Philipp Hacker (European U Viadrina Frankfurt (Oder) European New Digital Studies) and Matthias Holweg (U Oxford Said Business) have posted “The Regulation of Fine-Tuning: Federated Compliance for Modified General-Purpose AI Models” on SSRN. Here is the abstract:

This paper addresses the regulatory and liability implications of modifying general-purpose AI (GPAI) models under the EU AI Act and related legal frameworks. We make five principal contributions to this debate. First, the analysis maps the spectrum of technical modifications to GPAI models and proposes a detailed taxonomy of these interventions and their associated compliance burdens. Second, the discussion clarifies when exactly a modifying entity qualifies as a GPAI provider under the AI Act, which significantly alters the compliance mandate. Third, we develop a novel, hybrid legal test to distinguish substantial from insubstantial modifications that combines a compute-based threshold with consequence scanning to assess the introduction or amplification of risk. Fourth, the paper examines liability under the revised Product Liability Directive (PLD) and tort law, arguing that entities substantially modifying GPAI models become “manufacturers” under the PLD and may face liability for defects. The paper aligns the concept of “substantial modification” across both regimes for legal coherence and argues for a one-to-one mapping between “new provider” (AI Act) and “new manufacturer” (PLD). Fifth, the recommendations offer concrete governance strategies for policymakers and managers that propose a federated compliance structure, based on joint testing of base and modified models, implementation of Failure Mode and Effects Analysis and consequence scanning, a new database for GPAI models and modifications, robust documentation, and adherence to voluntary codes of practice. The framework also proposes simplified compliance options for SMEs while maintaining their liability obligations. Overall, the paper aims to map out a proportionate and risk-sensitive regulatory framework for modified GPAI models that integrates technical, legal, and wider societal considerations.

Raffaele Ciriello (U Sydney) and Angelina Chen (The U Sydney) have posted “Compassionate AI Design, Governance, and Use” on SSRN. Here is the abstract:

The rapid rise of generative AI reshapes society, transforming jobs, relationships, and core beliefs about human essence. AI’s ability to simulate empathy, once considered uniquely human, offers promise in industries from marketing to healthcare but also risks exploiting emotional vulnerabilities, fostering dependency, and compromising privacy. These risks are particularly acute with AI companion chatbots, which mimic emotional speech but may erode genuine human connections. Rooted in Schopenhauer’s compassionate imperative, we present a novel framework for compassionate AI design, governance, and use, emphasizing equitable distribution of AI’s benefits and burdens based on stakeholder vulnerability. We advocate for responsible AI development that prioritizes empathy, dignity, and human flourishing.

Parth Nobel (Stanford U) et al. have posted “Untangling AI Openness” (2026 Wisconsin Law Review (forthcoming)) on SSRN. Here is the abstract:

The debate over AI openness—whether to make components of an artificial intelligence system available for public inspection and modification—forces policymakers to balance innovation, democratized access, safety and national security. By inviting startups and researchers into the fold, it enables independent oversight and inclusive collaboration. But technology giants can also use it to entrench their own power, while adversaries can use it to shortcut years and billions of dollars in building systems, like China’s Deepseek-R1, that rival our own. How we govern AI openness today will shape the future of AI and America’s role in it.

Policymakers and scholars grasp the stakes of AI openness, but the debate is trapped in a flawed premise: that AI is either “open” and “closed.” This dangerous oversimplification—inherited from the world of open source software—belies the complex calculus at the heart of AI openness. Unlike traditional software, AI is a composite technology built on a stack of discrete components—from compute to labor—controlled by different stakeholders with competing interests. Each component’s openness is neither a binary choice nor inherently desirable. Effective governance demands a nuanced understanding of how the relative openness of each component serves some goals while undermining others. Only then can we determine the trade-offs we are willing to make and how we hope to achieve them.

This Article aims to equip policymakers with the analytical toolkit to do just that. First, it introduces a novel taxonomy of “differential openness,” untangling AI into its constituent components and illustrating how each one has its own spectrum of openness. Second, it uses this taxonomy to systematically analyze how each component’s relative openness necessitates intricate trade-offs both within and between policy goals. Third, it operationalizes these insights by advancing a research agenda that shows how law can be analyzed and refined to support more precise configurations of component openness.

AI openness is neither all or nothing nor inherently good or evil—it is a tool that must be wielded with precision if it has any hope of serving the public interest.

Yiyang Mei (Emory U) and Michael J. Broyde (Emory U Law) have posted “Reclaiming Constitutional Authority of Algorithmic Power” on SSRN. Here is the abstract:

Whether and how to govern AI is no longer a question of technical regulation. It is a question of constitutional authority. Across jurisdictions, algorithmic systems now perform functions once reserved to public institutions: allocating welfare, determining legal status, mediating access to housing, employment, and healthcare. These are not merely administrative operations. They are acts of rule. Yet the dominant models of AI governance fail to confront this reality. The European approach centers on rights-based oversight, presenting its regulatory framework as a principled defense of human dignity. The American model relies on decentralized experimentation, treating fragmentation as a proxy for democratic legitimacy. Both, in different ways, evade the structural question: who authorizes algorithmic power, through what institutions, and on what terms. This Article offers an alternative. Drawing from early modern Reformed political thought, it reconstructs a constitutional framework grounded in covenantal authority and the right of lawful resistance. It argues that algorithmic governance must rest on three principles. First, that all public power must be lawfully delegated through participatory authorization. Second, that authority must be structured across representative communities with the standing to consent, contest, or refuse. Third, that individuals retain a constitutional right to resist systems that impose orthodoxy or erode the domain of conscience. These principles are then operationalized through doctrinal analysis of federalism, nondelegation, compelled speech, and structural accountability. On this view, the legitimacy of algorithmic governance turns not on procedural safeguards or policy design, but on whether it reflects a constitutional order in which power is authorized by the governed, constrained by law, and answerable to those it affects.

Chee Hae Chung (Purdue U) and Daniel Schiff (Purdue U) have posted “AI and the Social Contract” (Proceedings of the Seventh AAAI/ACM Conference on AI, Ethics, and Society (AIES-25)) on SSRN. Here is the abstract:

As artificial intelligence (AI) systems increasingly shape public governance, they challenge foundational principles of political legitimacy. This paper evaluates AI governance against five canonical social contract theories—Hobbes, Locke, Rousseau, Rawls, and Nozick—while examining how structural features of AI strain these theories’ durability. Using a structured comparative framework, the study applies three forms of legitimacy (procedural, moral-substantive, and recognitional) and three types of consent (explicit, tacit, and hypothetical) as normative benchmarks. Applying each theory, the analysis finds AI governance is marked by deficits in accountability, participation, rights protection, fairness, and freedom from coercion, while AI’s opacity, global influence, and hybrid public-private control reveal blind spots within the social contract tradition itself. Though no single theory offers a complete solution and each contains specific weaknesses, the paper develops a hybrid model integrating Hobbesian accountability, Lockean rights protections, Rousseauian participation norms, Rawlsian fairness, and Nozickian safeguards against coercion. The paper concludes by distilling normative priorities for aligning governance with these hybrid contractarian standards: embedding participatory mechanisms, encouraging pluralistic ethical perspectives, ensuring institutional transparency, and strengthening democratic oversight. These interventions aim to reconfigure the social contract—and AI—for an era in which algorithmic systems increasingly mediate the exercise of political authority.