Machine Ethics Before the Revolution · Part 3

An Investigation of a Kantian Moral Machine

· 16 min read
philosophyai ethics

Written for Kant’s Ethics (PHIL 4260) at Columbia University, spring 2020. Part three of Machine Ethics Before the Revolution. Published as written, with typographical and citation corrections.

One year ago I began exploring the intersection between machine morality and Kantian ethics. With computational ethics commanding its place on the cutting edge of practical philosophy, it seemed that programmers and the machines they design were in dire need of a well-defined ethical framework to avoid replicating or extending the moral dilemmas that artificial intelligence has already encountered. Questions surrounding autonomous driving, use of unmanned weapons, implicit racial bias in surveillance equipment and the inherent risk of a singularity (the event at which artificial and human intelligences become indistinguishable) demand answers. In this paper I hope to expand upon the ideas presented in my initial bout with machine morality titled An Application of Kantian Ethics to Artificial Intelligence, and further clarify how machines may go about formulating a consistent and, more importantly, practical code of ethics. I will revisit why “machinekind” may in fact be better suited than humankind to the deontological tradition that Immanuel Kant rules over. I then move on to the difficult topic of artificial freedom. Additionally, I utilize the concepts of nonmonotonic logic and “explanatory extension” to show how computers may be capable of building cohesive and consistent ethical guidelines that mimic humanity’s practical applications of Kant’s formal moral law - the categorical imperative - by appealing to Jürgen Habermas’s discourse ethics. I conclude by suggesting the creation of a diverse network of autonomous moral agents (AMAs) of varying levels of predefined moral aptitude to collectively produce a code of ethics more reflective of the overlap between Kantian formalism and the authentic human psychological understanding of morality as a culturally defined language steering us away from humanity’s natural evil (Kant’s “stain of our species”).

Some Background in Kantian Ethics

Some brief background in Kantian ethics may be necessary for readers unfamiliar or in need of a refresher with the concepts mentioned above. The root of Kantian ethics is the intrinsic value of a good will.1 It follows then that the moral worth of a particular action is located in the rationale behind the agent’s willing of that action and not in the end it may produce, since the goodness of an end is never intrinsic. Moral actions are those in which the driving maxim of the action is represented as a law determining the will.2 This lawlike formulation of our maxims is one of Kant’s most famous contributions to philosophy: the categorical imperative (CI). Maxims that can pass the test of the CI are deemed moral actions. Kant provides a variety of formulas for the CI, each one bringing its true nature more into focus. They are examined and re-examined in the literature - you can find one such analysis in my paper An Application of Kantian Ethics to Artificial Intelligence. I will refer to them throughout the rest of the paper. These formulas3 each serve a different purpose in Kant’s exploration of the human moral faculty, and their law-like structure is the basis for his deontological ethic. It is important to note that he believes them to be equivalent, and merely different forms of the same underlying moral principle. Computers are well equipped to engage with the procedures that Kant outlines for human’s looking to determine the morality of their actions: questions of universalizability and systematicity are their bread and butter, so to speak.

The Mechanical Appeal of Deontological Ethics

Having (very briefly) outlined the core tenets of Kantian ethics, I now turn to the unique aptitude that machines have for the procedural moral evaluation Kant imagined for humanity. Thomas Powers put it quite well:

A rule-based ethical theory is a good candidate for the practical reasoning of machine ethics because it generates duties or rules for action, and rules are (for the most part) computationally tractable.4

In the most formal sense of the CI a maxim is determined to be moral simply by procedural analysis of its universalizability and general systematic compatibility within the preexisting moral framework. Moral maxims are those which, when universalized across all moral agents, do not produce a contradiction or absurdity - a task of simple logical contradiction between the proposed maxim and the default language G describing all acceptable moral propositions in the form of an imperative: I ought to do a particular thing. In no way is this rational methodology restricted to mankind. In spite of its imperative structure, the form of the categorical imperative is incredibly restrictive. A moral maxim dictates an action not from the ends it may bring, but out of declarative respect; I ought to do this action only for itself, for an action done only for itself is driven by a good, and therefore moral, will. This is clearly an incredibly pessimistic outlook, for mankind is fickle as philosophers throughout history have identified. Aristotle speaks of akrasia: a state of mind in which the soul’s appetitive partition actively fights against one’s better judgment through weakness of the will. Kant observes this very phenomenon in the Groundwork:

For, the will stands between its a priori principle, which is formal, and its a posteriori incentive, which is material, as at a crossroads; and since it must still be determined by something, it must be determined by the formal principle of volition as such when an action is done from duty, where every material principle has been withdrawn from it.5

The will must engage with both the formal principle of duty from the moral law and the material incentives and desires that so frequently drive action. Not only is the deliberation of action wrought with conflict internally; as Patricia Kitcher rightly points out “where moral motives and prudential ones would produce similar looking actions, it is virtually impossible for observers or even agents themselves to isolate the actual motive or determining ground.”6 Clearly there is a great deal of practical difficulty utilizing such a general form as the categorical imperative to guide moral behavior, since we so rarely can even identify the motivations of our actions.

Let us bring this procedural conflict more clearly into the light. First The CI insists on the utmost respect and reverence for our actions if we are to deem them moral. There can be no external “material principle” remaining to motivate the action. Such maxims are quite difficult to identify, namely because humans are constantly flitting between our moral imperatives and our material incentives. Kant further restricts the CI by insisting that the moral law determines the will. While humans are clearly not ideally suited as Kantian moral agents, an artificial moral agent (AMA), that is, a machine designed to behave morally in the most general sense, is well-equipped. The AMA does not have desires or passions in the Kantian sense, and is therefore not subject to the same akratic behavior diagnosed by philosophers from Aristotle to Kant. Additionally, the AMA is a superior classifier of moral contradiction and inconsistency. Where a human might miss (or ignore) a logical fallacy on account of self-interest or pure ignorance a machine is certain to formulaically establish closure and determinism of an ethical code. Finally, the AMA is blessed with John Rawls’ “Veil of Ignorance” since no current artificial intelligence is self-aware.

It cannot be that easy. In fact, to declare victory for the AMAs now would be to grant Kant an unearned victory. These broad strokes of morality aren’t particularly useful to us or an AMA. What is needed is a framework that allows for the more realistic and practical situational flexibility that the real world dictates to us. Onora O’Neill uses “spontaneity” to capture the more relaxed nature of practical reason and practical morality:

A first thought might be that it is often not necessary to reach highly, let alone wholly determinate judgements… indeterminate (generic, approximate, vague, schematic) answers are often enough… complete determinacy may be neither necessary, nor possible, nor useful. Only where we calculate rather than judge should we expect to reach wholly determinate results… Deliberation aims to guide action, and act-tokens (like other particulars) will be determinate in all respects. Indeterminacy, it seems, cannot be shrugged off in practical matters.7

Without a doubt our moral evaluation procedures are imperfect, regularly updated and revised, and open to subtle fluctuations tied to mood, feeling, and environment - and nonetheless they are far more practical and useful to analyze than the underlying principle that Kant claims must dictate them all. I believe it is due to this ability (one may even say tendency) for us to err in moral judgment, that the underlying moral principle is semantically structured as an imperative. Kitcher writes: “But ‘I ought not to testify falsely’ is not an action representation, but a prescription.”8 This structure allows mankind to escape the binding nature of the CI while still maintaining the faculty of moral learning and experiment. An “ought” allows me to follow through on the maxim or ignore it. Mankind’s moral faculty is twofold - it combines our rational nature with the freedom of our will. Their cohesion defines moral behaviour (we are autonomous only insofar as reason dictates the will via respect for the moral law) and their conflict implies internal dissent and lack of control (heteronomy) as we are driven away from our faculty of practical judgment by emotions and fleeting material incentives.

Artificial Freedom

Suddenly the tables have turned. AMAs were perfect for the generalized sweeping command of the pure moral law, but their apparent inability to make free decisions of action disqualifies them from the more practical framework identified by O’Neill. Ryan Tonkens articulates this difficulty well in his paper titled A Challenge for Machine Ethics. Pointing to the Metaphysics of Morals he argues that Kant believed the concept of duty is the suppression of our vice-ridden inclinations by pure practical reason, and since AMAs cannot generate such inclinations to suppress they are incapable of acting from duty. Although an AMA certainly respects and reveres the moral law its inability to experience and overcome internal conflict and (quite literally) preprogrammed nature makes it unworthy of praise or blame - a necessary consequence of moral and immoral action.

I intend to show that Kant’s conception of freedom is actually quite approachable to an AMA. Much like Thomas Powers and Patricia Kitcher I will appeal to a nonmonotonic logic and explanatory extension as the basis of machine ethics, for such a logic’s flexibility to qualifying scenarios is far more realistic (and practical) than the rigid CI. Just as Kant continues to hold that the moral law does serve as the default, a priori principle of human moral action even when the maxim’s moral judgement is far more spontaneous (and certainly less comprehensive) than the CI may dictate, we can imagine AMAs holding certain preprogrammed truths to be their own sort of a priori set of principles. David Poole wrote an extensive paper titled A Logical Framework for Default Reasoning that outlines the method by which axioms are submitted and redacted from some complex moral language. He aptly points out that our own common sense - that same common sense that Kant consistently appeals to in his egalitarian ways - is not reflected in monotonic logic. New axioms (moral laws) are added to the code generating additional logical consequences for each future submission. Even more importantly we seem to revise old opinions with new knowledge. This begs the question, how might we go about replicating this continuous learning and reframing of our laws of action in a computational structure? The initial parallels are clear: defining the CI and the intrinsic respect and determinism tied to the CI is easy to replicate in a computer’s lack of choice, and the procedure of contradiction identification is certainly an approachable problem for our AMAs. It is incorrect to view the nonmonotonic logic as a hindrance or obstacle to the determinism and irrevocability of the CI - it merely assigns this principle its appropriate place as the default logical reference against which all other maxims may be judged. An AMA is free to formulate its own entirely consistent moral code - it need not be anthropocentric in the slightest (beyond the assignment of the CI as a determining law of action) as long as it conforms with its default logic. This is scary - imagine robots being free to establish moral codes that are not human-centric. There is no denying that the goal of AI is not to reach this moral independence as fast as possible, but there are numerous reasons why such a project may prove useful to ethicists. (1) Machine ethics may prove a better framework for deontological moralities, and could even inform humanity of some of its own inconsistencies and ethical blunders (god knows we have many). (2) The design of anthropocentric moral machines, i.e. self-driving cars, autonomous weaponry, artificial intelligence assistants, will require an independent framework in order to learn humanistic morality in a deeper sense - one that is not entirely preprogrammed by the creator.

Development of Machine Morality

In An Application of Kantian Ethics to Artificial Intelligence I appealed to Jürgen Habermas and his conceptualization of discourse ethics. Taking a different approach to achieve the justice desired by John Rawls and his “veil of ignorance,” Habermas points out that complete circumstantial ignorance is not a realistic requirement for moral examination. Kant makes it very clear that human beings are not capable of discarding their self-serving incentives on a whim, and it is in fact the rational struggle of overcoming them that allows us to label a particular maxim as moral to begin with. While a single AMA’s respect for its programmatic “a priori” deliberation formula may infringe upon this very ability to “overcome” our vicious material incentives, a collection of AMAs would be capable of completing large scale ethical deliberations at lightning speed.

I proposed such a network of computers in my last paper, but I now hope to flesh out its conception with a major change: the network will be a collection of moral infants9 - AMAs with singleton axiom sets by which to operate. Let AMA X maintain the moral principle m1 as its foundational principle - according to X one can derive m1 from the CI, and thus it is a moral law that determines action. Let AMA Y maintain the moral principle m2 as its foundational principle. Now let X and Y engage in a type of ethical discourse in which each AMA runs the other’s foundational principle through the CI test to see if any volitional contradictions or inconsistencies arise. If none appear then each AMA adopts the other’s moral into their ethical code. X now holds both m1 and m2 to be moral maxims that determine action. Now let AMA Z maintain the moral principle m3. Upon ethical discourse with X and Y the AMA Z determines that m1 and m2 are incompatible with m3 on their own, but the set {m1, m2} provides an exception - the premises of m1 and m2 together allow for cohesion in the ethical set {m3, {m1, m2}}. This syntax implies that m3 remains in the default logic of Z while m1 and m2 rely on explanatory extension (a niche exception) to avoid contradiction and subsequent failure of the CI.

Now, to make this network productive in a practical sense, allow N moral agents with N foundational maxims to be exposed to a wide selection of test maxims, and let this ethical discourse play out over many iterations. All AMAs are related by their default logic - the CI - although the practical form that their default logics take are different. Furthermore, for the sake of realism, in each iteration some percentage of AMAs will act according to ~{M} where M is their set of moral codes. In this way we will introduce unpredictability into the network, thus encouraging learning as every AMA reacts to the sudden change of state introduced by this rogue behavior. This ethical discourse mirrors Kant’s suggestion in the Religion book that a moral agent combat their evil, vicious and deceptive maxims by forming ethical communities (churches) that collectively commit themselves to the maintenance of their members moral behavior.10 Our network is merely a computational church.

Has Kant’s egalitarian principle been broken down into a mere approval of diversity of thought? If single moral agents are incapable of reliably suppressing their evil maxims and staying straight on the narrow path of morality is a collection of varying principles and a logical, educational discourse the only way to achieve pure, rational morality as envisioned by Kant? Is God merely our programmer, and our epistemic pursuits a simple discourse intended to expand the principle of the CI into a more practical code? While such an ethical “wild-west” of a network as described above may veer too close to the terrifying edge of an AMA singularity it may very well be our best bet at unlocking the consistent, coherent, yet practical ethic that we (machine and humankind alike) so desperately need.

Works Cited

  • Kant, Immanuel. Practical Philosophy. Translated and edited by Mary J. Gregor. Cambridge: Cambridge University Press, 1996. Kant is cited throughout by Akademie pagination: Groundwork (volume 4), The Metaphysics of Morals (volume 6).
  • Kant, Immanuel. Religion within the Boundaries of Mere Reason. Part Three, on the ethical community; radical evil at 6:18–53.
  • Kitcher, Patricia. “Kant’s Ordinary Moral Agent.”
  • O’Neill, Onora. “Kant on Indeterminacy, Judgment and Interpretation.” 2016.
  • Poole, David. “A Logical Framework for Default Reasoning.” Artificial Intelligence 36, no. 1 (1988): 27–47.
  • Powers, Thomas M. “Prospects for a Kantian Machine.” IEEE Intelligent Systems 21, no. 4 (2006): 46–51.
  • Tonkens, Ryan. “A Challenge for Machine Ethics.” Minds and Machines 19, no. 3 (2009): 421–438.

Footnotes

  1. “It is impossible to think of anything at all in the world, or indeed even beyond it, that could be taken to be good without limitation, except a good will.” (Groundwork, 4:393) ↩

  2. “Nothing other than the representation of the law in itself - which of course can take place only in a rational being - in so far as it… is the determining ground of the will, can therefore constitute the pre-eminent good that we call moral,” (Groundwork, 4:401) ↩

  3. The Universal Law Formula, the Humanity Formula, etc. ↩

  4. Thomas Powers, “Prospects for a Kantian Machine,” IEEE Intelligent Systems (p. 46) ↩

  5. Groundwork (4:400) ↩

  6. Patricia Kitcher, “Kant’s Ordinary Moral Agent” (p. 13) ↩

  7. Onora O’Neill, “Kant on Indeterminacy, Judgement and Interpretation” (p. 2) ↩

  8. Patricia Kitcher, “Kant’s Ordinary Moral Agent” (p. 16) ↩

  9. I am stealing this terminology from Thomas Powers because I think it is the correct vocabulary for this case. ↩

  10. Religion within the Boundaries of Mere Reason, Part Three, on the ethical community. ↩