Intention, But Hybrid: A New Test for Posthuman Agents

Abstract

Lawmakers and courts are already confronting real harms and disputed evidence that flow from human-AI systems used in clinics, public transport and policing. Without clear rules we risk either over-attribution of guilt to the nearest human or under-attribution by blaming the machine. Both outcomes corrode legality and public trust. This paper offers a court-usable framework that preserves anthropocentric culpability while closing responsibility gaps that appear once adaptive systems shape human action.

The proposed concept of hybrid intention captures cases where a person initiates a purposive course of conduct, the realisation of that purpose is constitutively mediated by an artificial system that contributes more than a mere tool, and a limited yet assessable veto or control window is available to the human agent. From this structure follows a three-part attribution test that aligns with continental legality and guilt, refines agency, causation and culpability for distributed control, and avoids personifying machines. The test directs courts to examine initiation, technological mediation and the existence and use of a control window, drawing on evidence of foreseeability, traceability and realistic opportunities to intervene.

Moreover, the proposed test is supplemented by a set of operational criteria that help judges allocate criminal responsibility for offences involving human-AI systems. First comes explainability of the specific output so decision makers understand what the system found and why. Second comes prior validation with recorded version and configuration so reliability is proven before use and reproducible after the fact. Third comes complete logging so that inputs, processing, human interventions and outputs form an auditable chain from sensor to model state to actuation. These elements give investigators artefacts to collect, allow prosecutors to lay a proper foundation and let judges reliably examine human-AI system output.

The paper develops the test through a detailed hypothetical in surgical AI. A hospital deploys an AI-assisted system for laparoscopic procedures. During a standard operation the system proposes a manoeuvre based on real-time imaging and the action causes a vascular injury. Initiation is satisfied because the clinical team selected and activated the system for a defined therapeutic goal with known capabilities and risks. Constitutive mediation is present because adaptive processing generated proposals that materially shaped the act under uncertainty. Attribution then turns on the control window. If logs show a realistic override pathway, adequate training and alarms, and a clear alert that was ignored, individual culpability attaches within familiar categories. If due care and training were observed and no meaningful override existed in the relevant seconds, responsibility shifts from individual guilt toward organisational and system-level duties with documentation, validation and incident-response obligations carrying the weight.

The same decision rule generalises to neuroprosthetic control. Initiation lies in informed adoption and calibration. Constitutive mediation stems from the decoding and adaptive control loop that turns neural signals into actuation. The control window shifts upstream to configuration, safety locks and emergency disengagement. Where a realistic kill switch or safety regime existed but was negligently disabled, culpability individualises. Where no meaningful intervention was possible despite due care, accountability migrates to organisational duties in design, procurement, training, monitoring and incident response.

As deployment outpaces doctrine, courts need evidence-based criteria to admit or assign probative weight to AI-derived outputs, based on explainability, prior validation and complete logging. The framework supplies evidence-ready indicators that map to current governance duties. Foreseeability is grounded in documented capability envelopes, validation and known failure modes. Traceability is grounded in mandatory logging that ties sensor data, model states, thresholds, alerts and actions into a reproducible chain. The control window rests on interface salience, measured override latency and realistic single-action reach for intervention. Thresholds are defined ex-ante and scaled to risk, and all alerts and interventions are timestamped from alert onset to effective halt and recorded in tamper-evident logs. Together these criteria keep guilt with people where it belongs, allocate criminal liability to legal persons and responsible persons within them when design or oversight fail, and deter both speculative claims about machine agency and strict liability by default.

The contribution is immediate. It gives courts a compact rule for deciding who intended what inside hybrid control loops, offers regulators compliance-by-design guidance that improves safety without freezing innovation, and helps developers build systems whose outputs are intelligible and testable in legal forums. The framework is technology-neutral, usable across surgical AI, autonomous driving, defence and brain-computer interfaces, and it works under current law while remaining compatible with risk-based governance.