Phronēsis is a philosophical term that refers to conscious ethical reasoning or practical wisdom. It is argued that in most adult humans this capacity plays a primary role in high level ethical reasoning. If we want AI systems to have the capacity to reason on ethical problems in a way that is functionally equivalent to competent humans, then we will need to create machines that display phronēsis or practical wisdom in their interactions with human agents. It is argued here that this is the highest goal of AI ethics, but that this will not be a trivial problem since it may first require that the problem of artificial consciousness is solved. Achieving artificial phronēsis is necessary, since moral and ethical competence is required in order to develop ethical trust in the users of AI or robotics systems.
johnsullins.com
Ethical impact agents (EIA) need no explicit claims to consciousness to do what they do. These systems are notable only in that in their operations exhibit some autonomy and have the ability to impact human agents in ways that have ethical concern. For instance an autonomous car that, during the course of its autonomous driving operation, impacts and kills a pedestrian in an accident, (such as the accident that happened on March 18, 2018 in Tempe Arizona involving a pedestrian and a self-driving Uber car) does so completely unconsciously. The System’s actions are produced by its sensors and actuators working in accordance to a program, but it has no conscious experience of the situation at hand, suffers no moral remorse after the accident, and has no emotional reaction while the event is unfolding. While we might ask plenty of questions about the safety of the autonomous car, no one blames the car itself in a conscious moral sense. That moral condemnation and all questions of legal responsibility are reserved for the emergency backup driver present in the car, and the company that built the car and is testing it on public roads.1 Thus, there is no particular need to refer to consciousness when dealing with questions on how to ethically design and deploy EIAs.
Artificial Ethical Agents (AEA) are the next step up and differ from EIAs only in that they have explicit ethical considerations programed into their operation. Building on our earlier example, an autonomous car that was programmed to take into account some ethical calculus of value when deciding whether to risk occupants of another vehicle or its own occupants to increased risk during an otherwise unavoidable crash would be an AEA. At first look one might think that an AEA muddies the waters a bit and that the machine itself might deserve moral blame or legal responsibility, but that is just a trick of the light. The moral blame and responsibility for any adverse consequences is still fully borne by the human agents that built, deployed, licensed, and operated the vehicle. The key is that the AEA has never itself chose its own ethical standards, instead they were chosen and programed in by human agents, who therefore assume any blame or responsibility for any ethical decisions made by the system that they designed and/or deployed. The machine is not conscious of any of the events that occur based on its operations, even the ones that look to an outside party as if they were a conscious ethical choice.
It is only when we get to the level of the Artificial Moral Agent (AMA) that consciousness may play an important role. An AMA would have the ability to choose ethical behaviors that are appropriate to the situation at hand in a way that exhibits a form of practical reasoning similar to what can be seen in competent ethical reasoning found in most human agents. This means that the system either is a conscious moral agent or is functionally equivalent to one. Another way to say this is to claim that the system displays Artificial Phronēsis (AP). Of course this concept needs a lot more explanation and that is what the rest of this paper will discuss. However, at this point in the discussion we can make the conjecture that, while consciousness is not required for EIAs nor for many types of artificial ethical reasoning agents AEAs, it may play an important role in the development of much more sophisticated AMAs that would be 1 This list is not meant to be exhaustive, political decision makers who allowed the system to use public roads should also be under scrutiny as well as many other humans who took part in the chain of events that led up to this accident. much more useful in interacting with humans in complex social situations that tend to be bounded by shifting ethical norms.
Phronēsis is a term that many are not familiar with outside of philosophy and the word can seem a little off-putting. However, if one’s goal is to create AI and robotic agents that have the capacity to reason intelligently about ethical situations, then understanding this technical term will reward those who try, given that it is so relevant to the understanding of ethics.
Phronēsis has an ancient pedigree and has come down to us largely through the
tradition of virtue ethics as the skill of being able to “live well” [
Briefly stated, phronēsis refers to the practical wisdom that a conscious moral agent
uses when she is confronted with a difficult moral or ethical problem and attempts to
overcome these difficulties in an intelligent manner. Given that ethical problems are
always novel, no set of preconfigured answers will suffice to solve the problem, which
means that learning and creativity are the hallmarks of a phronētic agent. “There is no
general rule/procedure/algorithm for discerning which values, principles, norms,
approaches apply; rather, these must be discerned and judged to be relevant in the first
place, before we can proceed to any inferences/conclusions about what to do” [
AEAs might be successfully designed taking one or more ethical schools of thought
into account. One might design an autonomous system that makes ethical decisions
based largely on applied utilitarian or Kantian based calculations or rules, models of
human moral psychology, human religious traditions, or even on the three laws of
robotics developed by Isaak Asimov.2 While one could make AEAs using any of these
methods that might be useful in certain circumstances, they will all fall far short of an
AMA with artificial phronēsis. “For the virtue of practical wisdom or phronēsis
encompasses considerations of universal rationality as well as considerations of an
irreducibly contextual, embodied, relational, and emotional nature—considerations that
Kant and others have erroneously regarded as irrelevant to morality” [
It is understandable that systems designers will either want to ignore ethical
reasoning entirely attempting to avoid the construction of even EIAs. The slightly more
adventurous will attempt to apply the more computationally tractable rule based ethical
systems that could result in useful AEAs. Why get involved with AMAs that require
something like artificial phronēsis to work correctly? To succeed at that may require
solving problems in artificial consciousness, artificial emotion, machine embodiment,
etc., all of which may be computationally intractable. Let’s look at what might be the
reward for pursing the more difficult problem of building AMAs with artificial
phronēsis.
2 Introductory descriptions of all of these options and more can be found in [
Artificial Phronēsis (AP) is that claim that phronēsis, or practical wisdom, plays a primary role in high level moral reasoning and further asks the question of whether or not a functional equivalent to phronēsis is something that can be programed into machines.
If we want AI systems to have the capacity to reason on ethical problems in a way that is functionally equivalent to competent humans, then we will need to create machines that display phronēsis or practical wisdom in their interactions with human agents. This means that AP is one of the highest goals that AI ethics might achieve. Furthermore, this will not be a trivial problem since not all human agents are skilled at reasoning phronētically, so we are asking a lot from our machines if we try to program this skill into them. On top of this, the most difficult problem is that achieving AP may first require that the problem of artificial consciousness is solved, given that phronēsis seems to require conscious deliberation and action to be done correctly. Even so, the achievement of AP is necessary, since moral and ethical competence is required in order to develop ethical trust between the users of AI or robotics systems and the systems themselves. Achieving this will make for a future where humans can be comfortable inhabiting and not feel oppressed by the decisions made by autonomous systems that may impact their lives. 1.2
AP is a new concept but it is gaining some attention. What follows are some statements to help define this new, interdisciplinary area of research.
AP Claims that phronēsis, or practical wisdom, plays a primary role in high level moral reasoning and further asks the question of whether or not a functional equivalent to phronēsis is something that can be programed into machines.
AP is a necessary capacity for creating AMAs, however the theory is agnostic on the eventuality of machines ever achieving this ability but it does claim that achieving AP is necessary for machines to be human equivalent moral agents.
AP is influenced by works in the classical ethics tradition but it is not limited to only these sources. AP is not an attempt to fully describe phronēsis as described in classical ethics. AP is not attempting to derive a full account of phronēsis in humans either at the theoretical or neurological level. However any advances in this area would be welcome help.
AP is not a claim that machines can become perfect moral agents. Moral perfection is not possible for any moral agent in the first place. Instead AP is an attempt to describe an intentionally designed computational system that interacts ethically with other human and artificial agents even in novel situations that require creative solutions.
AP is to be achieved across multiple modalities and most likely in an evolutionary machine learning fashion. AP acknowledges that machines may only be able to simulate ethical judgement for quite some time and that the danger of creating a seemingly ethical simulacrum is ever present.
This means that AP sets a very high bar to judge machine ethical reasoning and behavior against. It is an ultimate goal, but real systems will fall far short of this objective for the foreseeable future. 1.3
If phronēsis was simply a concept only from ancient philosophy, it would be of limited
value to the project of AI. But it has evolved over time and one very interesting
development of the concept came from the philosopher John Dewey [
If he is correct, then phronēsis is part of what makes many of us competent, conscious, and conscientious beings. It follows then that AP is either essential for the creation of conscious machines, or vice versa. 2
Ethical Trust of AI and Robotic Agents Trust and AI and Robotic agents is its own complex topic, but here let’s limit our discussion to phronētic trust. Our AP agents will need to convince us that they have our moral character in mind when they are dealing with us and will make decisions accordingly. We learn from each other how best to develop our own moral character, so these machines will need to participate in that important social process as well. In some sense they have to be able to serve as a phronēmon, or moral teacher. In these kinds of relationships with our machines, we have to be warranted in reasoning that the machine we are trusting has a good ‘character’ that is deserving of our trust. Without a sufficiently developed AP, then this will be impossible and there will be no good reason to try to build artificial ethical agents and we will need to limit the applications in which we employ AI to only those with no ethical impact. 2.1
Mark Coeckelbergh [
Only if we wanted machines to join us in the phenomenological-social trust would we really need AP. At that level we would join them in a new kind of society. That is a very interesting eventuality to contemplate but it is also one that would come with rights and responsibilities that would be bidirectional, so we should proceed with cation, and if it turns out that machine consciousness is technically infeasible, then it is something that we could not approach at all.