AI do not qualify as criminal agents: machines lack the mental and normative capacities that ground mens rea and punishability, so accountability should rest with people and firms. Redirecting liability onto human actors concentrates incentives for safer design, affects compliance costs and insurance markets, and will shape innovation and market structure.

Main Finding

Applying the criminal-law concept of "agent" to current and foreseeable AI systems is a category mistake. No existing AI meets the cluster of capacities required for criminal responsibility or for legitimately bearing punishment. Redirecting attention away from imagining AI as criminal agents toward the human decisions, duties, and failures that cause AI-related harms is both conceptually correct and urgently important for designing effective regulation and accountability regimes.

Key Points

• The debate over AI and criminal responsibility is typically framed as whether AI can be a criminal law agent (a debate between denialists and inclusionists). This paper rejects that framing as mistaken rather than debating its specific conclusions.
• Why agentic framing is tempting:
  • Cognitive: humans readily anthropomorphize systems that exhibit goal-directed or humanlike behavior.
  • Linguistic: shorthand and legal idioms (e.g., “the system decided”) encourage reifying agency.
  • Legal: existing legal categories (e.g., corporate personhood) make it plausible to try to slot AI into agency roles.
  • Philosophical: the way we talk about responsibility and agency invites extension to novel entities unless we scrutinize category boundaries.
• Conditions for criminal-law agency (drawn from responsibility and punishment literature). An entity must, among other things:
  • Exercise relevant control over actions (not merely causally implicated but capable of intentional action in a legally salient sense).
  • Possess mental states that ground mens rea (capacity for intent, knowledge, recklessness, or culpable negligence understood in morally relevant terms).
  • Be reasons-responsive in a morally relevant way (capable of understanding and responding to moral and legal reasons).
  • Be capable of bearing punishment in ways that make retributive or consequentialist justification intelligible (e.g., deterrence, rehabilitation, desert).
  • Be appropriate objects of moral blame (capable of being morally motivated and normatively accountable).
• Why current and foreseeable AI fail these conditions:
  • No evidence AI has subjective understanding, moral consciousness, or genuinely normative reasons-responsiveness.
  • AI systems operate via algorithmic procedures and statistical optimization, lacking the kind of agency that grounds mens rea.
  • Punishment-dependent rationales (retribution, deterrence) cannot plausibly be applied to machines lacking capacities that punishment targets.
  • Treating AI as criminal agents obscures, rather than clarifies, where moral and legal responsibilities actually reside: with humans—developers, deployers, managers, regulators.
• Shifting focus to humans clarifies responsibility: design choices, deployment decisions, maintenance, oversight, regulatory failures, and corporate policies are the appropriate loci for criminal and regulatory accountability.

Data & Methods

• Conceptual and normative analysis: the paper relies on philosophical analysis of agency, responsibility, and punishment concepts (including discussion of “category mistakes”).
• Legal doctrinal analysis: examination of criminal-law doctrines and the constitutive elements of criminal liability (actus reus, mens rea, excuses/defenses, and aims of punishment).
• Literature synthesis: review of interdisciplinary literature spanning AI capabilities, legal scholarship on AI responsibility, and philosophical debates about moral agency.
• Argumentative structure: three-part reasoning—(1) diagnose why agentic framing persists, (2) articulate necessary conditions for criminal agency and show AI fails them, (3) draw regulatory and normative consequences for redirecting responsibility to humans.
• Note: the paper does not present empirical testing or novel quantitative data; it is a theoretical/legal-philosophical intervention aimed at reframing policy and doctrinal thinking.

Implications for AI Economics

• Liability allocation and incentives:
  • Treating AI as non-criminal agents concentrates legal exposure on humans and organizations (designers, deployers, operators). Economically, this creates clearer incentives for firms to invest in safe design, monitoring, and compliance because they (not the machine) bear legal costs.
  • Misattributing agency to AI risks diffusing responsibility and weakening incentives for firms to internalize harm externalities (moral hazard).
• Regulatory design and compliance costs:
  • A human-centered liability regime implies compliance costs fall on firms (personnel, safety engineering, audits, insurance). Smaller firms may face disproportionate burdens, potentially favoring incumbents and increasing market concentration.
  • Policymakers should anticipate increased demand for services: compliance consulting, third-party audits, certification, and liability insurance markets.
• Innovation incentives:
  • Clear, targeted liability rules that hold humans/organizations accountable while providing predictable standards encourage investments in safety-enhancing R&D.
  • Conversely, ambiguous or symbolic attribution of agency to AI could chill innovation (if firms fear unpredictable prosecution) or encourage underinvestment in governance (if firms think machines will absorb blame).
• Insurance and risk pricing:
  • Insurers will price in human and organizational risk factors (procurement choices, testing practices, deployment contexts) rather than treating systems as independent risk-bearers. This enables more granular risk pricing and underwriting linked to governance quality.
• Policy recommendations with economic effects:
  • Prioritize human-focused criminal and civil liability (negligence, recklessness, corporate liability) to ensure harms are internalized and deterrence channels remain effective.
  • Mandate transparency, logging, and auditability (reduces information asymmetries, lowers monitoring costs, and improves actuarial assessment).
  • Encourage or require safety standards and certifications (creates markets for compliance, can reduce uncertainty and lower liability costs).
  • Support liability insurance markets and possibly targeted subsidies for safety investments for smaller firms to avoid concentration effects.
  • Calibrate sanctions to deter negligent behavior without excessively chilling innovation: mix criminal liability (for severe, culpable human conduct) with civil/regulatory remedies and administrative fines for broader compliance enforcement.
• Broader economic externalities:
  • Properly attributing responsibility to humans helps internalize negative externalities of AI deployment, reducing social costs from harms.
  • It can also channel investments toward governance-capacity building (training, hiring legal/compliance staff), influencing labor demand and firm organization.

Overall, the paper implies that economic policy should aim to align incentives so that parties with control over design and deployment bear legal and financial accountability. That alignment encourages firms to invest in safety, improves risk allocation through insurance and markets, and avoids the perverse effects of treating machines as legal moral agents.