What is Kantian Ethics, and could it be used as the basis for training the ethical decision-marking arm of future AI systems?
Abstract
As Artificial Intelligence (AI) rapidly evolves and integrates into our lives, it is inevitable that AI systems will eventually require some capacity to navigate complex ethical situations.
This paper critically examines the suitability of Kantian ethics – a cornerstone of deontological philosophy – to serve as the foundation for the ethical component of Artificial Intelligence systems, and eventually determine that it is inadequate.
The following paragraphs will demonstrate that the precision necessary for an automated ethical system is lacking in Kant’s framework, and that Kant’s idea of the ‘categorical imperative’ is diametrically opposed to the existence and purpose of Artificial Intelligence. It will conclude that Kantian ethics is not a suitable framework for automated ethics, but a modified version – or some other deontological framework – is perhaps the most suitable basis for the ethics of an AI machine.
Overview of Kantian Ethics
Kantian ethics is a rule-based framework attributed to philosopher Immanuel Kant. It is a profoundly complex framework that is explained over multiple long and complex works. The following explanation is my attempt to distil it to its core.
Kant’s framework centers on the ‘categorical imperative’ – an absolute commandment that demands unwavering obedience in all circumstances. (Manna & Nath 2021, p. 147). The first distinct formulation of the categorical imperative posits that an action is deemed moral if the guiding principle (maxim) behind that action can be logically willed to become universally adopted.
The second distinct formulation of the categorical imperative concludes that all individuals must act in such a way to treat all persons as an ‘end’, and not as a means to further one’s own ‘ends’. Kant determines that we have a duty to obey both formulations of the categorical imperative in all circumstances, as it has intrinsic value independent of any ‘ends’ – it is an end in itself.
This contrasts with what Kant declares a ‘hypothetical imperative’ – an action that is necessary to fulfill some end (Manna & Nath 2021, p. 147). For example, “I must study to write a good essay.” The further defining paradigms of Kantian ethics is Kant’s elevation of persons based on their capability to act rationally, bestowing a ‘transcendental freedom’ (Manna & Nath 2021, p. 141) from their base instincts and underscoring a consequential duty to obey the categorical imperative.
These constitute the central tenets of Kantian ethics. It is an extremely interesting normative theory because it provides a universal and consistent framework for ethical decision making that is rooted in a simple set of rules. It offers a unique and detailed alternative to consequentialist normative theory.
However, I do not think Kantian ethics to be a complete nor infallible ethical framework. Kant’s notion of maxims and his formulations of the categorical imperative – while broad – are highly imprecise. As a result, there are many situations where Kantian theory is unclear, or it is so counter-intuitive to many people’s idea of morality that it is dismissed entirely.
Flaws of Kantian Ethics
Let us explore one such scenario. Envision yourself as a German citizen amidst the Nazi regime, and you are harbouring Jewish individuals in your home. The Gestapo knocks at your door and asks if you know of any Jewish people. We must look to the categorical imperative to guide our action. Is it morally permissible to lie? A popular interpretation of Kantian theory asserts an absolute prohibition of lying (Mazur 2015), as one cannot will the maxim – ‘lying to achieve a personal desire’ - universally, lest society be destroyed. You must disclose the concealed individuals to the authorities, or maintain silence, indirectly relinquishing them. This is an example of an action that is counter-intuitive to our intrinsic moral compass that Kantian theory determines we must perform, if we were to find ourselves in that scenario.
However, one could raise the level of specificity of the maxim from – ‘lying to achieve a personal desire’ to a more nuanced specification such as ‘lying to protect an ostensibly innocent life from harm’, further evolving to ‘lying to save a life from religious persecution’ and so on, infinitely. It is infeasible to compute the exact maxim, yet each level of specificity has vastly differing implications in the context of the categorical imperative. To circumvent this convoluted scenario, one could even argue that the Gestapo are not rational beings and hence one can lie to them. This is an example of a situation where Kantian theory is entirely unclear – and we can extend this uncertainty to nearly every single non-trivial moral conflict. We see this in the diverging interpretations of the categorical imperative: notably, neo-Kantian Christine Korsgaard subverts the ‘murderer-at-the-door’ thought experiment and claims that “Kant was wrong in thinking that it is never all right to lie” (Korsgaard 1986, p. 330), through reasons which I have not explained above, but which outline the intense ambiguity in the original work.
Explanation of the ineffectiveness of Kantian Ethics in an automated ethical system
This is one rationale that drives my contention that Kantian ethics is not suitable in serving as the basis to design a system of automated ethics. Before I explain further, I will present two assumptions: first, that the automated ethics that will be used for an Artificial Intelligence is a more advanced and commercialised version of today’s AI. Secondly, programming the ethical framework into the AI will be treated as a black-box and as such we do not need to understand its technicalities.
Contention 1: Lack of precision that is necessary for an automated ethics system
As I detailed above, Kantian ethics lacks precision – to design an automated ethics system founded on Kantian ethics, a human would need to contribute a non-negligible amount to the ethical framework by increasing its precision, scope, and details, ultimately creating a distinct normative theory.
Contention 2: Intrinsic incompatibility of Kantianism and Artificial Intelligence
Furthermore, training AI on Kantian ethics seems to be incongruent with Kantianism itself. The basis of Kant’s framework is predicated upon the notion of a ‘rational agent’ that is bound by duty to the categorical imperative. However, a machine with Artificial Intelligence will always be bound by its original, ‘lesser’ purpose – there will always be a given ‘hypothetical imperative’ which supersedes the categorical, undermining Kant’s theory completely (Manna & Nath 2021, p. 148). On the other hand, if the machine were purely trained on the categorical imperative and told to disregard any hypothetical imperatives, this could lead to disastrous results.
For example, let us analyse a self-driving vehicle with a passenger requiring urgent medical attention. Strict adherence to the categorical imperative would disallow the vehicle breaking traffic laws and consequently saving the passenger, as this cannot be universalised. This subverts the intended purpose of the self-driving vehicle – to serve humans by proffering a more convenient mode of transportation with a similar level of efficacy as a human-operated machine. If the automated ethics system were to decide that breaking laws for the benefit of the injured passenger is the morally correct decision, then it is allowing its hypothetical imperative to supersede the categorical, and hence violating Kantianism. We see that the result of training an automated ethics system on Kant’s theory leads to either compromising the intended purpose of the AI, or inadvertently creating an ethical framework that is not Kantianism.
Tangential Foray into AGI and Kantianism
This leads to an interesting exploration of the ramifications of training an Artificial General Intelligence (AGI) machine – that is, an AI which is cognitively equivalent to a person – on Kant’s theory to be used as a tool for humanity – in a self-driving car, trading system, and so on. This, too, is a violation of Kantianism. As Kant determines that humans have intrinsic ‘dignity’ as a direct result of our capability to act rationally, an AGI which equals humans in this facet must also have intrinsic dignity and hence must be regarded as a human. It is interesting that in all other ethical theories and human moral intuition, there is not a clear solution to the moral problem of AGI. Yet, Kantianism provides a simple logical answer to the rational status of AGI.
Conclusion
Having said this, using a consequentialist theory as the basis for an automated ethics system would require near-infinite computation power to operate (Singh 2021, p.16) and hence can be deemed infeasible. Using some rule-based ethical framework – like Kant’s – seems much more achievable. Although Kantianism itself is not suitable because of the logical flaws reasoned above, one could extend and adapt Kant’s work to create a theory that can be congruent with Artificial Intelligence.
The general idea of using a rule-based ethical framework to create an automated ethics system seems much more achievable than a consequentialist theory (Singh 2021, p.16), which would require near-infinite computational power to operate. Although Kantianism itself would not be a suitable basis to design an automated ethics system because of the reasons above, this is an opportunity to extend upon Kant’s work to create a more adaptable and intuitive theory that can be congruent with Artificial Intelligence – but this is a topic for another paper.
References
1. “The categorical imperative is the unconditional command, which is an end in itself, and one should follow it to perform one’s own ‘duty’. In Kant’s words: ‘So act as if the maxim of your action were to become through your will a universal law of nature’ (2002: 38).” (Manna & Nath 2021, p. 147, Kantian Moral Agency and the Ethics of Artificial Intelligence)
2. “There is another imperative that emphasises the consequence part, i.e., hypothetical imperatives. Hypothetical imperatives hold the format of ‘if-then’, which describes doing a specific action if someone wants to achieve a particular goal.” (Manna & Nath 2021, p. 147, Kantian Moral Agency and the Ethics of Artificial Intelligence)
3. “. In this regard, Kant claimed in Critique of Pure Reason that only human beings could conquer the pathologically necessitated causal connection between actions and effects and ‘transcendental freedom’ is intrinsic to rational beings.” (Manna & Nath 2021, p. 141, Kantian Moral Agency and the Ethics of Artificial Intelligence)
4. “The philosopher Immanuel Kant said that lying was always morally wrong…Lies are morally wrong, then, for two reasons. First, lying corrupts the most important quality of my being human: my ability to make free, rational choices. Each lie I tell contradicts the part of me that gives me moral worth. Second, my lies rob others of their freedom to choose rationally.” (Mazur 2015, https://scu.edu/mcae/publications/iie/v6n1/lying.html)
5. “These reflections might lead us to believe, then, that Kant was wrong in thinking that is never all right to lie. It is permissible to lie to deceivers in order to counteract the intended results of their deceptions, for the maxim of lying to a deceiver4 is universalizable. The deceiver has, so to speak, placed himself in a morally unprotected position by his own deception. He has created a situation which universalisation cannot reach.” (Korsgaard 1986, p. 330, Philosophy and Public Affairs),
6. “Now, as far as the performance of an AI agent is concerned, it works according to hypothetical rules. These rules are encrypted in the command system, and it navigates the system to work in a particular technique to accomplish specific goals. It is speculated that if AI robots can successfully execute moral actions as an intelligent computing device, they can be a ‘hypothetical moral agent’ at their best (Powers 2009).” (Manna & Nath 2021, p. 148, Kantian Moral Agency and the Ethics of Artificial Intelligence)
7. “In this section, I extend this argument and argue that Kantian ethics is more natural to formalize than the two other major ethical traditions, consequentialism and virtue ethics, because it requires little data about the world and is easy to represent to a computer.” (Singh 2022, p. 16, Automated Kantian Ethics: A Faithful Implementation)
