Punishing Artificial Intelligence: Legal Fiction or Science Fiction

Abstract

Whether causing flash crashes in financial markets, purchasing illegal drugs, or running over pedestrians, AI is increasingly engaging in activity that would be criminal for a natural person, or even an artificial person like a corporation. We argue that criminal law falls short in cases where an AI causes certain types of harm and there are no practically or legally identifiable upstream criminal actors. This Article explores potential solutions to this problem, focusing on holding AI directly criminally liable where it is acting autonomously and irreducibly. Conventional wisdom holds that punishing AI is incongruous with basic criminal law principles such as the capacity for culpability and the requirement of a guilty mind. Drawing on analogies to corporate and strict criminal liability, as well as familiar imputation principles, we show how a coherent theoretical case can be constructed for AI punishment. AI punishment could result in general deterrence and expressive benefits, and it need not run afoul of negative limitations such as punishing in excess of culpability. Ultimately, however, punishing AI is not justified, because it might entail significant costs and it would certainly require radical legal changes. Modest changes to existing criminal laws that target persons, together with potentially expanded civil liability, are a better solution to AI crime.

Definitions of intent suitable for algorithms

“…A-bots do present novel challenges to the law which cannot be answered by making to the past. The question as to whether criminal law is suitable for application to A-bots is called The Eligibility Challenge and debated at great length in Abbott and Sarch (2020).…”

“…Suppose an A-bot were to perform some sequence of actions that would be qualify for crime X if a human had performed them with the requisite mens rea. We will follow the terminology of Abbott and Sarch (2020) and term this state of affairs an AI-crime. At present the A-bot has not committed a crime because it is not a legal person regardless of its mental capabilities.…”

Definitions of intent suitable for algorithms

“…This is a question which we feel can only be answered legitimately by surveying the public in a rigorous scientific experiment. The question as to whether criminal law is suitable for application to A-bots is called The eligibility challenge and debated at great length in Abbott and Sarch (2020). One conclusion of our look at intent outside criminal law in Subsection 2.6, is that the concept of intent as developed in criminal law is relied upon elsewhere.…”

“…The possibility of an autonomous algorithm or AI possessing the Mens Rea for a crime, is tentatively suggested as a solution to the problem of 'Hard' AI crimes in (Abbott and Sarch 2020). Someone is criminally culpable if their behaviour shows insufficient regard for some legally protected norms or interests.…”

“…1: This paper proceeds under the assumption that intent is a definable concept that does not require a human brain to exist, that it arguably exists in other biological entities with demonstrable intelligence and can plausibly exist in an artificial intelligence. Images: Octopus -James Keuning, AI -Komkrit Noenpoempisut, The Noun project Sarch (2020) and henceforth refer to this state of affairs as an AI-crime. It is entirely possible that an autonomous algorithm can commit a AI-crime, with out having being instructed to by its programmers or owners, it instead learned to behave that way through some machine learning technique.…”