A new legal toolkit maps AI actions to human and corporate responsibility: Operational Agency and its causal graph let courts trace who should be held accountable for autonomous systems without treating machines as persons.

Main Finding

The paper introduces Operational Agency (OA) — a permeable legal fiction and ex post evidentiary framework — plus the Operational Agency Graph (OAG), a causal-mapping tool, to trace and apportion human culpability for harms caused by partially autonomous, multi-agent AI systems. By treating observable operational features of AI (goal-directedness, predictive processing, safety architecture) as proxies for mens rea, foresight, and standard of care, OA/OAG enable courts and regulators to assign liability to developers, fine-tuners, deployers, and users without granting legal personhood to AI.

Key Points

• Problem diagnosis
  • Modern agentic AI systems (what the authors call “fluid agency”) can autonomously plan, spawn subagents, and execute multi-step workflows, creating causal and epistemic distance between human actors and harmful acts.
  • Existing legal doctrines (mens rea/actus reus, products liability, secondary/vicarious liability) break down because AI is non-personal property with high operational independence; human culpability can become too remote to satisfy traditional standards.
• Operational Agency (OA)
  • OA is a structured evidentiary fiction that treats certain measurable operational characteristics of an AI as proxies for legal mental states and duties:
    • Goal-directedness → proxy for intent/mens rea
    • Predictive processing → proxy for foresight/foreseeability
    • Safety architecture → proxy for standard of care (duty)
  • OA is explicitly non-anthropomorphic and does not confer personhood; it is a tool for revealing and attributing human choices shaping AI behavior.
• Operational Agency Graph (OAG)
  • OAG is a methodological visual/causal tool that embeds the OA proxies into a directed causal graph connecting human actors, organizations, and AI components/agents.
  • OAG operationalizes the tracing of causal influence and evidentiary weight to apportion liability among creators, integrators, deployers, and end-users.
• Doctrinal integration
  • OA/OAG are designed to supplement, not replace, existing doctrines: they draw on corporate criminal liability, the innocent-agent doctrine, secondary liability principles, and respondeat superior to make culpability attribution practicable in agentic-AI contexts.
• Demonstrations
  • The framework is applied across multiple case studies (CreatorBot–ScraperBot, autonomous vehicle collision, algorithmic housing discrimination, biased hiring algorithms, policing alerts, and algorithmic cartel/cartel-like pricing) showing how OA/OAG produce actionable evidentiary paths to human accountability.

Data & Methods

• Methodological approach
  • Doctrinal legal analysis and synthesis of existing liability doctrines (tort, criminal, civil rights, constitutional, antitrust).
  • Conceptual construction of OA as an evidentiary proxy framework and of OAG as a causal-mapping method.
  • Integration of technical literature on agentic AI (e.g., LLM-powered agents, tool use, ReAct-style architectures) to ground the operational proxies in observable system features.
• Evidence and demonstrations
  • Illustrative, lawyerly case studies spanning tort, civil rights, constitutional law, and antitrust to demonstrate application and contours of OA/OAG.
  • Use of precedent and statutory materials (e.g., Thaler v. Vidal; Sony; Napster; Perfect 10) to show where existing doctrines fail and where OA can plug evidentiary gaps.
• Not an empirical dataset
  • The paper is conceptual and doctrinal rather than quantitative: it does not report new econometric or experimental datasets but relies on legal reasoning, prior technical publications, and applied case examples to validate the framework.

Implications for AI Economics

• Incentives and internalizing externalities
  • OA/OAG can change expected liability exposures across the AI value chain, creating stronger incentives for developers and deployers to invest in robust safety architectures, monitoring, and design choices that reduce downstream harms.
  • By making liability attribution more practicable, OA reduces the moral-hazard/shielding effect of autonomy that previously insulated actors from downstream externalities.
• Product design, deployment, and firm strategy
  • Firms may shift towards architectures that make goal-directedness and predictive-process traces auditable (e.g., explainability, logging, provenance), because those features will affect apportionment of liability.
  • The framework favors design practices enabling supervision controls, approval gating, and documented safety layers — increasing demand for compliance tooling, auditing, and logging services.
• Insurance and risk pricing
  • More tractable attribution enables insurers to underwrite AI liability risks with finer granularity, potentially leading to new insurance products priced by developer practices (safety architecture quality, monitoring capabilities) and by role (cloud provider vs. downstream fine-tuner vs. integrator).
  • Expected liability costs will feed into product pricing, contracting terms (indemnities), and investment decisions, especially for smaller firms whose compliance costs could be proportionally higher.
• Market structure and competition
  • Larger incumbents with resources to build auditable safety stacks could enjoy a liability-reduction advantage; conversely, clearer liability rules may reduce regulatory uncertainty and lower barriers to entry for firms that can credibly certify compliance.
  • The need for provenance, logging, and auditing could spawn new third-party verification and certification markets, potentially concentrating power among certifiers.
• Antitrust and systemic economic harms
  • OAG makes it easier to trace coordinated or cascade effects in multi-agent systems (e.g., algorithmic pricing across platforms), supplying evidentiary paths for antitrust enforcement against human actors who orchestrated or materially contributed to cartel-like outcomes.
  • This could raise enforcement intensity and thereby alter equilibrium strategies for pricing algorithms and coordination risk.
• Research and empirical directions
  • OA/OAG point to measurable operational indicators (degree of goal-directed planning, depth of predictive modeling, scope of safety controls) that economists can operationalize to model firms’ liability exposure and optimal investment in safety.
  • Empirical work could estimate how liability-risk-adjusted returns affect R&D allocation, prices, diffusion of agentic AI, and the welfare trade-offs between innovation speed and social harm mitigation.
• Policy trade-offs
  • By enabling liability without personhood, OA/OAG offer a middle path that preserves incentives for accountability while avoiding the complex consequences of granting AI legal status.
  • However, compliance costs and litigation risk could disproportionately burden smaller firms, so policymakers should consider complementary measures (safe-harbors for demonstrable safety practices, subsidized auditing, or tiered liability regimes) to avoid stifling beneficial innovation.

Overall, OA/OAG provide a practicable, legally informed way to translate observable AI operational traits into liability-relevant signals — a development that will reshape incentives, risk allocation, market behavior, and the economics of designing and deploying agentic AI systems.