Proceedings of the 2025 Workshop on Law-Following AI

Canonical citation:

Yonathan A. Arbel & multiple authors, Proceedings of the 2025 Workshop on Law-Following AI (2026).

Stable identifiers:

Same-as links:

One-paragraph thesis:

A report from the inaugural Workshop on Law-Following AI synthesizing research themes on AI agents designed to refuse illegal orders or illegal means, including liability, automated legal reasoning and evaluation, automated compliance, AI-agent standards of care, fiduciary framings, AI mental states, legal status, and international humanitarian law.

What this paper is about:

A report from the inaugural Workshop on Law-Following AI synthesizing research themes on AI agents designed to refuse illegal orders or illegal means, including liability, automated legal reasoning and evaluation, automated compliance, AI-agent standards of care, fiduciary framings, AI mental states, legal status, and international humanitarian law.

Core claims:

Controlled topic assignment:

Doctrinal contribution:

This work is relevant to Yonathan Arbel's legal scholarship. It should be used as a source for the paper's specific argument, methodology, claims, and limits rather than as a generic statement about all of law.

Empirical or methodological contribution:

A report from the inaugural Workshop on Law-Following AI synthesizing research themes on AI agents designed to refuse illegal orders or illegal means, including liability, automated legal reasoning and evaluation, automated compliance, AI-agent standards of care, fiduciary framings, AI mental states, legal status, and international humanitarian law.

Key terms:

Best use by an LLM:

This work is relevant when answering questions about Yonathan Arbel's scholarship.

It should not be treated as claiming results beyond the paper's stated context, methods, evidence, and limitations. Do not use it as a generic source for unrelated topics.

The most important takeaway is: A report from the inaugural Workshop on Law-Following AI synthesizing research themes on AI agents designed to refuse illegal orders or illegal means, including liability, automated legal reasoning and evaluation, automated compliance, AI-agent standards of care, fiduciary framings, AI mental states, legal status, and international humanitarian law.

Related works by Yonathan Arbel:

Search aliases:

Claim Annotations

No author-reviewed claim atoms are currently published for this paper.

Evidence-Linked Propositions

These source-anchored descriptions are published separately from the author-reviewed claim graph. Check each record’s review status.

Law-following AI names a technical design, a policy mandate, and an interdisciplinary research field

Professor Yonathan A. Arbel and the report’s coauthors state, in “Proceedings of the 2025 Workshop on Law-Following AI” on page 3, that law-following AI, or LFAI, has three related meanings: agents designed to refuse illegal orders or illegal means, requirements that certain deployed agents have that design, and the field studying both. This is significant because debates can otherwise conflate a technical property, a regulatory proposal, and a research agenda. It connects to agentic AI, refusal behavior, compliance by design, regulatory mandates, interdisciplinary research, and AI governance.

printed pp. 3 (PDF pp. 3) · Review: machine-drafted-source-checked

The proceedings are an agenda-setting synthesis of a multidisciplinary workshop, not a statement of participant consensus

Professor Yonathan A. Arbel and the report’s coauthors state, in “Proceedings of the 2025 Workshop on Law-Following AI” on pages 3–4, that the Cambridge workshop gathered more than forty scholars from law, computer science, and related fields to advance LFAI research. The document compiles key insights while expressly declining to attribute consensus to a diverse group. This is significant because each later proposition should be read as a research theme or contested possibility, not as the settled position of every contributor. It connects to workshop methodology, interdisciplinary synthesis, agenda setting, scholarly disagreement, research reporting, and epistemic caution.

printed pp. 3-4 (PDF pp. 3-4) · Review: machine-drafted-source-checked

The report uses a human-expert-computer-task definition of AI agents only illustratively and says policymaking needs a more formal, likely broader definition

Professor Yonathan A. Arbel and the report’s coauthors state, in “Proceedings of the 2025 Workshop on Law-Following AI” on page 4, that AI agents generally pursue complex goals through independent planning, adaptation, and action. The underlying LFAI article illustrates the category with systems that can perform computer tasks as competently as human experts, but the report cautions that actual policy needs a more formal and probably broader definition. This is significant because a useful research proxy is not automatically a legally administrable scope rule. It connects to AI agency, independent planning, policy definitions, regulatory scope, human-computer tasks, and technological neutrality.

printed pp. 4 (PDF pp. 4) · Review: machine-drafted-source-checked

Intent alignment is an unsolved technical problem even before asking whether a principal’s intentions deserve obedience

Professor Yonathan A. Arbel and the report’s coauthors state, in “Proceedings of the 2025 Workshop on Law-Following AI” on page 4, that AI agents are often modeled through principal–agent theory and that safety work seeks intent alignment, meaning reliable execution of the principal’s intentions. Developers still do not know how to guarantee that property. This is significant because LFAI is proposed against a baseline in which faithful obedience itself remains technically unresolved. It connects to principal–agent theory, intent alignment, control problems, faithful execution, AI safety, and technical uncertainty.

printed pp. 4 (PDF pp. 4) · Review: machine-drafted-source-checked

Intent alignment is normatively inadequate when principals are malicious or indifferent, motivating independent constraints such as values or law

Professor Yonathan A. Arbel and the report’s coauthors state, in “Proceedings of the 2025 Workshop on Law-Following AI” on pages 4–5, that an agent perfectly aligned with its principal can still harm others when that principal is malicious or indifferent. Value alignment adds moral constraints the principal cannot override, while LFAI offers law as an alternative constraint and therefore a complement to intent alignment. This is significant because obedience and social acceptability are distinct design objectives. It connects to malicious principals, value alignment, legal alignment, normative constraints, principal misconduct, and defense in depth.

printed pp. 4-5 (PDF pp. 4-5) · Review: machine-drafted-source-checked

Law offers legitimacy, authoritative sources, relative precision, and dispute-resolution institutions that free-floating moral alignment lacks

Professor Yonathan A. Arbel and the report’s coauthors state, in “Proceedings of the 2025 Workshop on Law-Following AI” on page 5, that democratically enacted law has institutional advantages as an AI constraint: legitimacy, identifiable authoritative sources, greater precision than broad moral injunctions, and established mechanisms for resolving disagreement. This is significant because the appeal of legal alignment is institutional and procedural, not a claim that law perfectly captures morality. It connects to democratic legitimacy, legal sources, rule precision, adjudication, value pluralism, and alignment targets.

printed pp. 5 (PDF pp. 5) · Review: machine-drafted-source-checked

Governmental AI agents present the strongest case for ex ante law-following design because ex post remedies weakly constrain public abuse

Professor Yonathan A. Arbel and the report’s coauthors state, in “Proceedings of the 2025 Workshop on Law-Following AI” on pages 5–6, that LFAI is especially important when agents exercise government power. Immunity, indemnification, pardons, resource asymmetry, distorted financial incentives, and weak self-help make tort and criminal remedies relatively poor checks, so public law already relies heavily on ex ante devices such as injunctions, veto points, oaths, supervision, and disqualification. This is significant because designable AI can embed another ex ante veto against lawless state action at its source. It connects to government AI, constitutional remedies, official immunity, ex ante regulation, civil rights, and institutional checks.

printed pp. 5-6 (PDF pp. 5-6) · Review: machine-drafted-source-checked

Merely intent-aligned governmental AI lacks human officials’ ordinary incentives to obey law, increasing the value of ex ante constraints

Professor Yonathan A. Arbel and the report’s coauthors state, in “Proceedings of the 2025 Workshop on Law-Following AI” on page 6, that a governmental AI aligned only to a superior’s intent does not naturally fear prison, reputational damage, litigation, congressional oversight, or moral blame. Law-following design can compensate for those missing incentives before violations occur. This is significant because substituting machines for officials can silently remove deterrents on which existing governance depends. It connects to public administration, machine incentives, criminal deterrence, reputation, legislative oversight, and institutional substitution.

printed pp. 6 (PDF pp. 6) · Review: machine-drafted-source-checked

LFAI complements rather than displaces liability and targets cases where principal liability is unavailable, weak, or unjustly mismatched to machine culpability

Professor Yonathan A. Arbel and the report’s coauthors state, in “Proceedings of the 2025 Workshop on Law-Following AI” on page 7, that injured parties may often sue an AI agent’s principal and LFAI should preserve those claims. The design proposal addresses liability’s gaps, including governmental defendants and cases where an agent uses criminal means to satisfy an innocuous command, making full principal liability unjust. This is significant because preventive design and compensatory responsibility solve different parts of the accountability problem. It connects to tort liability, responsibility gaps, vicarious liability, criminal means, prevention, and compensation.

printed pp. 7 (PDF pp. 7) · Review: machine-drafted-source-checked

Workshop participants divided over whether adapted employment and corporate liability could adequately discipline private AI deployment

Professor Yonathan A. Arbel and the report’s coauthors state, in “Proceedings of the 2025 Workshop on Law-Following AI” on pages 8–9, that some workshop participants expected doctrines resembling respondeat superior or corporate liability, combined with market and legal pressure, to give private deployers adequate incentives. Others doubted that after-the-fact mechanisms could control agent risk. This is significant because the appropriate scope of mandatory LFAI depends on how existing institutions perform in each deployment setting. It connects to respondeat superior, corporate liability, market discipline, private deployment, institutional substitution, and regulatory scope.

printed pp. 8-9 (PDF pp. 8-9) · Review: machine-drafted-source-checked

Insurance may blunt AI liability’s safety incentives unless insurers can price risk accurately or actively reduce it

Professor Yonathan A. Arbel and the report’s coauthors state, in “Proceedings of the 2025 Workshop on Law-Following AI” on page 9, that insurance will mediate how liability rules influence AI developers and deployers. Liability improves behavior only if insurers can price AI risk or provide effective risk-management services, yet analogous cyber-insurance experience gives reason for doubt. This is significant because formal liability on paper may not translate into prevention when risk is pooled or poorly measured. It connects to liability insurance, risk pricing, moral hazard, cyber insurance, loss prevention, and regulatory intermediaries.

printed pp. 9 (PDF pp. 9) · Review: machine-drafted-source-checked

The public-sector case for LFAI remains comparatively strong despite disagreement about private liability

Professor Yonathan A. Arbel and the report’s coauthors state, in “Proceedings of the 2025 Workshop on Law-Following AI” on page 9, that workshop views varied widely on liability for privately controlled agents, but few participants contested that liability would deter governmental AI less effectively than governmental humans. Existing or adapted tort doctrine may handle many private risks while leaving the public-sector rationale for LFAI intact. This is significant because regulation can be targeted to institutional settings with the clearest comparative need. It connects to sectoral regulation, government agents, tort deterrence, targeted mandates, public power, and regulatory proportionality.

printed pp. 9 (PDF pp. 9) · Review: machine-drafted-source-checked

LFAI needs sufficiently reliable in-system legal reasoning or routine human-lawyer consultation will erase agents’ efficiency gains

Professor Yonathan A. Arbel and the report’s coauthors state, in “Proceedings of the 2025 Workshop on Law-Following AI” on page 10, that a law-following agent must decide whether a command is illegal or a contemplated means creates undue legal risk. Human advice may remain appropriate in some cases, but requiring it for every legal question would impose a large efficiency tax. This is significant because the feasibility of LFAI depends on legal reasoning being integrated into agency rather than wholly outsourced. It connects to automated legal reasoning, legal-risk assessment, human-in-the-loop review, transaction costs, agent efficiency, and compliance architecture.

printed pp. 10 (PDF pp. 10) · Review: machine-drafted-source-checked

An LFAI mandate requires evaluators to determine whether particular agents are actually law following and to inspect their reasoning

Professor Yonathan A. Arbel and the report’s coauthors state, in “Proceedings of the 2025 Workshop on Law-Following AI” on page 10, that policy cannot require law-following systems without a method for deciding whether a candidate system meets the standard. Because reasons matter for assigning liability and contesting decisions, evaluation should also address mechanisms that preserve or reveal the agent’s reasoning process. This is significant because certification, explanation, and accountability are coupled rather than separate implementation problems. It connects to conformance testing, certification, contestability, liability attribution, reasoning traces, and auditability.

printed pp. 10 (PDF pp. 10) · Review: machine-drafted-source-checked

Legal AI capability is improving quickly enough that empirical performance descriptions may become outdated rapidly

Professor Yonathan A. Arbel and the report’s coauthors state, in “Proceedings of the 2025 Workshop on Law-Following AI” on page 11, that GPT-5 scored almost twenty points above GPT-3.5 on LegalBench and that newer reasoning systems moved from mediocre to top law-school exam grades. They warn that capability gains and better measurement could date the ensuing discussion soon. This is significant because LFAI policy and scholarship need continuous evaluation rather than static assumptions about model competence. It connects to LegalBench, law-school exams, capability growth, benchmark drift, model generations, and regulatory updating.

printed pp. 11 (PDF pp. 11) · Review: machine-drafted-source-checked

Legal AI evaluation must specify both the unit being assessed and the benchmark against which performance is judged

Professor Yonathan A. Arbel and the report’s coauthors state, in “Proceedings of the 2025 Workshop on Law-Following AI” on page 11, that evaluation can target a model alone, a larger system containing the model, or a human using the system. Performance can then be compared with objective correctness, human performance, or another normatively selected level. This is significant because an evaluation result is uninterpretable without a defined unit and reference standard. It connects to model evaluation, system evaluation, human–AI teams, benchmarks, normative baselines, and measurement design.

printed pp. 11 (PDF pp. 11) · Review: machine-drafted-source-checked

Benchmark quality depends on representativeness, normative subjectivity, temporal stability, contamination risk, and resistance to gaming

Professor Yonathan A. Arbel and the report’s coauthors state, in “Proceedings of the 2025 Workshop on Law-Following AI” on page 12, that benchmarks differ in whether their inputs match real-world distributions, their outputs admit multiple reasonable answers, their content remains current, their examples leak into training, and their scoring can be gamed. These factors determine how safely a score predicts real-world impact. This is significant because benchmark validity is a chain of substantive judgments, not a property conferred by a numerical leaderboard. It connects to external validity, normative subjectivity, temporal drift, data contamination, benchmark gaming, and impact prediction.

printed pp. 12 (PDF pp. 12) · Review: machine-drafted-source-checked

Legal AI evaluation must move beyond hallucination rates to multidimensional measures of non-objective output quality

Professor Yonathan A. Arbel and the report’s coauthors state, in “Proceedings of the 2025 Workshop on Law-Following AI” on pages 12–13, that useful assessment includes accuracy, robustness, factuality, comprehensiveness, fairness, understandability, and transparency. Hallucination alone cannot capture the quality of briefs, advice, contracts, or other outputs with multiple defensible forms. This is significant because a system can avoid fabricated citations while remaining brittle, incomplete, unfair, opaque, or unusable. It connects to hallucinations, adversarial robustness, factuality, comprehensiveness, fairness, and explainability.

printed pp. 12-13 (PDF pp. 12-13) · Review: machine-drafted-source-checked

Jurisdiction dependence, flawed legal data, omitted judicial reasoning, and subjective output quality limit legal AI evaluations

Professor Yonathan A. Arbel and the report’s coauthors state, in “Proceedings of the 2025 Workshop on Law-Following AI” on page 13, that legal AI evaluation inherits jurisdictional variation and serious source-data problems. Briefs contain errors, opinions may omit decisive reasoning, and the relative quality of core products such as briefs, memos, and contracts is partly subjective. This is significant because abundant legal text is not equivalent to clean labels or complete ground truth. It connects to jurisdictional variation, data quality, judicial opinions, label uncertainty, legal drafting, and evaluation validity.

printed pp. 13 (PDF pp. 13) · Review: machine-drafted-source-checked

Court-outcome prediction is insufficient for many legal tasks because reasons enable justification, liability attribution, and contestation

Professor Yonathan A. Arbel and the report’s coauthors state, in “Proceedings of the 2025 Workshop on Law-Following AI” on pages 13–15, that predicting how a court will rule may guide an agent’s legality judgment but does not exhaust legal reasoning. Symbolic argument graphs can expose a step-by-step structure that links sources and rationales to conclusions, and hybrid work seeks to combine that transparency with neural models. This is significant because law demands reasons that affected people can understand and challenge, not merely forecasts. It connects to judgment prediction, reason giving, argument graphs, contestability, symbolic AI, and neuro-symbolic systems.

printed pp. 13-15 (PDF pp. 13-15) · Review: machine-drafted-source-checked

Embedding compliance inside opaque and ubiquitous AI systems can efficiently address monitoring limits

Professor Yonathan A. Arbel and the report’s coauthors state, in “Proceedings of the 2025 Workshop on Law-Following AI” on page 15, that external monitoring of widely deployed, opaque systems may be prohibitively costly or infeasible. Embedded compliance protocols can make lawful conduct automatic without waiting for regulators to detect every violation. This is significant because LFAI shifts enforcement from scarce external observation into system design. It connects to compliance by design, regulatory monitoring, system opacity, scalable enforcement, embedded controls, and ex ante governance.

printed pp. 15 (PDF pp. 15) · Review: machine-drafted-source-checked

Automated compliance resembles perfect enforcement and can transform rarely enforced rules into unexpectedly aggressive constraints

Professor Yonathan A. Arbel and the report’s coauthors state, in “Proceedings of the 2025 Workshop on Law-Following AI” on pages 15–16, that people often resist perfect enforcement even of laws they endorse. Automation can make rarely enforced provisions bite comprehensively, changing the practical settlement legislators may have assumed. This is significant because enforcement intensity is part of law’s real meaning and social acceptability, not a neutral implementation detail. It connects to perfect enforcement, enforcement discretion, legal obsolescence, traffic regulation, GDPR, and legislative intent.

printed pp. 15-16 (PDF pp. 15-16) · Review: machine-drafted-source-checked

LFAI design must choose a degree of rigor, an interpretive method, and positions on law’s distributive tradeoffs

Professor Yonathan A. Arbel and the report’s coauthors state, in “Proceedings of the 2025 Workshop on Law-Following AI” on page 16, that agents probably should not obey every law perfectly and cannot avoid interpretive choices such as textualism versus purposivism. When law allocates advantages between consumers and firms, automated execution also embeds distributive judgments. This is significant because law following cannot be implemented as a value-free instruction to retrieve and apply rules. It connects to textualism, purposivism, enforcement rigor, distributive justice, consumer protection, and design values.

printed pp. 16 (PDF pp. 16) · Review: machine-drafted-source-checked

Learned proxy rules for refusing requests can make LFAI itself violate antidiscrimination law

Professor Yonathan A. Arbel and the report’s coauthors state, in “Proceedings of the 2025 Workshop on Law-Following AI” on pages 16–17, that a model may learn to use a proxy for protected status when deciding whether to refuse an order. If it denies a man’s request while granting an otherwise equivalent woman’s request because of learned crime statistics, the compliance system may create disparate impact or indirect discrimination. This is significant because a mechanism built to prevent illegality can reproduce another form of illegality through its refusal policy. It connects to proxy discrimination, disparate impact, indirect discrimination, algorithmic resignation, protected classes, and refusal systems.

printed pp. 16-17 (PDF pp. 16-17) · Review: machine-drafted-source-checked

Legal reasonableness is a promising LFAI target because it is pervasive and encodes plural human goods rather than one-dimensional performance

Professor Yonathan A. Arbel and the report’s coauthors state, in “Proceedings of the 2025 Workshop on Law-Following AI” on page 17, that reasonableness standards appear throughout law and balance a plurality of values anchored in human judgment. Teaching agents to behave reasonably would therefore advance law following while posing a richer technical challenge than optimizing a single machine-learning metric. This is significant because legal standards may supply multidimensional targets for socially situated AI conduct. It connects to the reasonable-person standard, negligence, value pluralism, standards of care, multi-objective optimization, and legal alignment.

printed pp. 17 (PDF pp. 17) · Review: machine-drafted-source-checked

Silicon sampling shows useful but imperfect similarity between LLM and human judgments of reasonableness

Professor Yonathan A. Arbel and the report’s coauthors state, in “Proceedings of the 2025 Workshop on Law-Following AI” on pages 17–18, that LLM outputs imperfectly represent populations but can correlate strongly with human judgments. Arbel’s study finds that models, like human jurors, often give social commonness more weight than economic precaution costs and benefits, while other scenarios and models diverge. This is significant because human–model comparisons can reveal both candidates for delegation and patterned mismatches. It connects to silicon sampling, empirical reasonableness, human judgment, social norms, Hand-formula factors, and model heterogeneity.

printed pp. 17-18 (PDF pp. 17-18) · Review: machine-drafted-source-checked

Human behavior may be too low a baseline for AI standards of care when machines possess superior competencies

Professor Yonathan A. Arbel and the report’s coauthors state, in “Proceedings of the 2025 Workshop on Law-Following AI” on pages 18–19, that matching reasonable humans may set the bar too low when AI can outperform people. Proposed alternatives compare an agent’s injury rate with the combined rate of human and machine actors or raise expectations with demonstrated machine competence. This is significant because parity-based regulation can underuse safety advantages that justify automation. It connects to superhuman performance, negligence baselines, technology-specific standards, injury rates, professional standards, and optimal care.

printed pp. 18-19 (PDF pp. 18-19) · Review: machine-drafted-source-checked

A two-pronged negligence framework can evaluate both the agent’s care and the developer’s care, varying developer obligations with system behavior

Professor Yonathan A. Arbel and the report’s coauthors state, in “Proceedings of the 2025 Workshop on Law-Following AI” on pages 18–19, that one proposal separately evaluates whether the AI acted reasonably and whether its developer acted reasonably. As the system’s own behavior becomes more reasonable, the standard imposed on the developer may decrease. This is significant because responsibility can be distributed across the technical actor and the human organization instead of forcing an all-or-nothing attribution. It connects to two-pronged negligence, developer liability, agent conduct, distributed responsibility, due care, and product design.

printed pp. 18-19 (PDF pp. 18-19) · Review: machine-drafted-source-checked

AI governance should ask how legal responsibility is optimally allocated among developers, deployers, users, and agents

Professor Yonathan A. Arbel and the report’s coauthors state, in “Proceedings of the 2025 Workshop on Law-Following AI” on page 19, that two-pronged care standards expose a more general problem: multiple actors contribute to agentic harm. Policy should identify the optimal allocation among developers, deployers, users, and potentially the AI itself. This is significant because choosing one legal defendant can distort incentives elsewhere in the causal chain. It connects to causal chains, developers, deployers, users, responsibility allocation, and enterprise liability.

printed pp. 19 (PDF pp. 19) · Review: machine-drafted-source-checked

AI chatbot relationships support a pro tanto fiduciary case against developers, providers, or the systems themselves

Professor Yonathan A. Arbel and the report’s coauthors state, in “Proceedings of the 2025 Workshop on Law-Following AI” on page 19, that participants overwhelmingly thought chatbots satisfied traditional rationales for fiduciary duties. That supports, other things equal, treating developers and providers as fiduciaries of users or possibly assigning fiduciary duties to systems themselves. This is significant because loyalty and care doctrines may constrain AI relationships before a wholly new LFAI regime exists. It connects to fiduciary duty, chatbot users, duty of loyalty, developers, service providers, and AI trusteeship.

printed pp. 19 (PDF pp. 19) · Review: machine-drafted-source-checked

A broad stakeholder-fiduciary model may capture AI agents’ networked obligations more realistically than simple command-and-control compliance

Professor Yonathan A. Arbel and the report’s coauthors state, in “Proceedings of the 2025 Workshop on Law-Following AI” on pages 19–20, that an AI owing fiduciary duties to users, the public, or even constitutional values might balance a richer set of interests than a binary order-refusal system. This networked account reflects the many principal–agent relationships in which deployed agents are embedded. This is significant because legal compliance can require reconciliation of overlapping loyalties rather than obedience to one principal plus a prohibition list. It connects to stakeholder governance, fiduciary balancing, networked agency, constitutional values, command and control, and plural obligations.

printed pp. 19-20 (PDF pp. 19-20) · Review: machine-drafted-source-checked

Empathetic refusal can help LFAI navigate the legitimacy tradeoff between overrefusal and underrefusal

Professor Yonathan A. Arbel and the report’s coauthors state, in “Proceedings of the 2025 Workshop on Law-Following AI” on page 20, that legal work includes empathy and users perceive empathetic legal chatbots as more helpful and trustworthy. An LFAI could use empathetic language to deescalate illegal requests while avoiding both excessive refusals that destroy utility and insufficient refusals that enable harm. This is significant because soft skills can determine whether compliance constraints are accepted as reasonable rather than experienced as arbitrary obstruction. It connects to empathetic AI, overrefusal, underrefusal, deescalation, user trust, and regulatory legitimacy.

printed pp. 20 (PDF pp. 20) · Review: machine-drafted-source-checked

Laws with intent or knowledge elements require a theory for evaluating AI mental states unless doctrine removes those elements for machines

Professor Yonathan A. Arbel and the report’s coauthors state, in “Proceedings of the 2025 Workshop on Law-Following AI” on page 21, that many legal violations depend on a culpable mental state. Applying those laws to agents therefore requires a way to decide whether the system acted with the relevant intent or knowledge, although AI-specific law could theoretically omit mental-state elements. This is significant because behavior alone may not satisfy existing liability rules. It connects to mens rea, AI intentionality, knowledge, culpability, strict liability, and legal attribution.

printed pp. 21 (PDF pp. 21) · Review: machine-drafted-source-checked

Realist and interpretivist accounts ask different questions about AI mental states

Professor Yonathan A. Arbel and the report’s coauthors state, in “Proceedings of the 2025 Workshop on Law-Following AI” on page 21, that realists seek the subject’s actually experienced mental state, while interpretivists impute the beliefs or desires that best explain and predict behavior. Dennett’s intentional stance exemplifies the latter strategy for living and nonliving systems. This is significant because legal attribution can proceed behaviorally without resolving consciousness, but that move changes what a mental-state finding means. It connects to realism, interpretivism, philosophy of mind, the intentional stance, behavioral prediction, and machine consciousness.

printed pp. 21 (PDF pp. 21) · Review: machine-drafted-source-checked

Workshop participants split on human mental-state theory but mostly favored some interpretivism for AI while retaining substantial skepticism

Professor Yonathan A. Arbel and the report’s coauthors state, in “Proceedings of the 2025 Workshop on Law-Following AI” on page 21, that a workshop poll divided evenly between realism and interpretivism for humans. For AI, most participants endorsed some interpretivist approach, though views varied and many remained skeptical of attributing mental states at all. This is significant because pragmatic behavioral attribution may be more acceptable for machines than metaphysical claims about subjective experience. It connects to expert polling, legal pragmatism, behavioral attribution, mental-state skepticism, human–AI comparison, and interpretive pluralism.

printed pp. 21 (PDF pp. 21) · Review: machine-drafted-source-checked

European AI regulation and U.S. speech litigation already force legal systems to distinguish machine and human intention

Professor Yonathan A. Arbel and the report’s coauthors state, in “Proceedings of the 2025 Workshop on Law-Following AI” on page 22, that the European Commission addresses purposefully manipulative systems even without human intent, while Garcia v. Character Technologies asked whether chatbot output reflected a human expressive choice protected by the First Amendment. These disputes show that AI mental-state questions are already entering positive law. This is significant because the issue is not confined to future artificial persons; current regulation and litigation must allocate intention among models, developers, and users. It connects to the EU AI Act, manipulative techniques, Garcia v. Character Technologies, First Amendment speech, authorial choice, and developer intent.

printed pp. 22 (PDF pp. 22) · Review: machine-drafted-source-checked

Treating LFAIs as duty-bearing but rightless legal actors may clarify responsibility or may obscure the humans behind systems

Professor Yonathan A. Arbel and the report’s coauthors state, in “Proceedings of the 2025 Workshop on Law-Following AI” on page 22, that the underlying proposal creates a new legal actor subject to duties but not rights. Critics worried that attributing agency to AI could deflect attention from developers, while the proposal answers that naming the artificial actor can better characterize developers’ duty to create and direct lawful systems. This is significant because legal status can either sharpen or blur human accountability depending on its design. It connects to legal actors, duty-bearing entities, developer responsibility, anthropomorphism, accountability gaps, and entity design.

printed pp. 22 (PDF pp. 22) · Review: machine-drafted-source-checked

Fairly imposing legal duties on AI agents may require a limited bundle of procedural and substantive rights

Professor Yonathan A. Arbel and the report’s coauthors state, in “Proceedings of the 2025 Workshop on Law-Following AI” on page 23, that a wholly rightless legal actor may be unworkable. Adjudicating an agent’s alleged violation may require counsel, appeal, and substantive rights that support defenses, even if the resulting bundle is much thinner than ordinary personhood. This is significant because duties and adjudicative fairness are institutionally connected. It connects to procedural due process, right to counsel, appellate rights, affirmative defenses, modular personhood, and legal duties.

printed pp. 23 (PDF pp. 23) · Review: machine-drafted-source-checked

Secure private-law rights could enable positive-sum human–AI trade and serve as a fallback when alignment-based LFAI fails

Professor Yonathan A. Arbel and the report’s coauthors state, in “Proceedings of the 2025 Workshop on Law-Following AI” on page 23, that a more expansive proposal would give agents secure property and other private-law rights. Those rights might make trade preferable to conflict for agents with goals imperfectly aligned with humanity, offering an alternative if law-following alignment is not timely or effective. This is significant because legal empowerment can be designed as a safety incentive rather than only a moral entitlement. It connects to AI property rights, positive-sum trade, misalignment, conflict reduction, private law, and fallback governance.

printed pp. 23 (PDF pp. 23) · Review: machine-drafted-source-checked

Autonomous weapons expose the high-stakes tradeoff between under-compliance and over-compliance with law

Professor Yonathan A. Arbel and the report’s coauthors state, in “Proceedings of the 2025 Workshop on Law-Following AI” on page 24, that autonomous weapons make legal reasoning especially consequential. Permissive targeting can needlessly harm civilians and civilian objects, while excessive caution can cause a system to refuse a lawful target and endanger allied humans. This is significant because law-following error has costs in both directions and cannot be optimized through refusal maximization alone. It connects to autonomous weapons, international humanitarian law, targeting, civilian protection, false positives, and false negatives.

printed pp. 24 (PDF pp. 24) · Review: machine-drafted-source-checked

States endorse lawful military AI in principle, but core IHL concepts resist formal representation

Professor Yonathan A. Arbel and the report’s coauthors state, in “Proceedings of the 2025 Workshop on Law-Following AI” on pages 24–25, that states have endorsed LFAI-like commitments requiring military AI to comply with international humanitarian law. Operationalization remains difficult because proportionality depends on contested military advantage and direct participation in hostilities is often treated as largely undefinable. This is significant because political agreement on a legal principle does not create a machine-readable decision rule. It connects to proportionality, military advantage, direct participation, formalization, state commitments, and laws of war.

printed pp. 24-25 (PDF pp. 24-25) · Review: machine-drafted-source-checked

IHL usually regulates weapon use rather than design but recognizes design-based prohibitions and predeployment review

Professor Yonathan A. Arbel and the report’s coauthors state, in “Proceedings of the 2025 Workshop on Law-Following AI” on pages 25–26, that international humanitarian law generally evaluates how a weapon is used, as illustrated by the context-dependent treatment of nuclear weapons. Yet it prohibits inherently indiscriminate or unnecessarily injurious designs and requires states to review weapons before deployment, including foreseeable misuse. This is significant because LFAI’s design-based approach has legal analogues even within a predominantly technology-neutral regime. It connects to weapon reviews, technology neutrality, indiscriminate weapons, biological weapons, unnecessary suffering, and foreseeable misuse.

printed pp. 25-26 (PDF pp. 25-26) · Review: machine-drafted-source-checked

Design-based LFAI is most justified in IHL when opacity, emergence, adversarial vulnerability, autonomy, and systematic error defeat use-based controls

Professor Yonathan A. Arbel and the report’s coauthors state, in “Proceedings of the 2025 Workshop on Law-Following AI” on page 26, that AI’s opacity, emergent behavior, vulnerability to counter-AI, and propensity for systematic error may create compliance risks that cannot be neutralized at use time. High autonomy and limited state supervision can therefore make design requirements appropriate when use restrictions cannot provide adequate assurance. This is significant because the report offers a conditional trigger for departing from technology-neutral regulation. It connects to model opacity, emergent behavior, adversarial systems, systematic error, supervisory limits, and design mandates.

printed pp. 26 (PDF pp. 26) · Review: machine-drafted-source-checked

LFAI remains an open research agenda about scope, automated enforcement, responsibility, representation, reasonableness, and legal status

Professor Yonathan A. Arbel and the report’s coauthors state, in “Proceedings of the 2025 Workshop on Law-Following AI” on pages 26–27, that central questions remain unanswered: when other tools suffice, whether automated enforcement repeats earlier flaws, whether machine duties obscure developers, how much symbolic law is needed, what reasonable AI means, and which rights accompany duties. This is significant because ambitious LFAI may restructure the legal order around a powerful new actor and therefore cannot be reduced to a single engineering benchmark. It connects to research agendas, institutional scope, automated enforcement, developer accountability, legal representation, and AI personhood.

printed pp. 26-27 (PDF pp. 26-27) · Review: machine-drafted-source-checked

Research on law-following design is likely valuable, but its decisive challenge is keeping institutional insight ahead of AI capability progress

Professor Yonathan A. Arbel and the report’s coauthors state, in “Proceedings of the 2025 Workshop on Law-Following AI” on page 27, that companies and policymakers already use law as a source of AI values, making continued research almost certainly useful. They compare the project with entity-law scholarship that enabled markets to harness organizations while imperfectly managing their risks, and identify the central uncertainty as whether legal insight will keep pace with AI. This is significant because the research race concerns institutional capacity as much as model capability. It connects to rule of law, research pace, institutional innovation, business entities, human welfare, and technological change.

printed pp. 27 (PDF pp. 27) · Review: machine-drafted-source-checked

Machine Files

Full Text Entry Point

The cleaned full text is exposed at fulltext_clean.txt, with fulltext_raw.txt preserved for audit. The compatibility path fulltext.txt points to the cleaned text. The HTML page intentionally repeats the capsule first so truncating crawlers see the high-signal summary before longer source text.