Wednesday, June 14, 2017

What is Law? - part 14


Mention has been made earlier in this series to the presence of ambiguity in the corpus of law and the profound implications that the presence of ambiguity has on how we need to conceptualize computational law, in my opinion.

In this post, I would like to expand a little on the sources of ambiguity in law. Starting with the linguistic aspects but then moving into law as a process and an activity that plays out over time, as opposed to being a static knowledge object.

In my opinion, ambiguity is intrinsic in any linguistic formalism that is expressive enough to model the complexity of the real world. Since law is attempting to model the complexity of the real world, the ambiguity present in the model is necessary and intrinsic in my opinion. The linguistic nature of law is not something that can be pre-processed away with NLP tools, to yield a mathematically-based corpus of facts and associated inference rules.

An illustrative example of this can be found in the simple sounding concept of legal definitions. In language, definitions are often hermeneutic circles[1] which are formed whenever we define a word/phrase in terms of other words/phrases. These are themselves defined in terms of yet more words/phrases, in a way that creates definitional loops.

For example, imagine a word A that is defined in terms of words B, and C. We then proceed to define both B and C to try to bottom out the definition of A. However, umpteen levels of further definition later, we create a definition which itself depends on A – the very thing we are trying to define - thus creating a definitional loop. These definitional loops are known as hermeneutic circles[1].

Traditional computer science computational methods hate hermeneutic circles. A large part of computing consists of creating a model of data that "bottoms out" to simple data types. I.e. we take the concept of customer and boil it down into a set of strings, dates and numbers. We do not define a customer in terms of some other high level concept such as Person which might, in turn, be defined as a type of customer. To make a model that classical computer science can work on, we need a model that "bottoms out" and is not self-referential in the way hermeneutic circles are.

Another way to think about the definition problem is in term of Saussure's linguistics[2] in which language (or more generically "signs") get their meaning because of how they differ from other signs - not because they "bottom out" into simpler concepts.

Yet another way to think about the definition problem is in terms of what is known as the descriptivist theory of names[3] in which nouns can be though of as just arbitrary short codes for potentially open-ended sets of things which are defined by their descriptions. I.e. a "customer" could be defined as the set of all objects that (a) buy products from us, (b) have addresses we can send invoices to, (c) have given us their VAT number.

The same hermeneutic circle/Sauserrian issue arises here however as we try to take the elements of this description and bottom out the nouns they depend on (e.g., in the above example, "products", "addresses", "invoices" etc.).

For extra fun, we can construct a definition that is inherently paradoxical and sit back as our brains melt out of our ears trying to complete a workable definition. Here is a famous example:
The 'barber' in town X is defined as the person in town X who cuts the hair of anyone in town who do not choose to cut their own hair.

This sounds like a reasonable starting point for a definition of a 'barber', right? Everything is fine until we think about who cuts the barber's hair[4].

The hard facts of the matter are that the real world is full of things we want to make legal statements about but that we cannot formally define, even though we have strong intuitions about what they are. What is a "barber"? What is the color "red"? Is tomato ketchup a vegetable[5]? What is "duty"? What is "ownership"? etc. etc. We all carry around intuitions about these things in our heads, yet we struggle mightily to define them. Even when we can find a route to "bottom out" a definition, the results often seem contrived and inflexible. For example we could define "red" as 620–750 nm on the visible spectrum but are we really ok with 619nm or 751nm being "not red"?

Many examples of computing blips and snafus in the real world can be traced to the tendency of classical computing to put inflexible boundaries around things in order to model them. What does it mean for a fly-by-wire aircraft to be "at low altitude"? What does it mean for an asset to be trading at "fair market value"? The more we attempt to bottom these concepts out into hard numeric ranges - things classical computing can easily work with - the more we risk breaking our own intuitions with the real world versions of these concepts.

If this is all suggesting to you that computational law sounds more like a problem that requires real numbers (continuous variables) and statistical calculations as opposed to natural numbers and linear algebraic calculations, I think that is spot on.

I particularly like the concept of law as a continuous, analog process as it allows a key concept in law to be modeled more readily - namely the impact of the passage of time.

We have touched on the temporal aspects already but here I would like to talk a little about how the temporal aspects impact the ambiguity in the corpus.

As time passes, the process of law will itself change the law. One of the common types of change is a gradual reduction in levels of ambiguity in the corpus. Consider a new law which needs to define a concept. Here is how the process plays out, in summary form.

  • A definition is created in natural language. Everybody involves in the drafting knows full well that definitions cannot be fully self-contained and that ambiguity is inevitable. In the interests of being able to actually pass a law before the heat death of the universe, a starter definition is adopted in the law.
  • As the new law finds its way into effect, regulations, professional guidance notes etc. are created that refine the definition.
  • As the new law/regulations/professional guidance impacts the real world, litigation events may happen which result in the definition being scrutinized. From this scrutiny, new caselaw is produced which further refines the definition, reducing but never completely removing, the amount of ambiguity associated with the defintion.

A closely related process - and a major source of pragmatic, pre-meditated ambiguity in the process of law - is contracts. While drafting a contract, the teams of lawyers on both sides of the contract know that ambiguity is inevitable. It is simply not possible, for all the reasons mentioned above, to bottom out all the ambiguities.

The ambiguity that necessarily will remain in the signed contract is therefore used as a negotiating/bargaining item as the contract is being worked. Sometimes, ambiguity present in a draft contract gives you a contractual advantage so you seek to keep it. Other times, it creates a disadvantage so you seek to have it removed during contract negotiations. Yet other times, the competing teams of lawyers working on a contract with an ambiguity might know full well that it might cause difficulties down the road for both sides. However it might cost so much time and money to reduce the ambiguity now that both sides let it slide and hope it never becomes contentious post contract.

So to summarize, ambiguity in law is present for two main reasons. Firstly there is ambiguity present that is inevitable because of what law is trying to model - i.e. the real world. Secondly, there is ambiguity present that is tactical as lawyers seek to manipulate ambiguities so as to favor their clients.


[5] https://en.wikipedia.org/wiki/Ketchup_as_a_vegetable

1 comment:

Brant Watson said...

Your Gödelian loops are not far from the surface here. Even the paradox of the barber is just another form of the liars paradox. All of these loops arise from some form of self reference.

We pretend to get around some of these issues using inheritance and composition. For example we might have:


class Person(object):
name = charField()


class User(Person):
username = charField()

class Senator(User):
chamber = 'senate'

class Representative(User):
chamber = 'house'

If you have something like an instance of "Senator" called x you can do:

isinstance(x, Representative)
>> False

isinstance(x, Senator)
>> True

isinstance(x, Person)
>> True

isinstance(x, User)
>> True

This eventually still relies on some form of bottoming out; it's just that we hide a lot of the complexity by exposing a form of hierarchy. And that's the crux of it all. So long as we have and can obey the hierarchy we can have things that fit multiple definitions. A person can be both a User and a Senator which is nice, but you still have to assume that you have a definition which encompasses both. Is a Senator still a Senator if they lose reelection? Do you add "former" as a flag?

The refinements to caselaw you mention are essentially an abstraction of the concept of business logic which creeps in over long periods of time as people begin to understand that their imagined hierarchy wasn't actually a good representation of the ambiguity and flexibility of reality. Over time experienced developers learn to be more abstract in the way they conceive of things. For instance we abandoned the concept of a LPID for a UUID and the reasons for that involve aspects that go beyond the original schema. Suppose jones_leslie_1 gets divorced and there is a lot of animosity. Suddenly the permanent identifier that encodes names painfully associated with that emotional time in their life becomes a source of consternation. It's unchangeable in the data model because too many relationships are defined by this data. But an abstract UUID which carries no semantic meaning other than to be a unique identifier is immensely helpful. Semantic information is lost but the gains in abstraction make a UUID inherently more powerful.

All of these are of course pipe dreams in the long run. There is no known way to bottom it all out perfectly any more than there is a known solution to the halting problem or the liars paradox. Not if you want any guarantee of consistency anyway.


And that is a long way to get at what I think is an even more interesting idea. That any one of these processes of categorization involve a single idea; things can be cut into one or another category. A bifurcation of reality if you will. Defining "Person" assumes that all things can fit into one of TWO categories. "Person" and "Not Person". Each positive category assumes necessarily that there is another category which is it's negation. And it's always *easier* to assume you can define a category of INclusion more easily than one of EXclusion. The whitelist vs blacklist idea.

To whitelist you simply define criteria and everything not meeting that is considered to be NOT in your category I.E. it doesn't meet criteria? Oh then it's not a Person. Reverse that though and you get the impossible task of defining everything that makes something NOT a person. This list would be infinitely long and impossible.

It's the same reason that the liars paradox only works one way:

This sentence is False. <---- Paradox due to infinite regress
This sentence is True. <---- Bottoms out immediately due to whitelisting

The trouble with whitelisting though is that definitions are context dependent and relative because the meaning is merely an invention itself. Bottoming out a concept which is actually black and white is different from the task of bottoming out a concept which has shades of grey. You can cut and cut but your task is effectively infinite.