Structure in AI, Law, Ethics, the World and the Mind
There are two kinds of (AI, Law, Ethics, Politics, World Models—take your pick), and that dichotomy profoundly affects our view of reality and the world we live in.
Just after writing the report on my recent work in machine ethics—on AI, the social role of automated systems, and ethics—I found myself in a couple of discussions of politics and the law, explaining some of the philosophy and history of the US’s legal system, and was struck by a common feature, a recurring dichotomy, one that seems quite important.
When writing about AI, it was important to distinguish between logicist and connectionist approaches. This was of particular importance in my recent work because the different approaches imposed very different burdens on the kinds of normative ethics that could be implemented. In discussions of the recent actions of both judges and Attorneys General, the distinction between Common and Statutory Law was key to understanding both the different perspectives of liberal and conservative critics and understanding the relationship between the law and ethics.
So what are these distinctions?
Logicist vs Connectionist AI: Shortly after WWII, the early cyberneticists, inspired by emerging neuroscience started to try to build automated systems modeled on the working of natural brains and nervous systems. A decade later, a new field, Artificial Intelligence (AI), emerged. AI researchers concentrated on creating data structures to represent knowledge and algorithms: systems of rules and computer code intended to duplicate human logic and reasoning. This style of logicist AI dominated for a couple of decades or more. Periodically the cyberneticists’ biology-inspired approach has returned under the labels of connectionism, neural nets, deep learning, artificial intuition and the like. Currently, the big successes in AI have all been on the side of Deep Learning and it is overshadowing the logicists’ “Good Old-Fashioned AI”.
The two approaches are quite distinct as can be seen in, for instance, Selmer Bringsjord’s “The Logicist Manifesto: At long last let logic-based artificial intelligence become a field unto itself” on the one hand, and Monica Anderson’s use of the term “Model-Free Methods” and her claims that “Machine Learning is the only kind of AI there is” on the other.
In the logicist world, intelligence is expressed in explicit formulas and algorithms written in mathematical programming languages according to the rules of formal logic. These operate on models of the world and our knowledge of it that are expressed in taxonomies and ontologies that explicitly capture the features and relationships of the things and actions that make up the world.
In the model-free connectionist world, the foundation of intelligence is very simple structures that use pattern matching techniques to discover connections between sensory signals, make predictions and alter their behavior so as to correct for unsuccessful predictions. Learning machines are not programmed, but trained, and exhibit artificial intuition or arrive at artificial opinions. Words, objects, actions and the like are not represented by explicit stored values, but exist as patterns or webs of connections.
Common Law vs Code Law: The US legal system is heir to, and based upon, two rather different legal traditions: British Common Law, and Imperial, particularly French Napoleonic Code Law, which is also called Civil or Statutory Law. Common Law is a particularly strong tradition in the States that were once British colonies and the Federal judiciary. Code Law is dominant in Louisiana, and strong in former French and Spanish colonies. Both traditions affect jurisprudence throughout the US.
In Common Law, judges and juries interpret the written law and pass judgement based on the facts, and especially the precedents set in previous cases. It is a system of accumulated wisdom, judgment and interpretation. In a common law system judges can even set aside the written law if it conflicts with the foundational Constitutional Law, and the precedents that rest upon it.
This primacy of accumulated precedent is particularly evident when you realize that while both Great Britain and the US recognize the primacy of Constitutional Law, only the US has a written Constitution. The English “Constitution” is, itself, a heritage of decisions and actions taken by judges, councils, parliaments and Kings and not a singular core constitutional document. Moreover, US Constitutional Law inherits from the British Constitution. Note, for instance, that while the US Constitution—the document—doesn’t define “habeas corpus”, it does say that the “Privilege of the Writ of Habeas Corpus shall not be suspended, unless when in Cases of Rebellion or Invasion the public Safety may require it.” It assumes this feature of British Constitutional and Common Law.
Code, Statutory or “Civil” Law, on the other hand, is based upon an organized body of laws passed by a legislature or dictated by a monarch. The job of magistrates in a civil system (or the lower civil courts in a common law system) is to merely administer the law. The roles of judges and magistrates in civil courts involve far less discretion, judgement, and interpretation than in common law courts. Authority rests in the hands of the legislature, or given the Imperial origins of Code law, frequently the King or Emperor, whose authority comes from God or from impersonal Natural Law.
Normative Ethics: The field of normative ethics is divided into three major branches which reflect a similar dichotomy to the ones above, though not in a precise one-to-one fashion. These are deontology, or rules-based systems, consequentialist systems, and virtue-based systems.
Historically, deontological systems start with lengthy sets of ethical rules often attributed to God or the gods. Judaic Law as found in Leviticus, the Ten Commandments and so on are examples. Over time, these tended to be distilled into a smaller number of principles such as Jesus’s Two Great Commandments or Kant’s deontological principle: the Categorical Imperative. The early rule sets work much like Code Law or the logicists’ computer code. In fact, Bringsjord’s team created a specific language, DCEC (the Deontic Cognitive Event Calculus), to encode ethical algorithms in and contributed a chapter entitled “The Divine-Command Approach to Robot Ethics” to the book Robot Ethics.
At the other end of the deontological spectrum, we have Jesus’s “Love your neighbor as yourself,” and even more so, Kant’s dictum to “act only according to that maxim whereby you can, at the same time, will that it should become a universal law.” While stated as universal rules, these principles are so general that they require extremely high levels of judgement and interpretation, not unlike the burden on a common law judge.
Dissatisfaction with deontology led philosophers during the Enlightenment to turn to consequentialism, wherein it is not a specific set of rules that determines morality, but the outcomes of one’s actions. “The ends justify the means” is the implicit principle. In Jeremy Bentham’s act utilitarianism, one of the best known consequentialist systems, we should perform the action that will create the greatest net utility—that which creates the best results for the most people. The result is that consequentialism is not so much about rules, prescriptions and ethical “algorithms” as it is about using judgement in applying general principles, potentially discarding a rule judged not to be applicable. Its ethics are situational and relative.
One of the earliest types of normative ethical systems is Virtue-based. It concerns itself not with following specific rules or being concerned with the consequences of one’s acts. Rather, it focuses upon the nature and character of the actor. A person who embodies certain values will act in accordance with those values. Which model of the world this is most like depends upon one’s understanding of the notion of “virtue”.
Aristotle, often seen as one of the strongest advocates of virtue-based ethics, defines virtue in terms of behaving in a manner that is in accordance to a mean or balance between two vices which are extremes of deficiency and excess. Courage is a balance between fear and cowardice on the one hand and foolhardy recklessness on the other. For him, the virtues require considerable judgement and weighing of principles.
On the other hand, many religions define virtue solely in terms of obedience to God’s will, and see some holy book as expressing that will precisely and inarguably. In this case, one or more virtues can be defined in very absolutist terms. For example, in some religions, the virtues of chastity and modesty, especially in women, can become as authoritative and compelling as deontological rules, and involve very little in the way of context, connection or judgement.
Thus, throughout normative ethics, we find absolutist systems, that work according to precise rules in ways similar to logicist AI and code law, and relational systems that share similarities with connectionist AI, and common law. It is not quite as simple as Deontology is the one, Virtue the other, but the dichotomy is present.
World views: In all three of these realms we can see two distinct approaches. One is based on clear and precise definitions and rules. The other is based upon context, connections, complexity, and balancing of influences. As with Aristotle’s virtues, one can choose to see the world according to either extreme or to some sort of balance. One’s world view can be heavily weighted to logicist, code law and absolutist ethics, or to connectionist, common law, and situational/relative ethics, or finally, one can see the world as a balance between models and connections, common and code law, rules and human judgement
Recognizing these as world views we can begin to see them in realms other than AI, the Law and Ethics. Take, for instance our understanding of the physical world. The classical Newtonian physics and Cartesian philosophy of the Enlightenment saw the world in terms of hard facts and rules. Our understanding was modeled upon clockworks and billiard balls, things that obeyed the simple mechanical and mathematical laws of nature. It’s a very logical and logicist view.
For the last century or two, science has been becoming much more “connectionist”. Relativity, Uncertainty, Incompleteness, and the wave/particle duality, are all examples of how much of contemporary science is becoming more “connectionist” and context dependent.
Coming from the perspective of one or the other of these world views explains a lot of the misunderstanding and outrage that we see in our politics as well. Here are the examples that I was discussing when I realized the power and the broad compass of these two world views. My point in presenting them is not to argue for one political or moral position in either case, but to show that the disagreement on these issues is strongly influenced by the dichotomy of world views.
Politics: Recently, a number of my Conservative friends have been aghast at the actions of the Massachusetts Attorney General who they see as arrogating the power of the legislature in “rewriting” the state’s assault rifle ban. Her actions make much more sense if you view them in the context of Common, not Statutory, Law. From this perspective, she has adopted a new interpretation of the existing statute that she thinks is consistent with existing case law and precedent, but which has not yet been tested before the judges who will have to rule that it is valid or not. Historically, Massachusetts’ jurisprudence has rested very heavily on Common Law, with very little emphasis on the Statutory tradition.
Conversely, a great number of my Progressive and Liberal friends have been aghast that recently, an Oklahoma appeals court ruled that neither that state’s Rape law nor its Forcible Sodomy law covers oral sex with an inebriated person. Oklahoma’s jurisprudence has a very strong Statutory Law tradition coloring their Common Law. Their law prohibits judges from criminalizing actions that are not explicitly criminalized in the state Code. The statues are pretty clear in this case. The Rape law does cover inebriation, but not oral sex. The Forcible Sodomy law covers oral sex, but not drunkenness. Additionally, the law treats sodomy as a crime, a crime against nature, and is specifically concerned with suppressing immorality rather than protecting the autonomy of the individual. If drunkenness is immoral, then it doesn’t necessarily obviate the responsibility for the immoral act of sodomy and so both parties share the moral guilt.
In all of these realms, we see these same two dichotomies, two approaches to seeing and understanding the world. They contribute to the phenomenon that I’ve written about before: our division into separate “worlds”, giving us very different notions of what is real and what is reasonable. When we hear someone who is operating from the other approach, they can sound absurd. “What world do they live in?” we can ask ourselves. These dichotomies are not the only source for that type of misunderstanding, but they contribute strongly.
On the one hand, we have logicist AI, Code law, absolutist ethics, classical mechanics, a world of hard facts, clear rules and definitions, hierarchy and delegated authority. The world and its workings can all be stated in objective language. Parts, people, things are interchangeable, and the systems work in an entirely predictable fashion. Clockwork, computers, and machines in general serve as excellent and valid models of the world.
On the other hand we have connectionist AI, Common Law, situational and relative ethics, contemporary physics and math where uncertainty, relativity and incompleteness are inescapable, probabilistic rules, dependent upon context and history, webs of interdependencies, emergent properties and consensus. The world and its workings require subjective language. The identity of participating entities and the context in which they interact are important in determining the outcome and meaning. Predictions are probabilistic and systems need to be self correcting. Living organisms, the weather, social relationships and chaotic systems are better models of reality than machines are.
Neither of these approaches is exclusively right. Both are valid descriptions of how our world works. Both are valid ways to operate. Neither solves all of our problems without reference to the other. Still, it is very hard not to emphasize one more than the other, even to the exclusion of the other. That is both natural, and OK. However, we do need to understand the role of both models, and to understand that the people around us may be viewing the world, explaining and understanding it based on a very different mix of these two approaches. Otherwise, we see them as irrational and frightening. We see them as the problem, rather than see our failure to communicate and to understand as the problem. We don’t all need to agree on everything, but it helps, a lot, to understand the worlds we each live in.