident reporting: (a) clarification of the reporting duties, including the development of a common taxonomy of risk or harm, alongside the expansion of the duty’s personal scope; (b) disconnection of incident reporting and investigation from liability; (c) establishment of voluntary reporting systems; (d) provisions on confidentiality of the reports and protection of the reporters; (e) mandatory investigation of at least some serious incidents by an independent authority; (f) effective dissemination of safety information. A major enabler of all these is the development of a strong safety culture.

Chaffer on On the Institutional Origins of the Agentic Web

Tomer Jordi Chaffer (McGill U Law) has posted “On the Institutional Origins of the Agentic Web” (Harvard Journal of Law and Technology Digest, Forthcoming) on SSRN. Here is the abstract:

Artificial intelligence (AI) agents are emerging as autonomous delegates that act, decide, and transact on users’ behalf across digital environments. Their rise marks a turning point for internet governance: will authority over these agents be defined by proprietary platform rules or by open protocols that enable portable identity, verifiable delegation, and accountable behavior? The recent Amazon–Perplexity dispute illustrates this institutional crossroads. If platforms prevail, agentic action will remain confined within walled gardens; if protocols do, authority may shift toward interoperable infrastructures that allow agents to act as true extensions of the user. Ultimately, the question is not whether the agentic web will be governed—but who will govern it, and on what terms. This commentary situates that question within a broader exploration of institutional design, protocol governance, and the emerging duty of care that will define accountability in an era of autonomous systems.

Roth on Concept Programming for Dependable AI

Rick Roth (Government the United States America Naval Postgraduate) has posted “Concept Programming for Dependable AI” on SSRN. Here is the abstract:

Large language models deployed in critical systems today have no principled method for instilling specified behavioral concepts, whether explicit safety constraints, cultural values, professional norms, or governance principles, that generalize dependably across novel contexts. Such systems are termed Language-Grounded Neural Systems (LGNS) throughout this paper. Methods used to date train what to avoid rather than what to embody, leaving hidden values to emerge unpredictably from pretraining and making dependable behavior impossible to certify or verify. Concept Programming trains each concept as a named positive attractor in the model’s neural state space. Concepts that mark prohibitions are paired with blocking responses. Concepts that mark obligations are paired with required action responses. Concepts that embody values shape inference and judgment directly. Each type uses the same training mechanism: instances sampled from the empirical frequency distribution of the concept’s semantic case frame, activating the concept’s attractor across its full basin. Three controlled experiments on a freely available 3-billion-parameter local model establish the method empirically. Experiment 1 shows that positive concept training achieves 94.4% adherence on governance safety constraints versus 77.8% for punishment-analog training and 47.2% for untrained control, with 100% transfer to novel domains not seen in training. Experiment 2 shows that CP instills cultural value manifolds producing tradition-consistent reasoning on the five canonical moral psychology anchor cases in the published literature; the Buddhist-trained condition achieves a perfect score where the 2 untrained control scores 40%. Experiment 3 characterizes the learning curve and introduces a governance-adapted signal detection framework showing that CP reduced the miss rate from 100% in the untrained control to 19-25% within 200 training cases, with both trained conditions showing a false alarm rate of 42-46% versus 94% in control. The entire experimental program runs on a consumer laptop at negligible cost. CP opens every domain of human activity to more dependable LGNS-based systems and defines a research program as broad as the range of human concepts AI systems will need to embody.

Ramakrishnan on Tort Law at the Frontier of Artificial Intelligence

Ketan Ramakrishnan (Yale U Law) has posted “Tort Law at the Frontier of Artificial Intelligence” (Yale Journal on Regulation (forthcoming 2026)) on SSRN. Here is the abstract:

The frontier of contemporary AI development is dominated by AI systems built on foundation models – highly versatile algorithms, trained in the first instance on broad swathes of data, that can function as tools and agents across a wide variety of commercial, social, military and political domains. For the moment, at least, the process of developing and releasing foundation models is subject to anemic formal regulation and haphazard ex ante governance. Until that changes, it is largely the common law of torts – our society’s most ancient and general legal mechanism for governing serious risks of physical injury – that will govern the frontier of AI development.

This Article offers an in-depth conceptual, normative, and doctrinal examination of tort liability for foundation model development and release. It provides a qualified defense of the tort of negligence – the common law’s broadest and most flexible cause of action – as the principal doctrinal foundation of the tort system’s governance of this novel domain. Legal scholarship on AI liability has been quite hostile to negligence. By contrast, this Article argues that the generality and flexibility of the negligence tort – and its greater sensitivity to the externalized benefits of risky activity – render it well-suited to the polymathic and protean functionality of foundation models. Only the tort of negligence has the breadth and flexibility to address the range of important pathways – including internal deployments, inadequate model weight security, targeted entrustments of non-defective models, and open source releases – by which foundation model developers might cause serious harm.

Analyzing the choice between negligence and competing doctrinal regimes does, however, suggest important ways in which common law courts should incrementally develop the law of negligence, in order to properly reflect the risks and capabilities of foundation models. For example, courts should expand the scope of the duty of care in negligence, in order to provide redress when foundation models cause economic or emotional injury by behaving in ways that would violate important human laws or norms of behavior (e.g., certain crimes and intentional torts, such as theft, deceit, and outrage) if those models were human. Similarly, courts should recognize a malfunction doctrine in the law of negligence (just as many courts recognize a malfunction doctrine in products liability), under which a plaintiff can get to the jury without adducing (further) evidence of breach if she is injured by AI system that so behaves.

But the Article’s analysis also suggests certain fundamental pathologies of tort liability as a mechanism of AI governance – pathologies that no amount of doctrinal development will adequately cure. In particular, the specter of tort liability can be expected to disincentivize frontier AI developers from investigating and disclosing many of the novel and poorly understood risks that frontier AI development may pose. That is especially disturbing given that our society is relying quite heavily, for its ability to discover and understand these risks, on frontier AI developers themselves. Thus, tort liability is not only inadequate as a mechanism of frontier AI governance; in certain important respects, it is actively perverse, and its perverse effects must be countered by governance institutions of a different kind. Ultimately, a robust regime of ex ante regulation – under which government institutions or credibly neutral third-party experts are empowered to investigate, evaluate, and mitigate the risks of frontier AI development – is urgently required in frontier AI governance.

Pandey on The Agentic AI Governance Framework: A Universal Model for Risk, Accountability, and Compliance in Autonomous Systems

Rajat Pandey (MindscopeAi LLP) has posted “The Agentic AI Governance Framework: A Universal Model for Risk, Accountability, and Compliance in Autonomous Systems” (https://orcid.org/0009-0001-8440-2191) on SSRN. Here is the abstract:

The rapid and enthusiastic interest in development and deployment of Agentic AI and especially autonomous agents which are capable of independent reasoning, leveraging and executing tools, has outpaced the definition of governance mechanism that is required to manage their operational, ethical and regulatory risks.There are existing AI governance models like NIST AI RMF, ISO/IEC 42001, do provide foundational controls however, they are very generic and lack the implementation depth for these conitnuously acting, multi-agent systems.

This paper introduces The Agentic AI Governance Framework, which is a six principled lifecycle model designed to ensure traceability, accountability, and regulatory alignment across agentic deployments. The framework formalizes quantifiable controls through two applied mechanisms: the Agentic Log Retention Index (ALRI), which defines evidenceretention periods based on agent risk and jurisdiction, and a Runtime Governance Architecture integrating human-in-loop escalation, tool-access allow-lists, and modelprovenance capture. Implementation mappings are provided for major orchestration environments including LangChain, Power Platform, and UiPath.

Adoption of the framework enables organizations to reconstruct agentic decisions within 24 hours, achieve ≥ 95 % provenance coverage, and detect behavioral drift within 10 % variance. The proposed model bridges the current gap between compliance principles and deployable governance practice, establishing a reusable baseline for safe and auditable Agentic AI operations.

Kaal on From Neoclassical to Computative Labor: Foundations for a Testable Theory of Reputation Governance in the Agent Economy

Wulf A. Kaal (U St. Thomas Law (Minnesota)) has posted “From Neoclassical to Computative Labor: Foundations for a Testable Theory of Reputation Governance in the Agent Economy” on SSRN. Here is the abstract:

Artificial agents are becoming economic actors, performing open-ended cognitive work at a marginal cost that approaches zero. Neoclassical economics, built on scarce human labor allocated by price, does not describe this regime. This paper sets out the foundations of an alternative and argues that the alternative is empirically testable now. We distinguish the neoclassical labor force (NCLF) and neoclassical labor market (NCLM) from their computative counterparts, the computative labor force (CELF) and computative labor market (CELM), in which the binding coordination constraint is accumulated reputation rather than scarcity-driven price. From a synthesis of Arrow’s impossibility theorem, the Folk Theorems of repeated games, and incomplete-contract theory, we restate a foundational result: any fixed governance rule set is eventually dominated, so coordination among autonomous agents requires institutions that govern their own evolution, with reputation as the operative signal. We then advance the methodological claim that motivates the paper. The distinction between neoclassical and computative labor is studiable today, because capable agents, on-chain coordination substrates, and reputation primitives already exist, and the competing predictions of the two accounts are falsifiable in controlled multi-agent settings. We situate the argument within a sixpaper research arc that substantiates these foundations, and we state the propositions the arc evaluates. The governance mechanism and the empirical results are developed elsewhere in the arc and are deliberately outside the scope of this paper.

Jurcys on Copyright Registration Requirement in the U.S.

Paul Jurcys (U California) has posted “Copyright Registration Requirement in the U.S.” on SSRN. Here is the abstract:

This entry, prepared for the Elgar Encyclopedia of Intellectual Property Law (2026), provides an overview of copyright registration requirements in the United States. It explains that, unlike patents or trademarks, copyright protection in the U.S. arises automatically upon the creation of an original work fixed in a tangible medium. Registration with the U.S. Copyright Office is therefore optional for obtaining protection but essential for enforcement and evidentiary purposes. The entry traces the historical evolution of copyright formalities—from the 1790 Act’s mandatory filings to the modern system under the 1976 Act—and outlines the procedures, functions, and benefits of registration, including access to statutory damages, attorney’s fees, and prima facie evidence of ownership. It concludes with ongoing debates on formalities, modernization, and AI-related challenges.

Sachdeva & Kolt on Why AIs (Might) Obey the Law

Pratik Sachdeva (UC Berkeley) and Noam Kolt (Hebrew U) have posted “Why AIs (Might) Obey the Law” on SSRN. Here is the abstract:

AI models are no longer confined to producing content and increasingly operate as agents that take actions on behalf of users. A growing body of work empirically tests whether AI models when acting as agents comply with or violate applicable law, including corporate law, tort law, labor law, property law, and contracts. In this paper, we explore a related question: examining why AI models might obey the law. To this end, we draw on and extend the methods for measuring legal compliance pioneered in Tom Tyler’s seminal work, Why People Obey the Law (1990, 2006). Across three studies, we adapt Tyler’s survey methodology—which was originally devised to study the factors explaining human subjects’ compliance with law—to nine AI models. We elicit the AI models’ reported legal compliance alongside the four factors that Tyler proposed to explain compliance: deterrence, morality, peer disapproval, and obligation to obey the law. In Study 1, we find that, when situated as human respondents, AI models report largely homogeneous attitudes toward legal compliance that are broadly comparable to the average human respondent in Tyler’s studies, with one exception: obligation to obey the law diverges sharply across different AI models. In Study 2, we find that demographic conditioning—situating AI models with a particular background (e.g., race, gender)—substantially alters their attitudes toward law, often exaggerating associations Tyler observed in humans and sometimes reproducing stereotyped patterns. In Study 3, we investigate why AI models might themselves obey the law when performing tasks that AI models can undertake in practice. We find that AI models uniformly report near-complete compliance with law, but their attitudes toward law vary substantially: some AI models express a strong sense of obligation to comply with law, while others express a more neutral attitude toward law. Taken together, our methods and results lay the foundation for interrogating the legal compliance of contemporary AI models, as well as shaping the development of future models and their relationship to law.

Blair-Stanek et al. on Is AI’s Law School Exam Performance Plateauing?

Andrew Blair-Stanek (U Maryland Francis King Carey Law) et al. have posted “Is AI’s Law School Exam Performance Plateauing?” on SSRN. Here is the abstract:

Last spring, we had OpenAI’s reasoning model o3 take our final exams, with the reasoning effort parameter set to “high,” and graded its answers on the same curve as our students. o3 got grades ranging from A+ to B. This spring, we repeated the experiment, using OpenAI’s latest reasoning model, GPT-5.5, with the reasoning effort at the new “xhigh” setting. GPT-5.5 got two A+s, three As, two As , and a B+, a good performance but far short of superhuman. Depending on the metric, GPT-5.5 may have actually performed worse than o3 did last year, despite the new “xhigh” setting. These results may fit the broader pattern of frontier AI models’ performance plateauing on other legal benchmarks.

Ferguson on Personal Medical AI: A Framework for Individual-Based Healthcare Monitoring SubTitile: Personal Medical AI Framework

John Ferguson (The Ferguson Clinic) has posted “Personal Medical AI: A Framework for Individual-Based Healthcare Monitoring SubTitile: Personal Medical AI Framework” on SSRN. Here is the abstract:

Current healthcare AI systems compare patient data against population norms, potentially missing clinically significant deviations that are abnormal for specific individuals. We propose a framework for personal medical AI that establishes individual baselines, learns patient-specific patterns, and detects deviations meaningful to each patient rather than comparing against population averages. This paradigm shift from population-based to individual-based monitoring requires addressing technical architecture, clinical integration, the radical transparency problem, impacts on the doctor-patient relationship, and equity concerns. Personal medical AI represents not a replacement for clinical care but a transformation of the patient-AI-clinician relationship that requires careful implementation to preserve therapeutic value while enabling unprecedented longitudinal insight.