Acting with AI: An Interaction-Based Framework for Agentic Tort Liability

Researchers propose a legal framework for allocating tort liability when autonomous AI systems cause harm, distinguishing between pure tool use, collaborative planning, and autonomous drift scenarios. The framework draws on human concerted action law and uses interaction logs as evidence to determine where responsibility attaches between users and developers.

The paper addresses a fundamental gap in tort law: existing frameworks struggle to assign liability when AI systems cause harm through paths neither fully chosen by users nor specifically foreseen by developers. This matters because agentic AI—systems that plan autonomously, use tools, and execute tasks over extended periods—operates in a legal grey zone where traditional product liability and negligence doctrines fail to capture the human-AI interaction dynamic.

The authors build their framework on Michael Bratman's planning theory and common law principles governing human-human concerted action, creating three interaction categories. Pure tool use remains straightforward: ordinary product-defect doctrine applies. Collaborative planning mirrors independent contractor liability and professional malpractice standards. Autonomous drift, where AI deviates from its authorized scope, invokes respondeat superior and strict product liability. This taxonomy reflects how responsibility shifts based on the degree of human control and foresight at each stage.

For the AI and emerging technology industry, this framework has significant implications. Developers face pressure to implement robust constraint verification, epistemic transparency, and forensic logging—features the authors propose as a "Reasonable Agent" standard. Companies deploying agentic systems must now consider stateful interaction logs as critical evidence and potential liability exposure. Insurance models may shift to reflect this new liability allocation.

Looking ahead, courts will likely adopt versions of this framework as agentic AI litigation emerges. The proposed standards create incentives for developers to build interpretable, auditable systems, potentially slowing deployment but reducing legal uncertainty. Regulatory bodies may incorporate these principles into oversight requirements, establishing baseline accountability standards before major incidents occur.

• →The framework distinguishes three AI-human interaction types, each triggering different liability doctrines from existing tort law.
• →Interaction logs become the primary evidence for determining where human-AI trajectories departed from authorized undertakings.
• →Developers must implement constraint verification, epistemic transparency, and forensic logging to meet emerging "Reasonable Agent" standards.
• →Liability allocation depends on the degree of human control and developer foresight rather than assuming developer responsibility by default.
• →The framework enables courts to apply existing tort principles to novel agentic AI scenarios without waiting for new legislation.