Day 1 (10 December 2026)
Plenary Session 1 - The Big Picture
The rapid advancement of artificial intelligence (AI), coupled with its inherent risks, necessitates the establishment of a robust global AI governance framework. Distinguished by its unique characteristics including rapid iterative development, black-box operational opacity, cross-sectoral scalability, and cascading systemic vulnerabilities, AI exacerbates the Collingridge Dilemma, a foundational paradox governing the regulation of emerging technologies, which presents a two-pronged governance impasse.
The Collingridge dilemma encapsulates the core challenge of imposing effective social control over technological innovation, and offers a cogent analytical framework for interpreting the regulatory predicament attendant to AI development. In the nascent stages of AI evolution, its societal ramifications are notoriously unforeseeable; yet once the technology matures and becomes deeply entrenched within societal infrastructures, its negative externalities evolve into increasingly intractable barriers to regulatory oversight and governance intervention.
Conventional doctrinal analyses predominantly situate the Collingridge Dilemma within the technical domain, emphasizing the inherent uncertainty and limited controllability of emerging technologies. However, this perspective is analytically incomplete. The dilemma is not only anchored in technological indeterminacy, but also is perpetually replicated and even amplified at the policy level by cognitive biases and strategic conduct among sovereign states. This phenomenon manifests principally in the goal distortion inherent to national AI legislative frameworks, thereby engendering novel risks to effective social control over AI systems. Accordingly, the Collingridge Dilemma manifests along dual dimensions: a technological dimension and a policy-induced dimension.
In recent years, states have implemented a litany of interventionist measures under the expansive guise of national security—a rubric that now encompasses supply chain resilience, data governance, technological risk mitigation, and semiconductor security. Such measures include export control regimes, foreign investment screening, industrial subsidization schemes, data security reviews and economic sanctions. While formally predicated on risk prevention and mitigation, these policies have in practice fostered to a de facto decoupling across trade, investment, financial and data governance regimes. This policy trajectory not only ex ante constrains the trajectory of technological innovation but also ex post distorts cross-jurisdictional regulatory learning, thereby exacerbating the Collingridge Dilemma along its policy dimension.
A core catalyst underpinning this policy-layered dilemma resides in the inherent subjectivity of risk and security assessments. Existing international law imposes only narrow constraints on the manner in which sovereign states define and operationalize these malleable concepts. Consequently, decision-making is frequently driven by cognitive biases and strategic competitive imperatives rather than rigorous, objective risk evaluation. In pursuit of maximizing relative strategic and economic gains, states tend to adopt zero-sum approaches to AI governance, which risk precipitating collective irrationality and systemic institutional fragmentation at the global level.
The international legal architecture governing AI remains deeply fragmented, with major powers engaged in competitive rather than collaborative rule-making endeavors. At the international level, AI governance is characterized by greater strategic competition than multilateral cooperation, deeper institutional fragmentation than normative integration, and more pronounced soft‑law attributes than hard‑law binding obligations. The evolution of international law applicable to AI has reached a regulatory impasse, standing at a decisive normative juncture. Major powers leverage domestic legislative instruments to entrench their technological superiority, curtail the technological advancement of geopolitical competitors, and further widen pre-existing digital and AI governance divides.
In particular, the United States and the European Union, drawing on their dominant market and technological leverage, are competing for normative primacy in the formulation of global AI rules and standards. Through the extraterritorial spillover effects of their domestic legal regimes, these actors shape the regulatory models and normative trajectories of AI governance adopted by third states. They further embed human rights, democratic governance and allied normative priorities into the global AI rule-making process, ostensibly to preserve their technological hegemony and supply chain security. This dynamic has contributed to the escalating securitization and instrumental weaponization of AI-related regulatory norms.
To redress the policy-induced amplification of the Collingridge Dilemma, this study advances a three-pronged remedial framework. First, the international legal regime governing AI must be strengthened to curb discretionary and unilateral state conduct that undermines multilateral governance. Second, the cultivation of transnational technical epistemic communities can foster normative coordination and mitigate security-driven policy divergence across jurisdictions. Third, sovereign states should develop proactive rule-making capacity to shape emerging global AI norms, thereby securing a more equitable and influential role in the construction of inclusive global AI governance.
AI agents represent a promising development in the trajectory of the progress of artificial intelligence. AI agents are able to do anything humans can do with a computer. Seen this way, AI agents can manage complicated tasks, including making autonomous choices and selections in order to follow through with its explicit or implicit goals. It can thus impact not only online environments but also offline ones, as individual and societal orderings are increasingly mediated through artificial intelligence and computational systems in general. As capable AI agents are being developed and deployed in real-world environments, the question of how to govern or regulate these agents becomes an ever more urgent inquiry.
The question of rights for artificial agents - machines or robots, has been the subject of scholarly investigation for some time. However, the question of rights for AI agents, with agency understood through the paradigmatic change introduced by generative AI, remains under-examined. This article aims to fill this gap - addressing the inquiry from the perspective of considering whether AI agents can have rights.
We examine four possible scenarios where this can happen. First, the derivative argument lifts the overarching legal and moral grounding of agency law and tests its applicability for AI agents. Second, the instrumental argument holds that AI agents are and increasingly will become more embedded in our social world. Its instrumental presence, namely its capacities to take actions that affect both online and offline environments mean that there is a need to address how rights (and therein duties) can help to navigate around inevitable challenges of responsibility and liability. While the first two scenarios look backwards, drawing from applicable law and literature from social agents such as robots, the remaining scenarios look ahead. The third scenario adopts a philosophical perspective of considering AI agents as a distinct entity deserving of rights protection. We highlight the role AI agents can play in advancing overall well-being, including its potential for addressing collective action problems that require agreement and high levels of social coordination. The fourth scenario looks into rights for AI agents as a means to counter the centralisation of (corporate) power.
With these scenarios, we aim to better inform academic and policy discussion around AI agents, including by decoupling the question of rights from considerations of (moral) personhood and its associated qualities, such as sentience and consciousness. We also address objections on how rights for AI agents could contribute towards the dilution of rights on the one hand, and its proliferation on the other. We also consider a further objection that rights are unnecessary because AI systems could simply be subjected to legal duties without being recognized as rights-holders. As AI agents increasingly feature in our social, professional and political lives, questions on rights for such agents will inevitably arise. This article contributes towards providing a more nuanced account of whether and how AI agents can have rights.
Artificial Intelligence (AI) is reshaping societies and economies, and in the absence of a unified global governance framework, regulations surrounding the governance of AI are siloed and fragmented. This undermines predictability in the governance, innovation and poses security challenges. Yet, a binding global treaty, as a unified framework, may be untenable given divergent national interests and technological capacities. The literature suggests that both imposition of goals does not produce greater participation whereas the absolute autonomy in setting the goals does not produce results either. Therefore, the paper proposes an alternative hybrid governance framework inspired by the Paris Agreement on climate change. The Paris Agreement balances state autonomy with international cooperation by combining voluntary commitments, peer accountability, and shared global goals. Drawing on this structure, the paper proposes a cooperative AI governance framework that would ensure equitable development and respect for national sovereignty.
The paper is based on doctrinal research on structure, strengths, and challenges with the Paris Agreement, and also looks into the literature around the existing frameworks on regulation of AI and the potential challenges in implementing the proposed framework for AI.
The paper presents the detrimental fragmentation and siloed development of global regulations surrounding AI. The hybrid structure of the Paris Agreement, legal character and form, and differentiation, with respect to its relevancy for the proposed AI governance framework is explained. The paper then critically evaluates the principal global AI governance models to identify the regulatory gaps that justify the proposed hybrid framework. Culminating the understanding of elements, an analysis is made whether the Paris Agreement model can be extrapolated to global AI governance. The analysis supports positive outlook provided certain steps such as phased implementation of the agreement, accommodations for developing nations, and technical assistance mechanisms are substantially considered. The potential challenges to such a governance framework is deliberated upon and a domain-specific multi-layered structure is proposed as a more practical as well as theoretically workable framework. Urging for a consensus-building exercise in the international community, the paper concludes that a domain-specific multi-layered structure, formed with an agreement similar to Paris Agreement, is a potential workable global AI governance framework which respects National sovereignty and equitable development.
Plenary Session 2 – AI, Human Decision-making and Legal Reasoning
The right to a human decision, enshrined in Art 22 GDPR, has a long pedigree - both as a constitutional right, and a technical requirement to ensure ‘humans in the loop’. But is it still fit for purpose in a world of high-dimensional automated decision-making? In this paper, I explore the different motivations for the right to a human decision, from ensuring accurate decisions to protecting individual dignity, and conclude that it has failed on all counts. What is required instead, I suggest, is a shift from the protection of human decisions to different mechanisms to foster human agency. This can no longer be achieved by a focus on humans in the loop alone: legal mechanisms need to provide additional and/or alternative mechanisms to frame and contest automated decision-making. Instead of a generalised right to humans in the loop, the article explores different ways of ensuring human involvement before, after, and above the loop.
Artificial intelligence (particularly generative AI and agentic AI) provides the opportunity to improve efficiency in the judicial system, and in reducing court costs, promote access to justice.
While the efficient resolution of disputes is a central function of the courts, courts are not just dispute resolvers. Courts have an institutional and constitutional role whereby they determine and apply the law in a fair manner, and in doing so uphold the rule of law.
For the courts to perform those roles there must be trust and confidence in the judiciary, which is achieved through both the just and efficient resolution of disputes. Just meaning both according to law, but also through a system that affords open justice, procedural fairness and impartiality.
There are significant concerns about the use of generative AI in courts, including wasting the courts’ resources, fairness concerns and undermining public trust in legal processes.
This paper examines how existing forms of governance applicable to the courts and the legal profession protect the administration of justice.
The paper is divided into two parts:
Protecting courts from within (including from government) such as when AI is suggested/mandated for use within courts. This part examines the nature of judicial power which must be exercised consistent with judicial process. This mandatory requirement provides guidance as to the standards AI would need to meet for use in the courts, or more likely the tasks within courts that AI use would be suitable for.
Protecting courts from without, such as when AI is adopted by court users in their interactions with the judicial system. Generative AI has been adopted by legal practitioners, parties, lay and expert witnesses. Court judgments have pointed to ‘hallucinations’ by generative AI tools that have led to court users referring to fictitious or ‘fake’ law. Concerns have also been raised about fictitious or ‘fake’ evidence (also referred to as ‘deep fakes’). This part examines the courts’ ability to protect their own processes, including through procedural and evidential requirements, sanctions, and lawyer’s ethical obligations.
As AI systems proliferate across high-stakes domains, courts increasingly face disputes involving opaque, adaptive technologies for which traditional liability concepts (such as fault, causation, foreseeability, and even product defect) lack workable evidentiary anchors. This paper argues that, in response, adjudication is “judicializing” data governance. Industry-led best practices in dataset documentation, provenance tracking, access controls, bias evaluation, and post-deployment monitoring are pulled into legal reasoning as practical benchmarks of reasonable care and product safety. Judicialization does not require courts to endorse any single technical standard. It is a process by which courts, confronted with informational asymmetry and the need to avoid arbitrary outcomes, translate governance practices into liability-relevant proxies for what a prudent actor would have recorded, tested, constrained, and reviewed at the data layer. The paper develops a theoretical account of this translation mechanism and its institutional consequences, showing how data governance can become the hinge between ex ante governance and ex post remedies, reallocating risk across developers, deployers, and integrators while reshaping incentives for responsible AI without bespoke AI tort doctrines.
Against the backdrop of growing deployment and experimentation with the use of LLMs in decision-making and courts worldwide, we explore an emerging mode of adjudication that we term the “simulated judgments” model - a form of AI-assisted opinion-drafting and adjudication which extends beyond administrative support or conventional legal research, in which LLMs contribute to the production of judicial opinions, while the human judge retains ultimate epistemic control and responsibility for the resulting decision. In this paper, we explore the objections and concerns regarding this emerging mode of AI use in adjudication - the technical, epistemic, procedural and legitimacy concerns - asking which of these concerns pose genuine limits to AI-assisted adjudication and which might instead be addressed through appropriate forms of human oversight, institutional design, and AI governance.
Plenary Session 3 – AI & Intellectual Property - Part 1
First articulated in influential work by Geoffrey Hinton, Oriol Vinyals, and Jeff Dean, knowledge distillation has become a widely adopted technique for training smaller, more efficient “student” models to emulate larger “teacher” models, often achieving performance close to far more computationally intensive systems. Their early results—and subsequent industrial practice—showed that distilling ensembles into a single model can deliver substantial efficiency gains and lower latency with minimal loss of accuracy. As a result, distillation is now a staple for model compression, deployment on resource-constrained devices, and transferring capabilities across domains.
Over the past two years, however, several U.S. AI companies (including OpenAI and Anthropic) and government officials have publicly alleged that certain Chinese AI firms—such as DeepSeek, Moonshot AI, and MiniMax—have engaged in large-scale misappropriation of proprietary model capabilities via knowledge distillation and related model-extraction techniques. Although the technique itself is general-purpose and widely used, the legal stakes turn on how the teacher model was accessed and what contractual or statutory limits govern that access. This paper examines whether, and under what circumstances, knowledge distillation could constitute copyright infringement, trade secret misappropriation, and breach of contract. It also considers how intellectual property law and policy should adapt as distillation becomes more prevalent and assesses the broader implications for the U.S.–China AI competition.
Copyright protects works recorded “by or under the authority of the author.” For the first fifty years of the 1976 Copyright Act, authority came from deliberate actions: authors directed newer and better machines to be their scribes. Whether it was pen to paper or wake word to Alexa, authors provided intentional direction to machines to record, or “fix” in copyright parlance, their ideas.
Ambient recording devices break that fixation model. When an author gives a public speech, it is now captured by smart glasses, AI recording devices, and always-on cameras operated by others. Device operators upload and exploit the work, not authors. Current doctrine rewards the operator: if an author did not intend to fix the work, it is not under their authority. The minimally creative operator gains protection for having an ambient device in the right place at the right time. And if the device operator is insufficiently creative, they still win: neither side has a copyrightable interest, but only the operator has the copy to exploit. They will upload the work and reap the profits from a brilliant public performance of another. This cannot be justified under any normative theory of copyright. Ambient devices create a constant authorship fixation gap.
This Article maps that authorship-fixation gap and proposes a new model for closing it: notice and adoption. Drawing from doctrines in cases, foreign jurisdictions, and digital platform infrastructure already operating at scale, the model requires those who would exploit ambiently fixed works to make efforts to notify the author. The author may then choose to adopt the fixation and claim copyright protection while potentially sharing rewards with the device operator. This is a natural extension of existing platform management of unauthorized works. In a world where ambient fixation is everywhere, copyright should stop rewarding whoever happens to operate devices and start rewarding those who create.
Integrated entertainment intersects at the narrative of technology law and cultural ethos. Its integration expands from streaming services to consoles to platforms or other immersive services that engages human creativity and ingenuity in evolving solutions to various challenges and needs. Gaming is a thriving industry, attracting huge investment and innovation. The online gaming market is expected to cross $9.2 billion by the year 2029. With the rise of gamer communities, gatherings, and festivals, gaming has evolved beyond an industry into a vibrant culture of creativity, competition, and knowledge exchange. Intellectual Property rights such as copyrights are important within creative sectors as it protects the right holders against infringement of their work and thus providing an incentive for their creativity. While copyrights have an important role in protecting creative works, its role in protection video games have been challenging as video games have not been recognized as a work under international and domestic copyright laws. With the technological advancements over the years and usage of Immersive technologies like Virtual Reality (VR) and Augmented Reality (AR) where VR headsets and Mobile AR games have changed the gaming experience of Players and equally the scope of Video Gaming industry. However, with the advent of synthetic intelligence, gamers are increasingly using generative AI to evolve assts such as a texture, voices, music etc. using various AI tools. In addition, new genera of games are beginning to evolve which are developed using procedural content generation such as no man’s sky and Dwarf fortress. This technique enables game developers to algorithmically generating game content, such as tasks or characters. The flexibility that procedural content generation offers to developers makes it more suitable to the needs of gaming community. While procedural content generation is inherently creative within game development, the unsettling aura surrounding generative AI works raises profound questions about whether copyright law can keep pace with the integrated entertainment ecosystems that is increasingly collaborative, algorithmic, and global in nature. More importantly questions of ownership, fair uses of synthetic works integrated into a complex gaming system. In addressing this the first part of the work examines the technological transformation and how copyright law had managed to keep pace with the change. The second part examines the comparative copyright law cases and their implications for generative AI based creations. More importantly it raises questions about player manipulation, AI autonomy, and digital rights and the final part tries to explore alternate means to address these challenges particular to video games and also offer sandbox environments for synthetic AI generation without risk of infringement. Acknowledging the role of intellectual property rights the work also finds a space for digital public domain in integrating technology with legal rights.
Plenary Session 4 – AI in Business: Competition, Employment and Corporate Liability
Dark patterns are deliberate design techniques for creating e-commerce environments inducing consumers into actions that benefit a business through some form of deception, coercion, manipulation or other undue influence. Fake urgency, guilt-tripping, visual-salience tricks are some typical examples. A large-scale ticketing system implementing such a scheme was urging consumers to obtain ticket insurance through terrorizing messages suggesting, among others, that consumers would be “responsible for whatever happens.” The context of the purchase also created the false impression that procuring ticket insurance is mandatory. Another example is when the process of taking an action that is detrimental to the economic interests of a business is made overly complicated. Finding the cancellation button to end a subscription may occasionally feel like an attempt to escape from a labyrinth.
Machine-learning models enable the creation of adaptive and personalised dark patterns increasing their perniciousness. Scarcity messages conveyed by a travel or retail platform may be algorithmically adjusted to target those consumers who most are susceptible to urgency-based pressure. AI sentiment models can analyse the consumers’ browsing behaviour and figure out their weak spots and moments. AI may also facilitate the design of the most labyrinthine menus for a given user.
The primary instrument for scrutinizing the legality of such practices under EU law is the Unfair Commercial Practices Directive (UCPD). The UCPD prohibits misleading as well aggressive practices and imposes obligations of due care with respect to commercial communications to consumers. Similar rules can be found in the UK Digital Markets, Competition and Consumers Act 2024. In the US, dark patterns have been deemed unlawful under section 5 of the FTC Act.
The approaches adopted in different states diverge significantly with respect to the enforcement of the relevant rules. In Germany, the system relies on private enforcement primarily driven by particularly strong trader and consumer associations. The Competition and Markets Authority (CMA) is the administrative body entrusted with the authority to enforce the relevant rules in the UK. Since 2022 the CMA has launched a dedicated enforcement programme on Online Choice Architecture. CMA’s wide administrative and investigative powers along with its specialized data analytics and technology unit offer particularly effective possibilities for deterring dark patterns. Notably, the CMA can be fast in reaching binding decisions, issue interim-stop orders, and impose fines. In the US, the Federal Trade Commission (FTC) constitutes a predominantly litigation-based enforcement authority, but which has some limited rule-making authority. As of current, the FTC is preparing the adoption of rules against manipulative AI-driven dark patterns.
What is the optimal enforcement structure for rules against unfair competition the primary purpose of which is to protect consumers against AI-driven unfair commercial practices?
The next generation of workplace AI may not merely replace employees; it may be trained to become them. Employers can now feed an employee’s emails, chats, documents, meeting transcripts, code reviews, and behavioral traces into AI systems that reproduce not simply discrete outputs, but working style, tacit judgment, institutional memory, and professional persona. This Article calls this practice employee distillation and cloning, and offers the first systematic legal account of this emerging practice that marks a new frontier in the propertization of labor.
Existing law appears, at first glance, to offer employees many possible avenues of protection: privacy and data protection law, copyright, trade secret law’s general skill, knowledge, and experience doctrine, the right of publicity, and contract law. Yet these protections largely dissolve upon closer inspection. Privacy law regulates only narrow forms of intrusive collection; data protection remains fragmented and uneven; copyright and trade secret laws often favor employers; right of publicity doctrine rarely reaches ordinary working identity; and contract law converts structural inequality into formal consent. The result, as this Article terms it, is a “rights mirage”: the appearance of legal protection without meaningful control.
This Article further argues that the mirage matters. Although unregulated distillation may generate short-term productivity gains, it may also impose costs that firms do not fully internalize. These costs include productivity losses and organizational friction resulting from employee resistance, degraded knowledge sharing, reduced human-capital investment, impaired mobility, weakened cumulative innovation, and distinct dignitary and autonomy harms. The Article therefore calls for future governance directed not at prohibiting employee distillation, but at structuring it through notice, meaningful consent or bargaining, use limits, compensation, and collective governance.
While corporate law has traditionally allowed a certain level of delegation, it has never had to address the problem of how to respond when the delegate isn’t a human being. Contemporary agentic artificial intelligence systems conduct negotiations, perform pre-employment screening, execute transactions, and distribute investment capital with very little human involvement in processes that are, by their very nature, human processes and fiduciary processes. If the system acts in the context of ordinary business activities, but violates the duties of care and loyalty owed by a human officer, there is no coherent legal theory for holding anyone accountable for the violation. This is not a purely theoretical problem for the future; discriminatory algorithms in recruitment, unauthorized use of algorithms in trading, and black box investment allocation mechanisms have already produced tangible harms that cannot be attributed to any human agent under current law. The thesis of this paper is that the corporate veil and the fiduciary framework that surrounds it provide the proper locus for addressing this problem, provided that courts and lawmakers are prepared to rethink the duties of director oversight rather than merely falling back on formalism or AI personhood.
Day 2 (11 December 2026)
Plenary Session 5 – Balancing Benefits & Risks: Setting Standards
As automation and AI reorganise social, political, and economic life, law’s reflex is often towards the human. By making human oversight a central feature of AI governance, policy makers hope to promote the safe realisation of automation’s economic and social promises. What exactly the human overseer is implemented to achieve, however, is less clear. Over time, the literature has described multiple goals such as system performance, imbuing systems with the normative value of human agency, and locating responsibility within automated workflows. But the efficacy and value of those rationales is difficult to demonstrate. In light of empirical studies consistently indicating that human oversight does not improve system performance, that human oversight cannot subordinate automation to human agency, and that human oversight improperly responsibilises humans, how it remains an entrenched regulatory orthodoxy is difficult to understand.
This paper suggests that the growing scholarly and policy literatures on human oversight makes understanding how human oversight interacts with those goals largely impossible. Empirical literatures attempt to draw generalisable findings about system performance from highly stylised experiments and simulations. Normative scholarship describes the value of human agency without any connection to real examples of automation. Studies that interface regulatory thinking with disciplines like human factors engineering and ergonomics become context agnostic, treat all examples of automation and human oversight as effectively equivalent, and propose that policy makers pursue sui generis strategies for whatever instance of automation is being governed. All three approaches treat human oversight as a generic and unitary concept, in ways that obscure its operational value and rationales.
To better understand the ways that human oversight interacts with the goals of efficacy, agency, and accountability, this paper rejects the genericisation and describes three different legal forms of human oversight, taxonomized by organisational and labour arrangements that elucidate the functions of human oversight in operation. These legal regimes alternatively demand a human overseer as a ‘final decision-maker’, as a ‘system operator’, and as a ‘supply chain input’, with each: producing different relationships between human overseers and the organisation that controls the means and purposes of automation; configuring human agency in different ways with different accountability consequences; and enabling different organisational forms, automation workflows, and labour relations for the strategic management of efficiency, controllability, and liability.
In defining these legal forms and offering case studies of how they are operationalised, the paper provides conceptual clarification of a complex and confused body of research that often treats human oversight as a generic and unitary regulatory strategy. And more importantly, it elucidates how regulation requiring human oversight functions, in different ways, as a managerial instrument for strategically coordinating competing organisational objectives, such as regulatory compliance, system efficiency, and the distribution of accountability and liability, though enabling discrete organisational and labour arrangements.
The governance of AI-enabled grey-zone operations suffers from a timing gap. Domestic liability regimes usually respond after deployment and harm, while international law often becomes practically actionable only when conduct crosses thresholds such as prohibited intervention or the use of force and can be attributed to a State. Yet AI-enabled influence operations, cyber operations, and autonomous surveillance may produce cumulative below-threshold effects that constrain another State’s choices without clearly triggering those rules. Existing cyber-law frameworks, including Tallinn Manual 2.0, help classify individual operations but do not fully address this form of cumulative coercion (Schmitt 2017, 30-38, 84-100, 312-338).
This paper shifts the inquiry from the legal classification of isolated conduct to the governance of the infrastructure through which high-risk AI capabilities are trained, scaled, accessed, and deployed. Compute providers and cloud service providers occupy a distinctive position because they can identify customers, preserve records, monitor certain uses, and restrict access in defined cases. Recent compute-governance literature therefore describes them as regulatory intermediaries and treats cloud infrastructure as a possible control point for upstream oversight (Egan and Heim 2023, 1-3, 7).
The paper develops an infrastructure-based due-diligence framework. It does not claim that private providers directly bear a general due-diligence obligation under international law. Rather, it asks how State due-diligence obligations, where applicable, may be operationalised through risk-based and proportionate regulation of infrastructure providers. The proposed framework calibrates regulatory duties according to the severity and likelihood of harm, knowledge of the risk, the degree of control over the relevant infrastructure, the means reasonably available, and the provider’s leverage over the customer or workload. Possible measures include customer identification, record preservation, reporting of defined high-risk activity, and targeted suspension or restriction of access (Ridings 2024, 146-147, 153; OECD 2023, 2, 44-46).
The central claim is modest. AI-enabled grey-zone operations should not be artificially reclassified as uses of force merely to close a governance gap. At the same time, they should not be left solely to ex post liability after harm has materialised. International due diligence provides the legal structure, while compute governance provides the operational tools, for a proportionate form of upstream risk management.
Plenary Session 6 – AI Use in Specific Scenarios: Is there a Common Thread?
While the full economic impact of Generative AI (GenAI) and Agentic AI has yet to unfold, their influence on the classroom is already profound. These models offer students unprecedented access to information and analysis, yet their impact on learning remains volatile. This essay argues that inappropriate reliance on AI creates a dangerous competency gap: a scenario where students graduate with advanced credentials but lack the foundational training to perform safely or effectively.
This essay explores the burgeoning liability landscape created by this shift. For educators and institutions, we examine the risk of regulatory scrutiny and common law claims—such as breach of contract or educational malpractice—when educational institutions produce AI-reliant but inadequately trained graduates. At the same time, graduates may incur personal and professional risks of passing off AI-generated competence that leads to findings of professional negligence. Analyzing current regulatory frameworks and the common law, this essay questions whether the legal conundrum facing universities and schools and provides a path forward.
AI promises to be one of humanity’s most consequential technologies. Much of the media attention has been focused on large language models, commonly referred to as chatbots, but more recently, it has turned to AI agents. While chatbots respond to prompts, AI agents can perform tasks autonomously, like a competent assistant. Companies claim that these AI agents can perform a wide range of functions and will make businesses and individuals more productive.
Yet, an important question remains: Who is liable for the acts of these agents?
AI’s vast potential and advanced capabilities have generated excitement but also trepidation among the public. Some industry insiders have even issued dire warnings of artificial intelligence systems running amok. Despite these concerns, companies that stand to profit from building AI continue to move at full speed.
Courts and legislators have yet to determine who will be responsible when AI agents act in a way that causes harm to their users or third parties. Polls have indicated that the public wants the government to regulate AI. However, it is probable that the most common way AI will be regulated in the immediate term is through unilaterally drafted terms of service.
The most dominant players in the tech industry wield fine print and click to agree terms to enact private laws and impose norms. The surveillance economy was enabled by terms of service because lawmakers and courts were cowed by fears of stifling innovation. These same corporations are now turning to terms of service to facilitate fast, widespread adoption of AI systems. These fine print terms reallocate rights from the user to the company and shift legal liability from the company to the user.
This Article proposes a framework for evaluating the enforceability of digital adhesive terms. This framework requires: first, categorizing terms depending on whether they assert the company’s property rights or diminish the other party’s existing rights; second, identifying the nature and cause of the harm; and third, applying property-asserting terms to user harms but applying strict liability to third party harms.
Companies should not be permitted to shape industry norms through fine print and escape liability for their offerings; however, they should be permitted to establish codes of conduct for the uses of their products and services. My proposal balances a company’s need to set codes of conduct and use for its products with society’s need to ensure corporate accountability and product safety for mass scale technology.
Criminal law has long confronted a responsibility gap in AI-mediated harm: AI systems learn, adapt, and act in ways no human actor can foresee or control closely enough for traditional attribution rules to apply (Matthias, 2004; Santoni de Sio & Mecacci, 2021). To resolve this gap, scholars have met a profound dilemma. Either they compromise the fundamental tenets of criminal law, or they accept that some AI harms fall outside its proper reach and surrender them to tort and regulation. The first horn takes three principal forms in the literature: attributing mens rea directly to the AI system (Hallevy, 2024), ascribing intentions through legal fiction and displacing subjective fault with objective standards of risk creation (Ayres & Balkin, 2024) or transposing the mens rea inquiry onto the deploying corporation (Weigend, 2024). The second horn surrenders the field. None of these positions, we submit, is wholly satisfactory. The existing literature, while closing the responsibility gap by locating a defendant, fails to close a distinct and deeper gap: the censure gap, the question of whether that defendant has the mental and moral status to be the proper subject of criminal moral judgment for that specific harm.
This contribution presents groundbreaking work on a novel ethical-legal framework for Artificial Intelligence systems (AI) for nonhuman animals. Existing AI law and ethics scholarship has focused largely on how AI impacts human beings. It is increasingly recognised that AI can also have a significant impact on environmental sustainability. In this paper we extend this nonhuman concern to how sentient animals are impacted by AI. AI technologies’ growing capacity to augment and automate human activities gives its creators and users extraordinary power to affect, for better or worse, countless animals in households, farms, science, entertainment, and the wild.
Much technical research is currently exploring ways of using AI to observe, analyse, and care for animals across four key domains. In industrial farming (Domain 1), precision tools and agents can increasingly monitor and automate farming activities for billions of land and sea animals while potentially reducing human oversight and intensifying production. AI deployed in automated vehicles like cars, trucks, delivery robots, and drones (Domain 2) could either kill or protect millions of wild and domestic animals according to how it is developed and governed. General purpose systems such as LLM chatbots and digital assistants (Domain 3) might either greatly promote or damage animal welfare in line with how they depict animals (e.g. by reproducing anti-animal prejudices from training data) and act on their wellbeing (e.g. through purchasing animal-friendly products).
Finally, AI is being designed for animals in conservation and habitat protection, healthcare, and interspecies communication (Domain 4); which could significantly assist many animals and species—and the ecosystems in which they live—or could create intentional or unintentional harm. As with humans, animal harm from AI can arise through bias, inaccuracy, opacity, misalignment, misuse, and abuse. To help ensure that AI improves rather than undermines animal welfare, human-animal relations, and environmental wellbeing, an effective ethical-legal response is urgently needed.
We are developing a first-of-its-kind ethical-legal framework for this purpose. Because law in the above areas is absent or limited, a substantial ethical analysis of AI’s impacts on animals is required to guide regulation and governance. Our ethical analysis adopts a pluralistic approach that incorporates major moral theories and adapts core principles from AI ethics. It also encompasses a fit-for-purpose conception of ‘animal welfare’, ethical lessons drawn from the emerging field of Animal-Computer Interaction (ACI), and insights from theories of relational entanglement with their focus on the interdependence of human, animal, and environmental wellbeing, and the way AI technologies reconfigure these interdependencies.
Our legal approach applies key concepts from animal law, multi-species justice and ecological regulation to identify opportunities in AI law and governance, as well as domain specific regulation (e.g. car safety design and farm animal welfare standards) to: (1) recognise that non-human animals have a legitimate interest in how AI systems are developed and used; (2) embed elements of animal-centred risk assessment in AI governance; (3) ensure those developing and deploying these technologies show responsibility and care for nonhuman animals; and (4) create institutions to permanently represent animals and their rights and interests in AI governance.
Plenary Session 7 – AI & Intellectual Property - Part 2
The concern over AI authorship has reached an all-time high. Regardless of creative domain, the fear that AI-generated content might directly or indirectly enter the market, has produced a whole set of formal and informal mechanisms directed at identifying and prohibiting the prevalence of AI in the creative process. The result has been the creation of a taboo around AI authorship. This AI authorship taboo is, however, predicated on a fundamental misunderstanding of the meaning, nature, and role of authorship. As a social institution deeply influenced by copyright law, authorship balances disparate normative criteria in developing an account of the appropriate human agency needed for different types of creativity. It thus works as a nuanced and contextual mechanism, which the AI authorship taboo risks flattening and oversimplifying. This Article argues for the rejection of the AI authorship taboo by looking to the working of authorship in copyright and beyond. Unpacking the different normative criteria underlying the working of authorship, it shows how the use of AI, to both mediate and generate expression, is fully compatible with human authorship and the mechanisms of collaborative creativity that it has come to embrace. Accepting AI authorship into the fold of authorship does not mean displacing, replacing, or undermining the human element in authorship; to the contrary it brings the meaning and significance of authorship into stark relief, allowing it to flourish and thrive in a world where AI is pervasive.
In some jurisdictions and courts, only certain forms of human–AI interaction are permitted to establish human control over AI-generated content. As a result, a user's copyright claim over AIGC may be rejected for lack of “sufficient control,” even where the user's creative contribution is substantial.
The essential problem with control theory lies in its inconsistency with the incentive theory of copyright. According to incentive theory, all other factors being equal, the more complex and information-rich the content becomes, the stronger the case for recognizing the human user's authorship, because such outputs are likely to embody greater human creative contribution. Under control theory, however, the opposite conclusion follows. As outputs become more complex, the user's relative degree of control is deemed to decline, since any increase in human contribution is outweighed by the even greater increase in machine-generated content. The more complex the output, the less likely the user is to succeed in an authorship claim.
The tension between control theory and incentive theory is irresolvable because the two approaches ask fundamentally different questions. Under incentive theory, the relevant inquiries are whether creative activity requires incentives and whether copyright protection would impose significant costs on the public domain. For complex content, whose creation is highly improbable and for which perhaps only thin copyright protection is sought, the answers will be “yes” and “no,” respectively. Incentive theory therefore favors protection.
Control theory, by contrast, replaces the assessment of incentive necessity with an assessment of human control. At the same time, it pays little attention to whether the resulting copyright claim would in fact pose a meaningful threat to the public domain. By substituting a proxy for the underlying objectives of copyright law, control theory risks producing results that diverge from the very goals copyright seeks to achieve.
The denial of copyright claims by AI systems should speak in favor of, rather than against, copyright claims by human users. At the de facto level, the “traditional elements of authorship” embodied in AI-generated content are jointly produced by human users and AI systems. The choice among labeling them as “human-made,” “machine-generated,” or “jointly produced” is not an ontological determination but a de jure judgment. Where recognizing users' claims would provide meaningful incentives without posing a substantial threat to the public domain, copyright law should favor recognition rather than denial.
By rejecting control theory, copyright law could save substantial cognitive resources that would otherwise be spent on the mandatory disclosure of the creative process. Instead, it could focus on the scope of protection. Courts should adopt a claim-based approach. The broader the protection sought, the heavier the burden on the user to separate human contributions from machine-generated contributions and to prove that the claimed subject matter falls within the scope of the human contribution.
There are at least two additional benefits to rejecting control theory. First, copyright law would maintain its theoretical coherence. Second, there is no need to introduce sui generis protection or to retreat to broad general clauses outside copyright law.
The development of Generative AI (GenAI) is disrupting the jurisprudence of copyright law. As AI-generated content becomes increasingly indistinguishable from human-authored works in appearance, style, and market function, copyright law faces pressure to decide whether, and how, such content should be protected. Existing proposals generally move in four directions. The first is to treat AI-generated content as AI-assisted work and interpret user involvement broadly enough to satisfy the human authorship requirement. The second is to grant only limited or “thin” protection, thereby distinguishing AI-generated content from human-authored works while still recognising some protectable interest. The third is to protect AI-generated content through neighbouring rights or a sui generis right, avoiding the hard question of human authorship. The fourth is to leave AI-generated content in the public domain on the grounds that it lacks human authorship and does not require copyright incentives.
Each approach, however, fails to answer the following question: If AI-generated content receives weaker protection than human-authored works, users will have strong incentives to present AI-generated content as human-authored. This creates an evidence problem that copyright law is poorly equipped to manage. Courts, copyright offices, platforms, and private parties would need to determine not only whether a work contains AI-generated elements, but also whether the level of human contribution is sufficient to justify copyright protection. As AI systems become more integrated into ordinary creative processes, this inquiry will become increasingly impossible. Differentiated protection may therefore produce the very misclassification problem it seeks to solve. However, AI-generated content has social consequences beyond copyright markets, including cognitive offloading, stagnation of human creativity, and risks to the authenticity of information. These consequences require a reliable legal and institutional separation between human-authored works and AI-generated content. Current labelling duties, largely imposed on platforms rather than individual users by certain jurisdictions, are insufficient. They risk becoming a Whack-A-Mole system in which regulators chase undisclosed AI outputs after dissemination. One needs a more coherent and workable copyright governance framework for such AI-generated content.
The paper therefore argues that, by shifting from the current opt-out model to an opt-in model for copyright protection, the risk of AI slop can be greatly reduced. The central question should not be how to exclude AI-generated content from copyright protection, but how claimants should affirmatively establish human authorship before receiving copyright protection. The most complete solution would be to reconsider the non-formality principle in international copyright law and permit formal authorship verification as a condition of protection. Because treaty revision is difficult, the paper also explores a more practical path: using copyright registration as an authorship-verification mechanism. On this view, registration would not impose a prohibited formality on copyright as such; rather, it would require the claimant to prove that the work is “authored” before the non-formality rule applies.
Plenary Session 8 – Living with AI: What’s Next?
Regulatory interoperability is a term more often used than clearly defined – perhaps less a clear goal than an aspiration: wouldn’t it be desirable if we could get regulatory systems to work together well, so that AI systems, products and services could flow across borders without flouting or bypassing local legal controls? If local regulators could more readily build upon each other’s compliance and regulatory work, audits – even their efforts at evidence gathering and enforcement?
These are all great aspirations to have, but how do we give substance and form to this rather vague, aspirational concept? Can we nail down a clear meaning; can we taxonomise it – clarify different types of interoperability; even develop some criteria for evaluating how far we are progressing towards an aspiration of interoperability and where more work is needed?
In this paper, we try to make some real progress along this path: we aim to start giving regulatory interoperability real substance, with the goal of taking it beyond mere aspiration, into a project we can pursue.
This involves significant analytical work. First, we make the case for the importance of regulatory interoperability, especially for countries like Australia and Singapore that are not major AI powers. We then develop and justify a definition, or least a sine qua non of regulatory interoperability, based around bi-directionality (interoperability is not the same as unilateral deregulation), and translation (interoperability requires that we respect, not seek to remove, different value and constitutional settings). We break down different dimensions of interoperability: normative, institutional, procedural and evidentiary, and set out some of the tools to achieve it. And we develop and justify some criteria for evaluating regulatory interoperability, and set out how we could go about the research and analytical program to perform that evaluation.
Our motivation is that this project of analytical heavy lifting is a necessary precondition to actually working on interoperability as a project. If, in the long term, middle and smaller powers are to continue to apply standards and protect their populations without pouring endless resources into compliance, assessment and enforcement work, they need better, more effective ways leverage each other's regulatory frameworks. Our paper is ultimately a blueprint for building that cooperation.
The emergence of Constitutional AI as a governance paradigm invites a question that its architects have not yet answered: if an AI system can have a constitution, can that constitution grow? Anthropic’s Constitutional AI framework trains models by reference to a set of written principles, known as a “constitution”, which is designed to guide self-critique, model training, and the production of outputs that are helpful and non-harmful.[1] This framework represents a significant departure from behavioral approaches to AI alignment, but it resolves one problem while generating another. It establishes what principles govern AI behavior, yet offers no account of how those principles should be interpreted, revised, or localized as technology evolves and deployment contexts change. This paper addresses that gap.
This paper argues that once the language of constitutionalism is adopted, constitutional law becomes an indispensable interlocutor. Drawing on the living tree doctrine that a constitution must be capable of organic growth within its fundamental structure,[2] this paper contends that Constitutional AI should be understood in analogous terms. Foundational concepts such as “harmlessness”, “fairness”, “privacy”, and “non-discrimination” do not carry fixed meanings but are contested across technologies, jurisdictions, and cultures. A principle drafted at a particular moment, or from a particular cultural standpoint, risks becoming substantively inadequate as values shift, technologies change, and new forms of human-machine interaction emerge. This inadequacy is both theoretically and practically significant given that large language models are deployed across diverse legal systems and communities in which speech, religion, gender, and public morality are differently configured, and what reads as a neutral principle in one jurisdiction may operate as a contested one in another.
Building on this analysis, this paper proposes a living tree framework for Constitutional AI. On this account, a well-designed Constitutional AI distinguishes between stable normative “roots” (including fundamental commitments to dignity, legality, equality, and non-arbitrariness) and the evolving “branched” application of those commitments to new technological and cultural contexts, through a process of principled revision conducted by identifiable institutional “gardeners” with defined roles and accountabilities (such as developers, regulators, ethicists, and affected communities). This paper’s claim is not that AI systems interpret their constitutions as courts do, but that Constitutional AI already converts principles into training signals, behavioral dispositions, and institutional choices, and that those conversions constitute exercises of normative authority that require a jurisprudence of their own.
This paper argues that Constitutional AI, while a significant advance in principled AI governance, lacks a theory of constitutional interpretation, and that supplying one is essential to addressing both an academic lacuna and a legitimate deficit with concrete consequences for how AI governance can be sustained, adapted and justified across jurisdictions and over time. This paper concludes that the living tree doctrine supplies the theoretical resources necessary to respond to the critical questions of by what process constitutional principles are reviewed, amended, localized, and reinterpreted; by whom; and against what standard, which remain structurally unaddressed by the current framework.
Most work on legal AI evaluation treats quality as binary (responses are either correct or incorrect) and objective (experts agree on what makes a response correct), relying on multiple-choice questions, classification tasks, and instance-level rubrics. Popular benchmarks in this tradition include LegalBench, LEXam, and Harvey's LAB. But much legal work resists this framing: quality is continuous—responses are better or worse, not right or wrong—and contested—experts may reasonably disagree about which response is better. Lawyers call this capacity judgement; in AI discussions, it is often called taste. We introduce LegalBench2, the first benchmark aimed at measuring judgement in the legal domain. LegalBench2 pairs tasks spanning transactional work, litigation, legal education, civil legal aid, arbitration, and other domains with responses ranked via pairwise preferences elicited from multiple expert attorneys. We describe the benchmark's construction and the research problems it opens up: building and validating faithful LLM judges, constructing personalized per-attorney leaderboards, and integrating validated judges into post-training pipelines.
Debates on artificial intelligence governance and liability have been dominated by regulatory models emerging from North America, Europe and East Asia. Yet AI systems increasingly generate social, political, environmental and economic effects beyond these jurisdictions, often in communities that have little influence over the design, deployment and governance of such technologies. This paper recentres contemporary discussions of AI governance by examining how Pacific perspectives can contribute to emerging frameworks of responsibility, accountability and technological justice.
The paper advances three principal arguments. First, it argues that contemporary AI governance relies extensively on soft law instruments, including ethical principles, voluntary standards and international guidelines. While frequently criticised for their limited enforceability, these instruments play a significant constitutive role in shaping expectations of responsible conduct. Drawing on Pacific legal experience and broader developments in international environmental and climate governance, the paper demonstrates that soft law can exercise substantial normative influence when grounded in relational and collective understandings of responsibility rather than narrow, fault-based conceptions of liability.
Secondly, the paper identifies important accountability gaps within existing AI governance frameworks. AI systems are largely developed, trained and controlled by actors located outside the Pacific, while many of their downstream impacts, including misinformation, extractive data practices, environmental costs associated with digital infrastructure, and threats to democratic participation, are experienced by communities with limited regulatory capacity and limited participation in governance processes. These asymmetries mirror challenges familiar in international climate and environmental law, where harms are often transboundary, responsibilities diffuse, and power unevenly distributed.
Thirdly, the paper argues that Pacific approaches to governance offer an alternative normative framework for understanding AI responsibility. Concepts of stewardship, relationality, collective wellbeing and intergenerational responsibility challenge dominant individualised and technocratic models of accountability. Rather than focusing on questions of AI personhood or machine autonomy, the paper conceptualises responsibility as a shared and ongoing obligation distributed across states, corporations, developers and institutions throughout the AI lifecycle. Such an approach foregrounds legitimacy, participation, consent and fairness as essential dimensions of effective governance.
The paper concludes that Pacific perspectives provide more than a regional case study. They offer a distinctive normative lens through which to evaluate emerging approaches to AI governance and liability. By highlighting relational responsibility and the continuing importance of soft law in shaping global standards of conduct, the paper contributes to the development of AI governance frameworks that are not only technically effective but also socially legitimate, inclusive and globally just.