AI and Criminal Liability

What is unfolding is a spectre of regulation by AI, rather than regulation of AI. The rise of “cyber-physical” infrastructure has made the old demarcation between online and offline artificial. Data-driven agency is converting phones, kitchen appliances, trains, bridges, hospitals and wearables into a neural lace of uninterrupted scrutiny and surreptitious adaptation among the “inforgs” who experience reality as “onlife.” Latin America and the Caribbean (LAC) must come to terms with delegated authority in the Age of Artificial Intelligence (AI).

When one data-driven agent is integrated with a set of discrete interacting artificial agents, a multi-agent system evolves. This meshwork of interacting artificial agents becomes smart in the sense that it contains a more fundamental measure of unpredictability as to how objectives can be achieved in a distributed problem-framing space.

It is likely that such systems may contribute to solution paths that programmers of the discrete systems could never foresee. These artificial agents may execute their own programmes and negotiate with each other to achieve goals not prescribed by developers in multiple jurisdictions.

Such interactions may generate emergent behaviours that may not have been planned or directed from a central node. The system as a whole can therefore develop “global agency,” meaning that it begins to behave as a unity of action within its ecology. This requires a regulatory paradigm on a scale and with a range and capacity that must be regional.

The data-mining and processing, pattern-recognition and outcome-generating capacities of some deep neural networks, and generative LLMs, can make predictions at speeds that far exceed human capabilities. The new knowledge those predictions may generate may present opportunities for highly targeted, dynamic interventions in new markets or conflict zones. They may also alter social relations at a global scale, and in real time. One consequence of this shift is the automated reorganisation of social life in a flow of consciousness linked to artificial life.

At this time, artificial artefacts are adroitly taking over tasks once performed by humans. In executing such tasks under delegated authority the systems may “intentionally” engage in behaviour which could be considered criminal if the task was completed by humans.

Despite its developer’s best intentions and actions-an autonomous agent may for example choose to manipulate markets while lacking the attributes that would make the artefact criminally liable under the rule of law. It is here that the regulation of human behaviour using judgements and jurisprudence, and the regulation of AI and Large Language Models (LLMs) collide.

In the above scenario a criminal responsibility lacuna emerges since no agent either human or artificial can be legitimately punished or sanctioned for the possible criminal outcomes. “AI crimes” is a fresh field. It may embrace the calculated performance of actions by AI assemblages which could constitute a crime if they were performed by a human having the required intention or “mens rea.”

Such offences may also be related to those AI crimes for which no human may be considered criminally responsible, within existing legal doctrines used for ascribing criminal liability. Such crimes could take place whenever the harmful consequence brought about by an artificial agent acting as a utility maximiser delivers an outcome that was not intended by any of the digital architects who contributed to the agent’s design.

To attribute “intent,” jurists defer to existing criminal doctrines, including oblique intent, or “dolus eventualis.” Consequently, no intentional criminal offence could be attributed to its developers anywhere in a multi-agent system that is distributed globally. As AI agents aggregate greater and greater autonomy, the less feasible it becomes for humans to anticipate an agent’s future criminal conduct.

It follows that it will be unlikely that the agent’s specific criminal behaviour may not be even loosely coupled to any reckless or negligent mental state of a citizen using a bot or portal. Simultaneously, the AI agents themselves cannot be held criminally responsible in any legitimate and meaningful way, because they lack legal personality, and in their current state of development, they may lack the capacities required for criminal responsibility.

It is therefore clear that technological management covers a broader gamut of regulatory targets than the legal regulatory paradigm. If the aim of technological management is to prevent certain problems, forms of conduct or actions from ever arising by making them impossible then the best form of regulatory response is to guarantee that they do not or cannot arise.

Since success will be limited, technological management requires that developers foreclose alternatives by architecture and code with barriers that are porous in only one direction, and software that will not function unless the user accepts the terms and conditions of use.

Cybersecurity is the new battlefield in LAC. Resilience in LAC turns on the ability to prepare for threat actors and hazards, to adapt to changing onlife conditions in the region, having at hand a recovery plan and when adverse conditions, disruptions, and denial of service incidents arise. Enhancing the security and resilience of critical infrastructure assets to existing threats and those that will arise in the future, requires coordinated action by governments across the region, including the private sector, and skilled individuals.