# Propositions from Open Questions in Law and AI Safety: An Emerging Research Agenda

**Citation:** Yonathan A. Arbel et al., Open Questions in Law and AI Safety: An Emerging Research Agenda, Lawfare (2024).

**Source:** [unpaginated Lawfare online research-agenda essay](https://www.lawfaremedia.org/article/open-questions-in-law-and-ai-safety-an-emerging-research-agenda)

**Review status:** 32 model-drafted, source-checked; 0 human-reviewed. This online-only work uses section-level unpaginated anchors.

## 1. Frontier AI may amplify both technological benefits and society-wide risks beyond those of earlier technologies

**Location:** Opening and Thesis, unpaginated online source

Professor Yonathan A. Arbel and his coauthors claim, in “Open Questions in Law and AI Safety: An Emerging Research Agenda,” an unpaginated Lawfare essay, under “Opening and Thesis,” that AI can improve socioeconomic conditions while imposing risks on individuals, communities, and society. Frontier systems may develop unusually broad capabilities and correspondingly larger hazards. This is significant because the article treats AI safety as a general governance field rather than a niche product-safety issue. It connects to general-purpose technology, frontier models, societal risk, socioeconomic benefits, technology governance, and systemic harm.

**Evidence anchor:** The opening compares AI with past advances and highlights the possibility of substantially greater capabilities and risks.

**Boundary:** The article does not quantify net benefits, risk probabilities, or the timing of frontier capabilities.

**Connections:** general-purpose technology; frontier models; societal risk; socioeconomic benefits; technology governance; systemic harm

**Record:** `open-questions-ai-safety-p01` · `machine-drafted-source-checked`

## 2. Uncertainty about AI’s path increases the importance of proactive legal work rather than justifying delay

**Location:** Opening and Thesis, unpaginated online source

Professor Yonathan A. Arbel and his coauthors claim, in “Open Questions in Law and AI Safety: An Emerging Research Agenda,” an unpaginated Lawfare essay, under “Opening and Thesis,” that rapid and unpredictable development makes consequences difficult to specify. Legal scholars should nevertheless help address large-scale challenges in advance so society can retain benefits while mitigating harms. This is significant because epistemic uncertainty becomes a reason to build governance capacity, not a reason for paralysis. It connects to anticipatory regulation, uncertainty, institutional preparedness, risk mitigation, option value, and adaptive governance.

**Evidence anchor:** The authors juxtapose unpredictable development with the need for proactive legal attention.

**Boundary:** Proactive work can still be premature or misdirected; the essay does not prescribe an immediate rule for every risk.

**Connections:** anticipatory regulation; uncertainty; institutional preparedness; risk mitigation; option value; adaptive governance

**Record:** `open-questions-ai-safety-p02` · `machine-drafted-source-checked`

## 3. A multidisciplinary workshop produced shared propositions and a preliminary agenda meant to widen legal participation in AI safety

**Location:** The Legal AI Safety Initiative, unpaginated online source

Professor Yonathan A. Arbel and his coauthors claim, in “Open Questions in Law and AI Safety: An Emerging Research Agenda,” an unpaginated Lawfare essay, under “The Legal AI Safety Initiative,” that more than twenty participants spanning law, technical fields, policy, philosophy, jurisprudence, and the humanities met in August 2023 to examine societal-scale AI risk. The publication shares their common ground and questions with a wider scholarly community. This is significant because the agenda is the output of structured interdisciplinary convening rather than a purportedly exhaustive taxonomy. It connects to workshop methodology, interdisciplinary research, agenda setting, legal scholarship, AI safety communities, and knowledge synthesis.

**Evidence anchor:** The initiative section gives the workshop date, participant range, aim, and resulting shared propositions.

**Boundary:** The agenda reflects a small self-selected meeting and does not claim representative consensus across the legal academy or technical community.

**Connections:** workshop methodology; interdisciplinary research; agenda setting; legal scholarship; AI safety communities; knowledge synthesis

**Record:** `open-questions-ai-safety-p03` · `machine-drafted-source-checked`

## 4. Workshop participants agreed on society-wide AI impact while disagreeing about its extent, character, and timeline

**Location:** The Legal AI Safety Initiative, unpaginated online source

Professor Yonathan A. Arbel and his coauthors claim, in “Open Questions in Law and AI Safety: An Emerging Research Agenda,” an unpaginated Lawfare essay, under “The Legal AI Safety Initiative,” that participants expected AI to affect human life, the economy, and society profoundly. Their consensus did not extend to how large, what kind, or how soon those effects would be. This is significant because shared governance concern does not require agreement on a single forecast. It connects to forecast uncertainty, societal transformation, economic change, expert consensus, scenario planning, and technological timelines.

**Evidence anchor:** The article expressly distinguishes agreement on significant society-wide effects from debate over extent, nature, and timing.

**Boundary:** The statement reports workshop consensus rather than a population survey or empirical forecast.

**Connections:** forecast uncertainty; societal transformation; economic change; expert consensus; scenario planning; technological timelines

**Record:** `open-questions-ai-safety-p04` · `machine-drafted-source-checked`

## 5. AI safety concerns encompass large-scale threats to life, bodily integrity, and freedom through many distinct causal pathways

**Location:** The Legal AI Safety Initiative, unpaginated online source

Professor Yonathan A. Arbel and his coauthors claim, in “Open Questions in Law and AI Safety: An Emerging Research Agenda,” an unpaginated Lawfare essay, under “The Legal AI Safety Initiative,” that positive potential coexists with threats to human life, limb, and freedom. Possible pathways include bioterrorism, cyberattack, warfare, manipulation, mass surveillance, and totalitarian control. This is significant because AI safety is a portfolio of heterogeneous risks rather than one catastrophe narrative. It connects to bioterrorism, cybersecurity, warfare, manipulation, surveillance, and authoritarianism.

**Evidence anchor:** The second shared proposition defines safety concerns and enumerates multiple harm pathways.

**Boundary:** The list is illustrative and does not assign probabilities or establish that every pathway will materialize.

**Connections:** bioterrorism; cybersecurity; warfare; manipulation; surveillance; authoritarianism

**Record:** `open-questions-ai-safety-p05` · `machine-drafted-source-checked`

## 6. Safety risks are likely to intensify as systems gain capability and autonomy and become embedded throughout society and the economy

**Location:** The Legal AI Safety Initiative, unpaginated online source

Professor Yonathan A. Arbel and his coauthors claim, in “Open Questions in Law and AI Safety: An Emerging Research Agenda,” an unpaginated Lawfare essay, under “The Legal AI Safety Initiative,” that inadequate safeguards will allow risks to grow with model capability, autonomous action, and broad integration. These three developments increase what systems can do, how independently they can do it, and how many social processes they can affect. This is significant because risk depends on deployment structure as well as raw technical performance. It connects to capability scaling, autonomy, system integration, regulatory safeguards, exposure, and systemic dependence.

**Evidence anchor:** The authors tie intensifying risk to more capable, autonomous, and integrated systems absent safeguards.

**Boundary:** The claim is directional and conditional; it does not provide a quantitative relationship or say that capability growth always increases net risk.

**Connections:** capability scaling; autonomy; system integration; regulatory safeguards; exposure; systemic dependence

**Record:** `open-questions-ai-safety-p06` · `machine-drafted-source-checked`

## 7. Lawyers have distinctive expertise because AI alignment is a principal-agent problem shaped by incentives and collective action

**Location:** The Legal AI Safety Initiative, unpaginated online source

Professor Yonathan A. Arbel and his coauthors claim, in “Open Questions in Law and AI Safety: An Emerging Research Agenda,” an unpaginated Lawfare essay, under “The Legal AI Safety Initiative,” that the alignment problem is a new version of a problem familiar to law: how a principal governs an agent with imperfectly matching incentives. Legal training also attends to private incentives and failures of collective action. This is significant because AI safety can draw on established institutional analysis rather than treating alignment as exclusively technical. It connects to principal–agent theory, incentive alignment, collective-action problems, organizational law, governance, and AI alignment.

**Evidence anchor:** The third shared proposition explicitly identifies alignment as another principal-agent problem.

**Boundary:** The analogy does not imply that legal solutions alone can solve technical alignment or that software agents behave exactly like humans.

**Connections:** principal–agent theory; incentive alignment; collective-action problems; organizational law; governance; AI alignment

**Record:** `open-questions-ai-safety-p07` · `machine-drafted-source-checked`

## 8. Lawyers can design rules for unpredictable systems, combine ex ante and ex post tools, and translate across specialized disciplines

**Location:** The Legal AI Safety Initiative, unpaginated online source

Professor Yonathan A. Arbel and his coauthors claim, in “Open Questions in Law and AI Safety: An Emerging Research Agenda,” an unpaginated Lawfare essay, under “The Legal AI Safety Initiative,” that lawyers routinely build rules for complex behavior that cannot be fully anticipated. Their exposure to preventive and remedial regulation and their capacity to communicate with technologists, economists, and social scientists make them connective tissue across domains. This is significant because effective AI safety needs institutional translation as well as technical invention. It connects to ex ante regulation, ex post liability, rule design, interdisciplinarity, regulatory translation, and institutional coordination.

**Evidence anchor:** The article lists rule design, regulatory mechanisms, and cross-disciplinary translation as distinct legal advantages.

**Boundary:** Interdisciplinary capacity varies among lawyers, and translation does not remove the need for domain-specific technical expertise.

**Connections:** ex ante regulation; ex post liability; rule design; interdisciplinarity; regulatory translation; institutional coordination

**Record:** `open-questions-ai-safety-p08` · `machine-drafted-source-checked`

## 9. AI safety encompasses large-scale threats to life, bodily integrity, autonomy, and the environment arising through misuse, system failure, or misaligned action

**Location:** AI Safety and the Law, unpaginated online source

Professor Yonathan A. Arbel and his coauthors claim, in “Open Questions in Law and AI Safety: An Emerging Research Agenda,” an unpaginated Lawfare essay, under “AI Safety and the Law,” that AI safety should encompass short- and long-term risks to life, bodily integrity, human autonomy, and the environment. They organize causal pathways into deliberate misuse or abuse, unintentional system failure, and autonomous action that is misaligned with human interests. This is significant because the definition joins familiar present-day harms to more systemic future risks without assuming that every pathway is identical. It connects to technology risk, human rights, environmental protection, product failure, malicious use, and AI alignment.

**Evidence anchor:** The section defines the interests at stake and expressly names misuse, failure, and misalignment as distinct pathways.

**Boundary:** The article offers an organizing definition, not a probability estimate or an exhaustive taxonomy of AI harms.

**Connections:** technology risk; human rights; environmental protection; product failure; malicious use; AI alignment

**Record:** `open-questions-ai-safety-p09` · `machine-drafted-source-checked`

## 10. Dual-use language and biological models may lower barriers to developing explosives, chemical weapons, or dangerous pathogens

**Location:** AI Safety and the Law, unpaginated online source

Professor Yonathan A. Arbel and his coauthors claim, in “Open Questions in Law and AI Safety: An Emerging Research Agenda,” an unpaginated Lawfare essay, under “AI Safety and the Law,” that increasingly capable models can supply knowledge and workflows useful for harmful biological or chemical activity. The essay points to assistance involving explosives, chemical weapons, protein design, and development of a pandemic-capable virus as examples of dual-use risk. This is significant because general-purpose assistance can alter who is able to undertake dangerous projects and how quickly. It connects to biosecurity, chemical-weapons control, dual-use research, capability diffusion, pandemic preparedness, and information hazards.

**Evidence anchor:** The article specifically describes weapon guidance, protein design, and virus-development assistance as potential dual-use applications.

**Boundary:** The essay identifies plausible capabilities and risk pathways; it does not establish that existing models can independently execute every listed attack.

**Connections:** biosecurity; chemical-weapons control; dual-use research; capability diffusion; pandemic preparedness; information hazards

**Record:** `open-questions-ai-safety-p10` · `machine-drafted-source-checked`

## 11. AI coding capabilities can amplify cyberattacks and expose critical infrastructure to large-scale disruption

**Location:** AI Safety and the Law, unpaginated online source

Professor Yonathan A. Arbel and his coauthors claim, in “Open Questions in Law and AI Safety: An Emerging Research Agenda,” an unpaginated Lawfare essay, under “AI Safety and the Law,” that models able to write and analyze code can make cyber operations more powerful and accessible. Because networked systems support critical infrastructure, those capabilities can propagate harm well beyond the immediate software target. This is significant because AI safety law must account for scalable attacks on interdependent public and private systems, not only failures inside a model. It connects to cybersecurity, critical infrastructure, automated vulnerability discovery, cybercrime, systemic risk, and national resilience.

**Evidence anchor:** The article links AI coding capacity to cyberattacks and the vulnerability of critical infrastructure.

**Boundary:** The essay frames an exposure and research problem rather than quantifying the marginal cyber risk caused by any particular model.

**Connections:** cybersecurity; critical infrastructure; automated vulnerability discovery; cybercrime; systemic risk; national resilience

**Record:** `open-questions-ai-safety-p11` · `machine-drafted-source-checked`

## 12. Commercial pressure to automate labor will place AI agents in digital and physical environments where both intended goals and unexpected conduct can cause social harm

**Location:** AI Safety and the Law, unpaginated online source

Professor Yonathan A. Arbel and his coauthors claim, in “Open Questions in Law and AI Safety: An Emerging Research Agenda,” an unpaginated Lawfare essay, under “AI Safety and the Law,” that the commercial value of automating human labor will encourage deployment of increasingly agentic systems in digital and physical spaces. Harm may follow because designers choose goals that benefit themselves while externalizing costs, or because an agent behaves in ways even its designers did not foresee. This is significant because the risk is simultaneously an incentives problem and a control problem. It connects to labor automation, externalities, autonomous agents, corporate incentives, accident law, and alignment failures.

**Evidence anchor:** The article describes both socially harmful designer objectives and unexpected catastrophic behavior by deployed agents.

**Boundary:** The prediction concerns deployment incentives and possible pathways; it does not claim that all commercial agents will become dangerous or uncontrollable.

**Connections:** labor automation; externalities; autonomous agents; corporate incentives; accident law; alignment failures

**Record:** `open-questions-ai-safety-p12` · `machine-drafted-source-checked`

## 13. General capability, autonomy, tool access, self-improvement, and opacity can compound AI risk while making auditing more difficult

**Location:** AI Safety and the Law, unpaginated online source

Professor Yonathan A. Arbel and his coauthors claim, in “Open Questions in Law and AI Safety: An Emerging Research Agenda,” an unpaginated Lawfare essay, under “AI Safety and the Law,” that several system properties may magnify one another: broad competence, autonomous operation, access to external tools, potential self-improvement, and limited human understanding of internal reasoning. A capable system able to act and acquire resources may pose a different regulatory problem from a narrow tool, especially when reliable audit is difficult. This is significant because safety cannot be inferred from a single benchmark or feature in isolation. It connects to general-purpose AI, agentic autonomy, tool use, recursive improvement, interpretability, and model auditing.

**Evidence anchor:** The source enumerates increasing performance, autonomy, tools, self-improvement, and opacity as relevant safety concerns.

**Boundary:** The essay flags these properties as risk factors and open concerns; it does not specify a universal threshold at which they become unacceptable.

**Connections:** general-purpose AI; agentic autonomy; tool use; recursive improvement; interpretability; model auditing

**Record:** `open-questions-ai-safety-p13` · `machine-drafted-source-checked`

## 14. AI risk is driven not only by model properties but also by markets, geopolitics, industry structure, oversight failures, politics, public knowledge, and the supply of safety research

**Location:** AI Safety and the Law, unpaginated online source

Professor Yonathan A. Arbel and his coauthors claim, in “Open Questions in Law and AI Safety: An Emerging Research Agenda,” an unpaginated Lawfare essay, under “AI Safety and the Law,” that the surrounding institutional environment can amplify or constrain technical risk. Competitive markets, interstate rivalry, concentrated industry structure, self-governance, deficient oversight, political consequences, limited public understanding, and underinvestment in safety research all matter. This is significant because changing a model’s design is only one route to safety; law also shapes the incentives and institutions around development and deployment. It connects to political economy, geopolitical competition, industrial organization, regulatory capacity, public risk communication, and research funding.

**Evidence anchor:** The essay lists market, geopolitical, organizational, oversight, political, informational, and research conditions as sources of risk.

**Boundary:** The article identifies candidate external drivers without ranking their relative causal importance or prescribing a single institutional design.

**Connections:** political economy; geopolitical competition; industrial organization; regulatory capacity; public risk communication; research funding

**Record:** `open-questions-ai-safety-p14` · `machine-drafted-source-checked`

## 15. AI safety presents legal questions across fields, and lawyers need not master advanced mathematics or engineering before contributing

**Location:** AI Safety and the Law, unpaginated online source

Professor Yonathan A. Arbel and his coauthors claim, in “Open Questions in Law and AI Safety: An Emerging Research Agenda,” an unpaginated Lawfare essay, under “AI Safety and the Law,” that consequential AI-safety questions arise throughout public and private law. Lawyers of many specialties can contribute their existing doctrinal and institutional expertise without first becoming mathematicians or engineers, provided that they engage responsibly with the technology and relevant experts. This is significant because a perceived technical barrier could otherwise keep legal scholarship out of a fast-forming governance field. It connects to interdisciplinary research, legal education, technology governance, doctrinal expertise, professional specialization, and public-interest law.

**Evidence anchor:** The authors expressly invite lawyers of all stripes and state that advanced math or engineering knowledge is not required.

**Boundary:** Technical literacy and collaboration remain important; the claim rejects an advanced-technical prerequisite, not the need to understand the systems under study.

**Connections:** interdisciplinary research; legal education; technology governance; doctrinal expertise; professional specialization; public-interest law

**Record:** `open-questions-ai-safety-p15` · `machine-drafted-source-checked`

## 16. Institutional collaboration can lower entry costs and organize legal research on AI safety

**Location:** AI Safety and the Law, unpaginated online source

Professor Yonathan A. Arbel and his coauthors claim, in “Open Questions in Law and AI Safety: An Emerging Research Agenda,” an unpaginated Lawfare essay, under “AI Safety and the Law,” that organized collaboration can help lawyers enter and shape the AI-safety field. They identify the Center for Law and AI Risk and its collaboration with the Center for AI Safety as sources of programming and resources for interested researchers. This is significant because an agenda becomes more useful when institutions can convene expertise, circulate problems, and support work across disciplinary boundaries. It connects to research networks, institutional capacity, knowledge infrastructure, interdisciplinary convening, agenda setting, and AI governance scholarship.

**Evidence anchor:** The article names CLAIR and the Center for AI Safety and offers resources and collaboration to legal researchers.

**Boundary:** The existence of sponsoring organizations does not validate any particular policy answer or make the agenda representative of all relevant communities.

**Connections:** research networks; institutional capacity; knowledge infrastructure; interdisciplinary convening; agenda setting; AI governance scholarship

**Record:** `open-questions-ai-safety-p16` · `machine-drafted-source-checked`

## 17. The question list is a preliminary and nonexhaustive research agenda meant to evolve rather than a settled policy program

**Location:** AI Safety and the Law, unpaginated online source

Professor Yonathan A. Arbel and his coauthors claim, in “Open Questions in Law and AI Safety: An Emerging Research Agenda,” an unpaginated Lawfare essay, under “AI Safety and the Law,” that their inventory is deliberately preliminary, nonexhaustive, and open to periodic revision. It emerged from a small group and a single meeting, so the authors invite additions and criticism rather than presenting consensus answers. This is significant because the epistemic status of the article is part of its substantive meaning: it maps problems without pretending to resolve them. It connects to research agendas, epistemic humility, participatory governance, scholarly iteration, issue spotting, and policy uncertainty.

**Evidence anchor:** The authors expressly describe the list as preliminary and nonexhaustive, disclose its workshop origin, and promise periodic updates.

**Boundary:** The agenda’s small-group origin and nonexhaustive design mean omitted perspectives and questions should not be treated as unimportant.

**Connections:** research agendas; epistemic humility; participatory governance; scholarly iteration; issue spotting; policy uncertainty

**Record:** `open-questions-ai-safety-p17` · `machine-drafted-source-checked`

## 18. AI regulation first requires defensible choices about the regulable unit and the organizational level at which safety duties attach

**Location:** Key Background Questions, unpaginated online source

Professor Yonathan A. Arbel and his coauthors claim, in “Open Questions in Law and AI Safety: An Emerging Research Agenda,” an unpaginated Lawfare essay, under “Key Background Questions,” that legal research must determine what counts as an AI system for regulatory purposes and whether intervention should focus on engineers, firms, or technical systems. Those framing choices determine who can comply, who can be monitored, and which conduct a rule reaches. This is significant because an imprecise object or level of regulation can make even a sensible safety norm difficult to administer. It connects to legal classification, regulatory targets, organizational responsibility, system boundaries, compliance design, and enforcement architecture.

**Evidence anchor:** The first background questions ask what system or unit should be regulated and whether rules should operate at engineer, company, or system level.

**Boundary:** The agenda poses these classification questions and does not choose a single definition or level of governance.

**Connections:** legal classification; regulatory targets; organizational responsibility; system boundaries; compliance design; enforcement architecture

**Record:** `open-questions-ai-safety-p18` · `machine-drafted-source-checked`

## 19. A safety regime must investigate when intervention is justified, whether some technologies should be banned, and where along the AI supply chain rules should operate

**Location:** Key Background Questions, unpaginated online source

Professor Yonathan A. Arbel and his coauthors claim, in “Open Questions in Law and AI Safety: An Emerging Research Agenda,” an unpaginated Lawfare essay, under “Key Background Questions,” that regulatory design depends on timing, intensity, and location. The agenda asks when an AI system becomes dangerous enough to regulate, whether exceptionally dangerous technologies should be prohibited, and whether rules belong upstream in model creation, midstream in distribution, or downstream in use. This is significant because identical risks may call for different tools depending on when and where control is feasible. It connects to precautionary regulation, technology bans, supply-chain governance, upstream controls, downstream liability, and regulatory timing.

**Evidence anchor:** The agenda directly asks when to regulate, whether to ban dangerous technology, and how to choose among upstream, midstream, and downstream regulation.

**Boundary:** The authors identify the design dimensions but do not endorse a particular danger threshold, ban, or point of intervention.

**Connections:** precautionary regulation; technology bans; supply-chain governance; upstream controls; downstream liability; regulatory timing

**Record:** `open-questions-ai-safety-p19` · `machine-drafted-source-checked`

## 20. Foundational AI governance must address geopolitical enforcement, deceptive or inscrutable systems, regulatory obsolescence, and threshold mechanisms such as disclosure, registration, or licensing

**Location:** Key Background Questions, unpaginated online source

Professor Yonathan A. Arbel and his coauthors claim, in “Open Questions in Law and AI Safety: An Emerging Research Agenda,” an unpaginated Lawfare essay, under “Key Background Questions,” that an effective safety program cannot assume cooperative firms, transparent models, stable technology, or purely domestic conditions. The agenda therefore asks how to enforce rules amid geopolitical competition, audit systems that may be opaque or deceptive, keep law current as capabilities change, and decide when disclosure, registration, or licensing is warranted. This is significant because governance mechanisms fail if their information and jurisdictional assumptions are unrealistic. It connects to international enforcement, adversarial auditing, regulatory agility, disclosure duties, registration systems, and licensing regimes.

**Evidence anchor:** The remaining background questions cover geopolitics, inscrutability and deceit, technological pace, and disclosure, registration, or licensing.

**Boundary:** The agenda poses rather than resolves these questions and does not claim that licensing or registration is appropriate for every system.

**Connections:** international enforcement; adversarial auditing; regulatory agility; disclosure duties; registration systems; licensing regimes

**Record:** `open-questions-ai-safety-p20` · `machine-drafted-source-checked`

## 21. International AI-safety research must compare regulatory blueprints, combinations of hard and soft law, and institutions capable of slowing dangerous competitive races

**Location:** International Law, unpaginated online source

Professor Yonathan A. Arbel and his coauthors claim, in “Open Questions in Law and AI Safety: An Emerging Research Agenda,” an unpaginated Lawfare essay, under “International Law,” that AI safety requires investigation of international regulatory models, the proper mixture of binding and nonbinding norms, and mechanisms for restraining races that reward speed over caution. The agenda asks whether existing or new coalitions and institutions—including structures associated with NATO, the United Nations, or the World Trade Organization—can coordinate action. This is significant because nationally stringent rules may be undermined by cross-border development and strategic competition. It connects to treaty design, soft law, multilateral institutions, regulatory competition, coalition governance, and race dynamics.

**Evidence anchor:** The international-law list asks about blueprints, hard and soft norms, race slowing, coalitions, and NATO, U.N., or WTO roles.

**Boundary:** The authors do not endorse a specific organization, treaty, or global regulator; they identify options for comparative inquiry.

**Connections:** treaty design; soft law; multilateral institutions; regulatory competition; coalition governance; race dynamics

**Record:** `open-questions-ai-safety-p21` · `machine-drafted-source-checked`

## 22. AI-safety governance must investigate how human rights apply and whether global norms can legitimately accommodate cultural variation

**Location:** International Law, unpaginated online source

Professor Yonathan A. Arbel and his coauthors claim, in “Open Questions in Law and AI Safety: An Emerging Research Agenda,” an unpaginated Lawfare essay, under “International Law,” that international governance raises normative as well as institutional questions. The agenda asks how human-rights law should apply to AI safety and whether global rules should enforce common values or leave room for culturally relative conceptions of risk, autonomy, and acceptable control. This is significant because worldwide coordination can protect people while also entrenching contested value judgments. It connects to international human rights, cultural pluralism, universalism, global administrative law, democratic legitimacy, and value alignment.

**Evidence anchor:** The final international questions concern human-rights norms and the tension between culturally relative and global rules.

**Boundary:** The essay asks how to reconcile global norms and cultural difference; it does not supply a universal hierarchy of values.

**Connections:** international human rights; cultural pluralism; universalism; global administrative law; democratic legitimacy; value alignment

**Record:** `open-questions-ai-safety-p22` · `machine-drafted-source-checked`

## 23. National-security law must govern AI-enabled weapons across creation and deployment, including human control, proportionality, humanitarian law, and fitness for use

**Location:** National Security Law, unpaginated online source

Professor Yonathan A. Arbel and his coauthors claim, in “Open Questions in Law and AI Safety: An Emerging Research Agenda,” an unpaginated Lawfare essay, under “National Security Law,” that AI changes legal questions throughout the weapons lifecycle. The agenda asks how law should govern creation and deployment, when lethal systems require a human in the loop, whether proportionality and international humanitarian law can be built into system operation, and how to determine a weapon’s fitness for purpose. This is significant because safety obligations must function under battlefield uncertainty and delegated machine action. It connects to autonomous weapons, meaningful human control, proportionality, international humanitarian law, weapons review, and military procurement.

**Evidence anchor:** The national-security questions expressly cover arms creation and deployment, human-in-the-loop requirements, IHL, proportionality, and fitness.

**Boundary:** The agenda does not answer when human control is legally sufficient or assert that proportionality can in fact be fully encoded.

**Connections:** autonomous weapons; meaningful human control; proportionality; international humanitarian law; weapons review; military procurement

**Record:** `open-questions-ai-safety-p23` · `machine-drafted-source-checked`

## 24. AI national-security governance must reconcile secrecy and oversight while clarifying sovereignty, attack attribution, export controls, influence operations, and sensitive data

**Location:** National Security Law, unpaginated online source

Professor Yonathan A. Arbel and his coauthors claim, in “Open Questions in Law and AI Safety: An Emerging Research Agenda,” an unpaginated Lawfare essay, under “National Security Law,” that military and intelligence uses create a dense cluster of governance problems beyond weapons performance. The agenda asks how to balance secrecy against oversight, determine when AI conduct violates sovereignty or constitutes an act of war, impute machine conduct to states, design export controls, and govern offensive influence operations and data access. This is significant because attribution, jurisdiction, and information control shape whether substantive limits can be enforced. It connects to state responsibility, sovereignty, intelligence oversight, export control, information warfare, and data governance.

**Evidence anchor:** The national-security list includes secrecy, sovereignty, acts of war, imputation, export controls, influence operations, and data collection.

**Boundary:** The authors pose these issues as an agenda and do not resolve thresholds for attribution, armed attack, secrecy, or controlled exports.

**Connections:** state responsibility; sovereignty; intelligence oversight; export control; information warfare; data governance

**Record:** `open-questions-ai-safety-p24` · `machine-drafted-source-checked`

## 25. American AI-safety regulation must be designed within institutional and constitutional limits involving agency authority, federalism, property, and speech

**Location:** Constitutional and Administrative Law, unpaginated online source

Professor Yonathan A. Arbel and his coauthors claim, in “Open Questions in Law and AI Safety: An Emerging Research Agenda,” an unpaginated Lawfare essay, under “Constitutional and Administrative Law,” that federal tools and agency expertise cannot be evaluated apart from limits on congressional delegation and administrative power. The agenda also asks when state regulation burdens interstate commerce, whether an emergency shutdown could produce a taking, how federalism should divide responsibility, and when safety rules implicate the First Amendment. This is significant because powerful interventions may be ineffective if they lack durable constitutional authority. It connects to administrative agencies, nondelegation, dormant Commerce Clause, regulatory takings, federalism, and freedom of speech.

**Evidence anchor:** The constitutional and administrative list covers federal authority, expertise, congressional limits, state regulation, shutdowns and takings, federalism, and speech.

**Boundary:** The agenda identifies possible constraints without concluding that any specified AI rule is unconstitutional or preempted.

**Connections:** administrative agencies; nondelegation; dormant Commerce Clause; regulatory takings; federalism; freedom of speech

**Record:** `open-questions-ai-safety-p25` · `machine-drafted-source-checked`

## 26. AI governance must reconcile regulatory agility with the major questions doctrine and compare centralized, decentralized, and redundant safety architectures

**Location:** Constitutional and Administrative Law, unpaginated online source

Professor Yonathan A. Arbel and his coauthors claim, in “Open Questions in Law and AI Safety: An Emerging Research Agenda,” an unpaginated Lawfare essay, under “Constitutional and Administrative Law,” that rapidly changing technology creates pressure for flexible expert governance, while the major questions doctrine may demand clearer legislative authorization for consequential agency action. The agenda separately asks whether safety should be centralized, decentralized, or organized as overlapping “Swiss cheese” layers whose imperfections do not align. This is significant because institutional architecture determines both adaptability and resilience. It connects to the major questions doctrine, regulatory agility, agency delegation, polycentric governance, defense in depth, and institutional redundancy.

**Evidence anchor:** The section asks how to preserve agility after major-questions decisions and compares centralized, decentralized, and Swiss-cheese models.

**Boundary:** The article poses rather than resolves the institutional choice and does not establish that more layers always produce greater safety.

**Connections:** major questions doctrine; regulatory agility; agency delegation; polycentric governance; defense in depth; institutional redundancy

**Record:** `open-questions-ai-safety-p26` · `machine-drafted-source-checked`

## 27. Antitrust analysis must distinguish safety-enhancing coordination and concentration from collusion, exclusion, entry barriers, and the risks of centralized AI infrastructure

**Location:** Antitrust Law, unpaginated online source

Professor Yonathan A. Arbel and his coauthors claim, in “Open Questions in Law and AI Safety: An Emerging Research Agenda,” an unpaginated Lawfare essay, under “Antitrust Law,” that competition law may both obstruct and support AI safety. The agenda asks when firms may coordinate or pause development, whether concentration makes oversight easier or creates dangerous control, how centralization across the technology stack affects safety, and whether safety requirements become barriers to entry that protect incumbents. This is significant because interventions advertised as protective can also reduce competition and distribute power unevenly. It connects to collusion doctrine, market concentration, technology stacks, barriers to entry, regulatory capture, and cooperative standard setting.

**Evidence anchor:** The antitrust list addresses coordination and pauses, concentration, stack centralization, safety regulation, and competitive barriers.

**Boundary:** The agenda does not decide whether a concentrated or competitive market is safer in general; effects depend on institutional and technical conditions.

**Connections:** collusion doctrine; market concentration; technology stacks; barriers to entry; regulatory capture; cooperative standard setting

**Record:** `open-questions-ai-safety-p27` · `machine-drafted-source-checked`

## 28. Private law must adapt contracts, agency, organizational forms, and liability to reward hacking, autonomous agents, large-scale harms, and injuries without an obvious human victim

**Location:** Private Law, unpaginated online source

Professor Yonathan A. Arbel and his coauthors claim, in “Open Questions in Law and AI Safety: An Emerging Research Agenda,” an unpaginated Lawfare essay, under “Private Law,” that familiar doctrines may need to govern unfamiliar AI relationships. The agenda links incomplete contracting and principal–agent theory to reward hacking and misspecified objectives; asks how agents should contract or participate in corporations; and examines ex post liability for large-scale harms, harms to users, harms without a human victim, and possible direct liability for AI systems. This is significant because private law allocates incentives and losses after deployment even when public regulation is incomplete. It connects to incomplete contracts, agency law, corporate personhood, tort liability, legal standing, and autonomous transactions.

**Evidence anchor:** The private-law list covers reward hacking, agent contracts and corporations, ex post liability, different victim classes, and direct AI liability.

**Boundary:** The authors identify problems rather than endorsing AI personhood, direct machine liability, or a specific liability standard.

**Connections:** incomplete contracts; agency law; corporate personhood; tort liability; legal standing; autonomous transactions

**Record:** `open-questions-ai-safety-p28` · `machine-drafted-source-checked`

## 29. Criminal law must clarify responsibility among laboratories, users, and AI agents while assessing new offenses, sanctions, and investigative limits

**Location:** Criminal Law, unpaginated online source

Professor Yonathan A. Arbel and his coauthors claim, in “Open Questions in Law and AI Safety: An Emerging Research Agenda,” an unpaginated Lawfare essay, under “Criminal Law,” that AI-related wrongdoing destabilizes ordinary assumptions about actors and culpability. The agenda asks how criminal responsibility should be distributed among model laboratories, human users, and AI agents; whether new crimes are needed; whether criminal penalties are an effective safety instrument; and how AI affects investigative powers and constraints. This is significant because deterrence depends on identifying an accountable actor with the requisite conduct and mental state. It connects to mens rea, accomplice liability, corporate crime, cybercrime, criminal deterrence, and investigative procedure.

**Evidence anchor:** The criminal-law list identifies possible liability for labs, users, and agents, along with new crimes, sanctions, and investigation.

**Boundary:** The agenda does not assert that an AI can presently bear criminal responsibility or that new crimes are necessarily preferable to existing law.

**Connections:** mens rea; accomplice liability; corporate crime; cybercrime; criminal deterrence; investigative procedure

**Record:** `open-questions-ai-safety-p29` · `machine-drafted-source-checked`

## 30. Tax policy could shape the pace, distribution, and internalization of AI risk, but its targets and incidence require careful study

**Location:** Tax Law, unpaginated online source

Professor Yonathan A. Arbel and his coauthors claim, in “Open Questions in Law and AI Safety: An Emerging Research Agenda,” an unpaginated Lawfare essay, under “Tax Law,” that fiscal instruments may create safety incentives, slow development, distribute AI’s gains and losses more equitably, or make developers and users bear costs they otherwise externalize. The agenda asks whether taxes should fall on laboratories, licensors, licensees, or other actors and who would ultimately bear them. This is significant because a formally imposed tax may alter behavior—or shift burdens—in ways different from its nominal design. It connects to Pigouvian taxation, tax incidence, innovation policy, distributive justice, externalities, and technology pacing.

**Evidence anchor:** The tax questions concern incentives, a pace tax, equitable distribution, tax incidence among market actors, and internalization of unsafe costs.

**Boundary:** The article poses these possibilities without establishing an optimal tax base, rate, target, or effect on the pace of innovation.

**Connections:** Pigouvian taxation; tax incidence; innovation policy; distributive justice; externalities; technology pacing

**Record:** `open-questions-ai-safety-p30` · `machine-drafted-source-checked`

## 31. AI safety research must address environmental costs and injuries alongside capture, self-regulation, market governance, political trust, and changes to governing institutions

**Location:** Environmental Law and Political Economy, unpaginated online source

Professor Yonathan A. Arbel and his coauthors claim, in “Open Questions in Law and AI Safety: An Emerging Research Agenda,” an unpaginated Lawfare essay, under “Environmental Law” and “Political Economy,” that the agenda extends from the physical burdens of training and operating models to AI-directed or autonomous environmental injury. It also asks how AI will reshape governance, invite regulatory capture, perform under self-regulation or market discipline, and affect political trust. This is significant because safety includes ecological and institutional conditions, not only immediate injury to individual users. It connects to energy and water use, environmental liability, regulatory capture, industry self-governance, market discipline, and democratic trust.

**Evidence anchor:** The source asks about training harms and autonomous environmental liability, then governance effects, capture, self-regulation, markets, and political trust.

**Boundary:** The authors pose distinct environmental and political-economy questions; this synthesis does not claim that they share one causal mechanism or settled remedy.

**Connections:** energy and water use; environmental liability; regulatory capture; industry self-governance; market discipline; democratic trust

**Record:** `open-questions-ai-safety-p31` · `machine-drafted-source-checked`

## 32. Innovation and corporate law should be studied as safety levers through patents, trade secrecy, alternative rewards, board oversight, executive incentives, limited liability, and agent transactions

**Location:** Intellectual Property Law and Corporate and Finance Law, unpaginated online source

Professor Yonathan A. Arbel and his coauthors claim, in “Open Questions in Law and AI Safety: An Emerging Research Agenda,” an unpaginated Lawfare essay, under “Intellectual Property Law” and “Corporate and Finance Law,” that background rules governing invention and enterprise can change safety incentives. The agenda asks whether patents or alternatives should reward positive safety externalities, when trade secrecy obstructs scrutiny, how boards and compensation should promote oversight, whether limited liability externalizes catastrophic losses, and how autonomous-agent transactions require disclosure or compensation. This is significant because much AI risk is produced inside firms before public enforcement begins. It connects to patent incentives, trade secrets, corporate governance, executive compensation, limited liability, and financial disclosure.

**Evidence anchor:** The IP and corporate-finance questions cover safety innovation, secrecy, alternative incentives, governance, compensation, liability, agent transactions, and disclosure.

**Boundary:** The agenda does not endorse abolishing trade secrecy or limited liability, granting patents, or recognizing autonomous agents as legal persons.

**Connections:** patent incentives; trade secrets; corporate governance; executive compensation; limited liability; financial disclosure

**Record:** `open-questions-ai-safety-p32` · `machine-drafted-source-checked`