AI and Liability under Uncertainty

In ancient Rome, the determination of accountability and responsibility turned on a basic formula, “cuius commoda eius et incommoda,” (the one who derives an advantage from a situation must also bear the inconvenience). This rule is the core of the Polluter Pays Principle (PPP).  In Trinidad and Tobago, it is reflected in the amended Water Pollution Rules (WPR), 2019, and the Water Pollution (Fees) Regulation 2019. The PPP is a straightforward code of international, and domestic environmental law.

The PPP is designed to achieve the “internalization of environmental costs.” To do this, it guarantees that the costs of pollution control and remediation are borne by those who cause the pollution, and thus reflected in the costs of their goods and services, rather than borne by the public at large (OECD Council 1972 Recommendation of the Council on Guiding Principles concerning International Economic Aspects of Environmental Policies; and the Rio Declaration 1992 Principle 16). The PPP now serves as a signpost as we develop frameworks to regulate Artificial Intelligence (AI) use.

In the speed to establish ethical principles for socially beneficial AI, governments have been besieged by tenets, declarations, partnerships, codes, expert group papers, and joint select committee reports. These expositions have added much to the public discussion of AI. But they have also amplified ambiguity, confusion, repetition, and redundancy. To circumvent the misadventure of a “market for principles”, a core set of principles has been agreed upon.

To build Good AI Societies, beneficence, non-maleficence, algorithmic justice, and autonomy must be embedded in a default set of practices to regulate the ethical use of AI. These four core principles are used in bioethics. A fifth principle: explicability, understood as combining both the epistemological sense of intelligibility (as an answer to the question “How does it work?”) and the ethical sense of accountability (as an answer to the question: “Who is liable for the way it works?”) has been added to extend the bioethics framework. This additional principle is needed if AI is to be construed as a form of agency and not as a novel form of intelligence.

One flaw of recent legislation framed to regulate AI risk is a set of risk categories for AI systems that are pigeonholed using risk groupings. A taxonomy of risk groupings is disconnected from a clear method for assessing risks in concrete real-world situations that combines specific risk factors influencing real-world AI application scenarios.

To address this legislative lacuna, consideration has been given to exploring the heuristic utility of the methodological framework developed by the Intergovernmental Panel on Climate Change (IPCC). The climate change model offers a nuanced analysis of risk by exploring the interplay among

(b) individual drivers of determinants, and

While some tort law arguments have been derived from a reading of tort liability under uncertainty, the research and policy reports on climate change risks continue to be an abundant source of inspiring ideas, especially the framework developed by the IPCC working groups. Accordingly, the risk of an event is assessed by the interplay among

(1) determinants of risk such as hazard, exposure, vulnerability, and responses,

(2) individual drivers of determinants, and

(3) other types of risk, including extrinsic, and ancillary risks.

The climate change framework offers a more accurate prototype to determine the risk magnitudes of AIs under a specific scenario. Such measurements will be based on hazard chains, the compromise among impacted values, the aggregation of vulnerability profiles, and the contextualization of AI risk with ripples of risks transferred and amplified by the Butterfly Effect.

In Latin America and the Caribbean, we should be committed to developing AI assemblages and technologies that guarantee people’s trust, serve the public interest, strengthen shared social responsibilities, and protect the environment. However, pre-digital experience points to a delay of decades before society catches up by rebalancing rights and protections to restore trust.

This requires an assessment of the capacity of existing institutions, including national civil courts to prescribe remedies for mistakes made or harms inflicted on citizens by AI systems.  The appraisal process should evaluate the existence of sustainable, majority-agreed fundamentals for liability from the origin and design stages onward, to reduce negligence and conflict.

At the outset, an assessment of which tasks and decision-making functions should not be delegated to artificial authority is urgent. This creates alignment with public opinion and some orientation with societal values. This assessment calls for the inclusion of institutions like Law Review Commissions which promote reform and keeping the law under review, AI developers, civil society, and the Organization of Commonwealth Bar Associations.  The outcome may be a framework of key principles that apply to urgent and/or unanticipated problems.

A deeper step would be to develop appropriate legal procedures and practice guidelines and to improve the digital infrastructure of the justice system to allow the scrutiny of algorithmic decisions in court proceedings linked to the principle of “explainability.” This would entail a requirement for procedures to include disclosure of sensitive information in litigation. To deal with traceability, unfair outcomes, unintended consequences, misguided evidence, inconclusive evidence, and inscrutable evidence, countries will have to develop auditing mechanisms for AI systems.