Deontology, Utilitarianism, and Truth-seeking AI

Immanuel Kant (1724-1804)

The classic question, “what matters more, actions or their consequences?” has persisted to be a highly contested subject within the realms of philosophy and ethics. The essence of the question can be portrayed through a fictional scenario: imagine that as you walk toward a bank, you see another person exiting, carrying a large duffel bag. You think little of it, but as an upstanding citizen, you hold the bank’s door for him as he exits. Later, after doing your business at the bank and arriving home, you see a story on the news. A robbery took place at the bank you were just at, and you see a picture of the man for whom you just held the door. The question arises: did you aid a criminal by doing something you thought was kind? 

It seems unfair to blame a person for acting in a way that they perceived to be moral under the given circumstances. This is the position of deontology in ethics. Deontology states that the intention that went into acting matters significantly more than whatever outcome was produced by the action. The philosopher Immanuel Kant formalizes this in his Categorical Imperative, a moral command that must be followed by all rational beings. In Groundwork of the Metaphysics of Morals, he writes, “act as though the maxim of your action were to become, through your will, a universal law of nature.” In simpler words, if you are deciding whether or not you should do something, consider if you could accept the result if everyone in the world started doing it. 

In contrast to this idea and methodology, utilitarianism prioritizes the contributions to collective happiness that are produced by a given action. Through this ethical perspective, a wrongful act could in fact be morally justified, so long as it produces a net increase to the happiness of everyone involved. John Stuart Mill, one of the most influential utilitarian thinkers, wrote, “actions are right in proportion as they tend to promote happiness, wrong as they tend to produce the reverse of happiness,” illustrating a moral philosophy based on the consequences of actions, not the intention that went into performing them. 

For the purpose of developing ethical AI, a clear distinction must be made between these two frameworks, and work needs to be done to determine the different situations and circumstances in which one would be preferable over the other. Before diving into the outcomes of implementing either one, we should consider if AI even merits the ability to adhere to moral principles. Traditionally, rationality has been the benchmark for a being’s ability to participate in ethics through moral principles. During the times of philosophers like Kant and Mill, humans were the only beings thought to possess rationality. According to the following definition of rationality,  “the ability to use reason and logic to make decisions and achieve goals,” AI appears to check many of the boxes required. If we analyze AI essentially as an synthesis of many number-crunching algorithms, the “decisions and goals” that they produce are outcomes of various mathematical processes, and those outcomes would be produced “rationally” if in accordance with a mathematically optimized recommendation. 

Regardless of whether AI is rational or not, it ultimately appears necessary to encode certain moral principles into their decision-making frameworks, purely for practical purposes. Since AI has proven to have a tremendous impact on society, it therefore must be given proper and justified moral guidance. A benefit of choosing deontology to guide AI systems is ethical consistency. Under the deontological lens, an action is unchangingly right or wrong, no matter the situation or circumstance. As a result, a deontological AI would never do anything it perceives to be wrong, no matter how small. Instead it would act within the ethical boundaries of what it perceives to be right. If this approach is taken to extreme lengths, AI can be very impractical. As a funny example, the AI model GOODY-2 avoids fulfilling any user request, attributing its incapability to a moral imperative never to be harmful or offensive. On the other hand, a utilitarian AI also has its drawbacks. It could commit wrongful actions in pursuit of what it deems “the greater good,” with little regard for the desires of humanity. For this reason, many consider utilitarianism to be a philosophy that is too cold and inhuman.

Once a theory is decided upon, much work still remains. To quote Mill,

"But when he begins to deduce from this precept any of the actual duties of morality, he fails, almost grotesquely, to show that there would be any contradiction, any logical (not to say physical) impossibility, in the adoption by all rational beings of the most outrageously immoral rules of conduct.”

Mill argues that the applications of a fundamental principle of morality (like Kant’s Categorical Imperative) are limited to hypothetical scenarios. The principle provides an overarching sense of what morality looks like, but does not tell us how to be moral, making it highly unusable for an ordinary person. This challenge of determining actionable duties holds true for AI. Solving the problem of which principle to use is only one piece of the puzzle. Extrapolating these real-life moral duties is the rest, and arguably the more challenging endeavor. So, while we may be able to envision what a moral AI looks like (thate it does not treat other rational beings only as means, for example), actualizing the stepping stones that take us from the current state of AI to this hypothetical ideal AI proves difficult. 

A certain fundamental principle could help get AI to this higher strata of development. That is the principle of truth-seeking, a principle that should be among the highest priorities of AI developers. AI models have already shown an aptitude for evaluating truth in logical statements. 

Channeling the computational processing power of an AI model internally to seeking its own true moral duties could initiate progress toward the ideal ethical AI model. During this iterative training process, the use of a fundamental principle such as Kant’s Categorical Imperative or Mill’s Happiness Principle could aid the AI as it seeks its own duties. Then, as the AI learns these actionable moral principles, it can incorporate them into its processes, moving itself closer to the hypothetical stage of ideally ethical AI.