Autonomous AI Agents and Criminal Liability: Attribution, Digital Evidence, and Human Accountability

This paper argues that criminal law does not need to recognize criminal personhood for machines in order to address harms caused by autonomous artificial intelligence agents. The real problem lies elsewhere: how to allocate, prove, and limit human responsibility when systems capable of planning, using tools, maintaining context, and executing multistep actions operate with increasing autonomy. The analysis distinguishes agentic AI from ordinary generative systems, revisits the literature on the so-called responsibility gap, and argues that in most cases the core difficulty is not a vacuum of accountability but a problem of tracing and distributing responsibility among developers, integrators, deployers, operators, and decision-makers. In Brazilian criminal law, familiar doctrines remain useful when AI functions as an instrument of a human plan, while more autonomous cases shift attention toward ex ante duties of design, supervision, documentation, and risk containment. The paper also examines emerging governance duties in Brazilian and comparative regulation and shows why digital evidence - including model versioning, prompts, tool permissions, approval logs, and deployment records - is central to criminal attribution and adversarial scrutiny. The most promising path is not criminal personality for AI, but stronger human governance and more demanding evidentiary standards.