Veale et al. on The Obligations of Providers of General-Purpose AI Models

Michael Veale (U College London Laws) and João Pedro Quintais (U Amsterdam Institute Information Law (IViR)) have posted “The Obligations of Providers of General-Purpose AI Models” on SSRN. Here is the abstract:

During the legislative process, the EU Artificial Intelligence (AI) Act was amended to include provisions related to general-purpose AI (GPAI) models. These broadly relate to transparency towards downstream users and relevant regulators, in addition to obligations connected to intellectual property. In this paper, we provide detailed analysis of these new provisions in the context of current technological applications and emerging trajectories, connecting them to computing literature and practice, and the broader context of connected and adjacent legal regimes, in particular copyright and relevant emerging case law. We find that there are a significant number of inclarities, tensions and contradictions both within the text, between the text and other legal regimes, and between the text and guideline documents, such as the Code of Practice on General-Purpose AI and recent guidelines by the European Commission. We identify a range of issues with the scoping of the provisions which may undermine its policy goals and create loopholes for regulatory avoidance, such as those relating to non-commercial models, open-source models, and model finetuning along the value chain. We find that the Code of Practice contains significant omissions and misstatements, some of which may present a compliance risk for an entity choosing to rely on the Code. We do not consider the provisions on GPAI models which present a systemic risk, which are dealt with elsewhere in the volume which this work will form a part of.

Shahidullah et al. on Intelligent Fraud Detection: Applying Advanced Analytics and Cybersecurity Insights in U.S. Finance

Mohammad Shahidullah (International American U (IAU)) et al. have posted “Intelligent Fraud Detection: Applying Advanced Analytics and Cybersecurity Insights in U.S. Finance” (Journal of Posthumanism, volume 4, issue 3, 2024 [10.63332/joph.v4i3.3593]) on SSRN. Here is the abstract:

Fraud detection in financial transactions is a major and crucial problem that does not cease to exist, mainly because of the enormous imbalance in the datasets obtained and the very high requirement for an accurate distinction between legitimate and fraudulent activities. In the following study, we assess the performance of three common machine learning models: Logistic Regression, Random Forest, and Gradient Boosting, for the detection of fraud, using a real-data set of transactions (284807 of which only 0.173% are labelled as fraudulent). The models were thoroughly evaluated with respect to critical metrics of performance including the precision, recall, F1-score and Area Under the Receiver Operating Characteristic Curve (AUC) to try to understand which of the models may be appropriate for dealing with class imbalance and false positives. Of the analyzed models, Random Forest was the best, with AUC being equal to 0.98, being superior to Logistic Regression (AUC = 0.97) and equal to Gradient Boosting (AUC = 0.98), while enabling more superior recall (0.88) and precision (0.44). This implies a higher capacity of detecting fraud cases without compromising the rate of false alarm too much. Feature importance analysis further noted that V14, V10, and V4 features were most predictive and most responsible for model classifying accuracy. Furthermore, calibration analysis revealed that Random Forest was the most reliable in estimating probabilities, outputs closely conformed to the ideal calibration curve implicating better reliability in practical applications. These findings suggest the effectiveness of the ensemble machine learning models especially Random Forest in promoting the efficacy of fraud detection systems. The study supports future research on real-time deployment and integration with deep learning methods to enhance the strength of fraud detection in the constantly changing financial spaces.

Schwarcz on Distributing Risk in an Age of AI: Procedural Bad Faith and AI Claims Handling

Daniel Schwarcz (U Minnesota Law) has posted “Distributing Risk in an Age of AI: Procedural Bad Faith and AI Claims Handling” on SSRN. Here is the abstract:

Written in honor of Kenneth Abraham and his foundational contributions to insurance law, this Essay argues that the rise of AI-driven insurance claims handling exposes a significant gap in first-party bad faith law. Traditional bad faith doctrine has focused primarily on outcomes, asking whether an insurer wrongfully denied or delayed payment of benefits owed under the policy. But increasingly automated claims processes create a distinct procedural injury when insurers deny, reduce, or delay claims without meaningful human review, adequate explanation, or a genuine opportunity for the insured to be heard. Drawing on procedural justice theory, the Essay shows that such practices can undermine voice, dignity, neutrality, and trustworthiness in a relationship defined by vulnerability and dependence. It therefore argues that courts should give procedural fairness substantially greater weight within the bad faith inquiry and should treat heavily automated claims denials without meaningful human oversight as powerful evidence of procedural bad faith. Doing so would adapt bad faith law to the distinctive risks posed by AI while preserving insurance’s core promise of fair, respectful, and accountable claims resolution.

Pollanen on The Probabilistic Foundations of Surveillance Failure: From False Alerts to Structural Bias

Marco Pollanen (Trent U) has posted “The Probabilistic Foundations of Surveillance Failure: From False Alerts to Structural Bias” on SSRN. Here is the abstract:

For decades, forensic statisticians have debated whether searching large DNA databases undermines the evidential value of a match. Modern surveillance faces an exponentially harder problem: screening populations across thousands of attributes using threshold rules rather than exact matching. Intuition suggests that requiring many coincidental matches should make false alerts astronomically unlikely. This intuition fails. Consider a system that monitors 1,000 attributes, each with a 0.5 percent innocent match rate. Matching 15 pre-specified attributes has probability 10^(-35), one in 30 decillion, effectively impossible. But operational systems require no such specificity. They might flag anyone who matches any 15 of the 1,000. In a city of one million innocent people, this produces about 226 false alerts. A seemingly impossible event becomes all but guaranteed. This is not an implementation flaw but a mathematical consequence of high-dimensional screening. We identify fundamental probabilistic limits on screening reliability. Systems undergo sharp transitions from reliable to unreliable with small increases in data scale, a fragility worsened by data growth and correlations. As data accumulate and correlation collapses effective dimensionality, systems enter regimes where alerts lose evidential value even when individual coincidences remain vanishingly rare. This framework reframes the DNA database controversy as a shift between operational regimes. Unequal surveillance exposures magnify failure, making “structural bias” mathematically inevitable. These limits are structural: beyond a critical scale, failure cannot be prevented through threshold adjustment or algorithmic refinement.

Lubin on Out of Time: Artificial Intelligence, Self-Help, and International Law’s Temporal Logic

Asaf Lubin (Indiana U Maurer Law) has posted “Out of Time: Artificial Intelligence, Self-Help, and International Law’s Temporal Logic” (The Cambridge Handbook of Public Law and Artificial Intelligence, Gregory, Williams, & Zerilli eds., Cambridge Univ. Press, forthcoming 2027.) on SSRN. Here is the abstract:

The rule of law has a tempo. It unfolds in time, organizes through time, and relies on time as a condition of its own functioning. Legality is sustained through an iterative practice of claim and counterclaim, argumentation and contestation. That exchange has a temporal structure. Time is needed for any claim to be formulated, any premise to be tested, any objection to be heard, and any judgment to be revised. In this sense, law is more than just a system of rules, institutions, and sanctions. It is also a system of pauses and intervals: the space between power’s first impulse and law’s second thought.

Law’s discursive and temporal character is most acute in international law. In a decentralized system lacking compulsory jurisdiction, states make law meaningful by arguing through it. Nowhere is this more evident than in the international law of self-help. The doctrines of self-defense, countermeasures, and retorsions govern the circumstances in which each state may invoke its own interests as grounds for unilateral protective action. The risk, of course, is that self-help will become self-licensing; that states, moving first, moving alone, and moving in secret, will be incentivized toward abuse and vigilantism. International law responds to this concern by forcing structured moments of argumentation around unilateral uses of power. The necessity, imminence, and proportionality of a use of force must be assessed and defended. The attribution of a prior wrong, together with the notice, purpose, and reversibility of a countermeasure, must likewise be tested and justified. These requirements are what prevent the law of self-help from becoming a vocabulary for disobedience.

And yes, these assessments also require time. Lots of time. Time for inquiry and for dissent. Time for simmering and for digestion. Time for contemplation and for meditation. Time for sleep, even. Indeed, some of the world’s greatest inventors and creators credit their creativity to the power nap. The sleeping mind consolidates, reconsiders, and sometimes resolves what the waking mind cannot.

But AI does not sleep. AI is built to increase efficiency through automation. It promises to offload cognition onto the machine, thereby removing the supposed waste and cost of “human friction.” AI does not deliberate, simmer, or meditate. It does not toss and turn at night. It does not wake at three in the morning with a reconsidered judgment. It does not pray or consult a loved one. It only processes. And it processes fast. Extraordinarily fast. So fast, in fact, that it outpaces law’s intervals. As this book chapter reveals, the very features that make AI attractive to national security decision-makers—its speed, scale, and tirelessness—are precisely the features that make it dangerous as a substitute for legal judgment. The chapter therefore asks not whether AI must be kept outside the national security state, for it will not be, but whether the state can use it without surrendering the pauses, frictions, and second thoughts through which law makes power answer.

Chouldechova et al. on Race-Conscious Admissions Algorithms and the Law

Alexandra Chouldechova (Carnegie Mellon U H. John Heinz III Public Policy and Management) and Daniel J. Hemel (New York U Law) have posted “Race-Conscious Admissions Algorithms and the Law” on SSRN. Here is the abstract:

In 2023, the U.S. Supreme Court held in Students for Fair Admissions v. Harvard that higher education institutions cannot admit students “on the basis of race.” This article addresses what it means for an admissions algorithm to operate on the basis of race. We develop a taxonomy of race consciousness in the algorithmic decision making context that provides lawyers and machine learning researchers with a shared vocabulary for exploring the implications of the Court’s ruling. We distinguish between “first-order” and “second-order” race consciousness at both the training and predictive phases of machine learning, and we argue that each category of race consciousness raises distinct legal and normative issues. We go on to explain why the Court’s decision need not be read as a flat-out ban on all types of race consciousness in admissions, and why certain forms of race consciousness might even advance the goals of justices who voted to strike down affirmative action policies inSFFA.

Bronsther on When the Cheapest Cost Avoider Is the Machine: Direct Sanctions for Autonomous AI

Jacob Bronsther (Michigan State U College Law) has posted “When the Cheapest Cost Avoider Is the Machine: Direct Sanctions for Autonomous AI” on SSRN. Here is the abstract:

The scholarship on artificial intelligence and legal liability assumes that the cheapest cost avoider is always a human being: a designer, deployer, or user. This Article identifies the conditions under which that premise fails. As an AI system’s behavior becomes less observable to its developer and more autonomous from human direction, the system may become the actor best positioned to foresee and forestall harmful outcomes. When such a system is also sensitive to the threat of legal penalties, the economic logic of AI-liability theory requires sanctions to reach the system itself. To the extent the system is judgment-proof, those sanctions must take nonmonetary form: limits on the computational resources it can use, the capabilities it can exercise, or its continued operation, calibrated to the severity of harm and the difficulty of detection. Human liability remains for the upstream risks humans could efficiently prevent; direct sanctions apply only to the residual conduct-level choices the system is best positioned to control.

Ard et al. on Technology Law Chapter 6: Upset Equilibria

Bj Ard (U Wisconsin Law) and Rebecca Crootof (U Richmond Law) have posted “Technology Law Chapter 6: Upset Equilibria” on SSRN. Here is the abstract:

Based on years of experience teaching the subject, we have produced a first draft of a “Technology Law” coursebook. It teases out fundamental concepts, introduces our methodology for resolving tech-fostered legal uncertainties, and identifies the strengths and weaknesses of different regulatory choices. Through a mixture of readings, exercises, and discussion questions, it helps readers develop facility in:

– Recognizing the recurring techlaw and policy questions and discerning the application, normative, and institutional uncertainties associated with a particular technology;

– Working through the process of resolving legal uncertainties, which includes consciously selecting a regulatory approach, identifying legally salient characteristics and relevant analogies, and weighing the benefits and drawbacks of various regulatory choices (law-by-analogy, creating new law, or reconfiguring legal institutions); and

– Developing familiarity with employing and countering common rhetorical strategies for advancing, opposing, or shaping regulation.

This course is designed to be accessible and useful to all students, regardless of career interests or prior experience with technology. New technologies challenge every area of the law, and the regulatory and rhetorical strategies we’ll explore are transferable across subjects.

This posting includes Chapter Six: Upset Equilibria. Future chapters will be posted bi-monthly.

We welcome feedback at the link included in the document; additional chapters will be updated regularly. If you are interested in teaching from this text, in whole or in part, please let us know, as we would be happy to share our class notes and slides.

Yu on Toward a New Research Agenda for Artificial Intelligence and International Law

Peter K. Yu (Texas A&M U Law) has posted “Toward a New Research Agenda for Artificial Intelligence and International Law” (UCLA Journal of International Law & Foreign Affairs, Vol. 30, 2026, Forthcoming) on SSRN. Here is the abstract:

Over the past few years, artificial intelligence—in particular, generative AI—and the changes brought about by this new technology have garnered priority attention from policymakers, commentators, and the mass media. Issues such as job displacement, the unauthorized use of personal information and likeness, the protection of intellectual property rights, and environmental sustainability have attracted analyses through an AI lens. What is less explored, however, is how AI will change the future development of international law. 

To provide an overview of the issues that have received attention thus far and that will continue to emerge at the intersection of AI and international law, this short essay, which was expanded from the opening remarks on the “Artificial Intelligence and the Boundaries of International Law” panel at International Law Weekend—West 2026, identifies eight issues that deserve greater policy and scholarly engagement. Together, these issues help demonstrate the need for a new research agenda for AI and international law and provide content to fill this agenda.

Solow-Niederman on Clickwrap Accountability

Alicia Solow-Niederman (George Washington U Law) has posted “Clickwrap Accountability” (95 Fordham L. Rev. (forthcoming 2026)) on SSRN. Here is the abstract:

Picture a public-facing generative AI chatbot on a government website that can answer questions about benefits eligibility.  Systems like these expose a sociotechnical-legal divide: they route around the standard due process framework, because there has been no formal government determination, yet they induce reliance from users, above and beyond past rounds of automated legal guidance.  Moreover, if there is an error, current doctrine on government errors and equitable estoppel makes relief unlikely.