*2 This characterization is certainly true of
private international law and civil procedure. Both areas are
complex, and some may consider them to be boring; yet, procedure
is crucial to law practice. Substantive rules stem from
national, international, regional, federal and local sources.
These substantive rules must then be applied in domestic courts
subject to equally diverse procedural rules, rapidly resulting
in dizzying complexity. This complexity is somewhat offset by
the mechanical and straightforward nature of the rules of civil
procedure: Although there are many procedural rules, the rules
are determinate (few in number and reaching precise results).
The procedural rules, at least, follow basic mechanical formulas
with Boolean true/false outcomes that result from conjunctions
and disjunctions of conditionals. Such formulas lend themselves
well to modelling by computer. This paper discusses modelling
the law by computer.
*3 Computer applications for legal
problem-solving have progressed from mere text editors to case
law research to automated form generation.
1
Today we see computers used as intelligent agents
2
tasked with solving specific legal problems.
3
Can artificial intelligence solve legal problems? Will the
ability of computer programs to solve legal problems have real
life applications, or is it merely an intellectual curiosity? To
what extent and in what ways can artificial intelligence help
real lawyers with real legal problems?
*4 Computer programs can indeed solve legal
problems. The fact that computer programs can model law is not
necessarily simply of academic interest. Automated case research
is one potential application of intelligent programs. When
artificial intelligence determines a solution to a legal
problem, it could then automatically fetch relevant cases from
online or off-line statutory and case law databases.
4
"Spiders"
5
crawl through online databases all the time; why not adapt this
technology to law?
*5 Efforts have been made to use computer
programs for automated search and retrieval from legal
databases.
6
Computer intelligence can also be used as a backstop to keep
lawyers from missing obvious issues and to provide potential
lines of argument and defences to the litigator. Using the
computer as a backstop is far from using the computer as a
judge; however, automated search and retrieval, as well as
checklisting a lawyer's work, are tasks well within the
computational power of contemporary machines.
7
*6 Because procedural rules are mechanical,
they lend themselves to computer modelling. The complex yet
mechanical nature of procedural laws, particularly in the
context of international law, explain why computer modelling of
complex mechanical rule structures such as civil procedure,
conflicts of laws/private international law may be a useful tool
for practitioners. The computer is less likely than a human to
overlook any of the Byzantine exceptions or exceptions to
exceptions that may result in the application or non-application
of a foreign or domestic procedural or substantive law.
Computers are not more intelligent than humans. Humans are far
more creative than the computer programs that they write.
Computers, however, are more systematic and less prone to error
in simple repetitive tasks than humans.
8
This author is of the opinion that artificial intelligence can
play a useful legal role as a diagnostic and a checklist.
Artificial intelligence can act as backstop for human reasoning
to prevent human error, such as oversight or omission of
potential claims and defences, and guide potential lines of
argument.
*7 Artificial intelligence programs can be
divided into programs that are general or expert systems of
intelligence. General systems are computer programs that attempt
to simulate intelligence generally, or with no fixed limited
class of problems.
9
Consequently, programming a general system can be very
difficult. Further, because general systems are relatively
impractical, they are rare.
10
In contrast, an expert system is a computer program geared
toward solving one limited class of problems. Expert systems
infer implications from a given knowledge base.
11
This knowledge base may be static, pre-programmed and
unchanging, or dynamic and capable of evolution.
12
Dynamic rule bases may be better at representing intelligence
since the evolution of the rule base reflects the program's
ability to "learn." Programs that play chess generally use
static rule bases, though some chess-playing programs use
dynamic rules and adapt themselves to their opponent.
13
Most artificial intelligence applications, including law
applications, are formulated as rule-based expert systems.
14
But just what is that intelligence trying to model? How does the
human brain actually work, and to what extent does it work
differently than the computer?
DPA1
*8 Neuroscience
15
has now determined what computer science has surmised,
16
that human brains and most computers operate quite differently.
17
Specifically, human brains appear to be analog,
18
whereas contemporary computers are nearly always digital. While
today's microprocessors almost universally represent knowledge
in binary states (true/false; yes/no; on/off), humans represent
knowledge in analog states (warmer/colder; brighter/darker).
19
While analog computers are possible (for example, a slide rule
is an analog computer),
20
virtually all of today's microprocessors are digital because a
sufficiently fine digital representation is indistinguishable
from an analog representation, and it is also easier to store
and transmit.
21
*9 A computer and the human brain are not only
different because the right hemisphere of the brain functions
using analog principles but also because the human brain is a
massive parallel processor ("MPP").
22
While it is possible to emulate parallel processing using
several networked Central Processing Units ("CPUs"),
23
none of the major desktop CPUs use parallel processors.
24
In parallel processing, one part of the brain (or one CPU) works
to solve a problem at the same time as another part (or a
different CPU) works on the same problem.
25
The parts of the brain then compare answers, and if they agree,
the brain then moves to the next step.
*10 Although the above explanation of the human
brain and parallel processing is simplified, it does explain how
the human brain works. The brain tries to get an answer. If it
finds no answer to the current problem, it either backtracks to
an earlier answer or skips forward to a new problem, hoping that
by solving the other problem it will gain insights on the
skipped problem. At the same time the brain is forward and
backward chaining its search tree, the brain is also comparing
search strategies by a dialogue between the left (execution) and
right (creative) hemispheres.
26
Thus, the brain, unlike most computers, is engaging in parallel
processing.
27
*11 The vast majority of computer processors
today are not parallel processors; instead, they are serial
processors.
28
In fact, a microprocessor is simply a very fast and perfectly
accurate adding machine (the CPU) with several abaci
29
attached to store results (the "registers").
30
Microprocessors, at present, are not at all creative. On the
other hand, microprocessors tend not to forget, at least until
you pull the plug.
31
*12 Unlike the brain which has at least two
processors (namely the left and right hemispheres), computers
today do not generally assign a problem to two different CPUs,
32
skip backwards and forwards in aleatory searches for tentative
solutions to interrelated problems, or periodically compare the
processing to other CPUs.
33
The right hemisphere of the brain handles creative, holistic
tasks and the left hemisphere is dedicated to linear
computation.
34
Most computing is not done in parallel. Instead, one main chip
and possibly a math co-processor do all the calculations in a
linear fashion. The machine will always return to whatever it is
told to return. Current chip technology and software do not
include native creative functions other than pseudo-random
numbers generated by reference to the computer's clock.
35
Contemporary CPUs, like their predecessors twenty years ago, are
simply blindingly fast and nearly infallible adding machines
that are able to compare and store values.
*13 Of course, it is possible to do parallel
processing with software using networked computers. Although
this was not the origin of computing, it may be the future.
36
*14 Serially processing data,
37
or thinking like a traditional serial microprocessor, is
essentially a linear function. The serial processor steps
through each command sequentially. Commands are run only
sequentially, and results are not compared to the results of
outside processors. Computers may evolve toward parallel
processing, as we can already see in distributed computing
applications such as SETI.
38
However, very little work has been done on programming computers
to emulate human creativity, other than generating random art
39
or random poetry.
40
Perhaps this is due to the fact computer scientists tend to
think sequentially, whereas artists tend to think holistically.
*15 Artificial intelligence ("AI")
41
has evolved sporadically and, despite remarkable initial work,
has stagnated to some extent. AI guru Marvin Minsky recently
stated in a speech at Boston University that "AI has been
brain-dead since the 1970s."
42
AI's "brain-death" is not due to any computational limits, but
is simply due to the fact that other problems were more
profitable. However, profitable areas of AI, such as machine
translation, have indeed kept pace with other programming
achievements of the last several decades.
43
*16 Early computer scientists originally
thought that artificial intelligence would be the defining
characteristic of computational power.
44
Alan M. Turing proposed that machine intelligence would be
considered "intelligent" to the point where a user would not
know the difference between the machine and a person.
45
The "Turing Test" has since generated much scholarship
46
and some criticism for concealing as much as it reveals.
47
The ability to mimic a human successfully has not, in fact,
turned out to be the
sine qua non of computer
intelligence. The famous computer program "Eliza" demonstrates
this development.
*17 "Eliza" was one of the first successful
attempts at creating a machine that could interact with a human.
48
Eliza was intended to simulate a psychiatrist by mirroring the
information provided to it by the client.
49
Eliza is an intellectual curiosity because, despite being rather
primitive, it does meet the Turing Test, as people often believe
that Eliza is "intelligent" and "human."
50
This anthropomorphization is pre-scientific, and it also shows
that Turing's Test is not as objective as we might first think.
Brighter people are much less likely to be "fooled" into
thinking that the computer is a person. In addition to being an
achievement as a successful language parser, Eliza has
successfully demonstrated the limits of Turing's Test.
*18 Arthur Clarke, like Alan M. Turing,
51
also focused on artificial intelligence as a key definitional
characteristic of the future of computer science. In the 1960s,
Clarke thought that computers in 2000 would still be very big
mainframes and would have vast memory banks that would allow
them to be self-aware and able to interact in natural language.
52
Instead, we see today a global network of small, powerful
computers that are rarely parallel processed to create a
super-computer. Because existing super-computers do rely on
massive parallel processing
53
and could rely on neural networks, but do not even attempt to
emulate human processes,
54
the initial vision of artificial intelligence was clearly
erroneous.
*19 Clarke was correct, however, in predicting
a quantum leap in computational power. Computers today literally
have around 60,000 times more dynamic storage capacity ("RAM")
than computers of the mid-1980s.
55
Programs such as A.L.I.C.E.
56
and Babel Fish are able to communicate in natural language.
57
These programs have easily over 1000 times more static memory
storage, or hard drive space, than computers of the early 1990s.
58
Computers of the 1960s only had 1/1000th of the storage capacity
of a computer of the 1990s.
59
Processor speed has also increased by several hundred times
since the 1980s, while storage capacity has increased even more
rapidly.
60
Clark's prediction was least accurate as to size and network
capability. With the exception of industrial strength servers,
today's computers are small and globally networked. This is
because modem speed has increased from 300 bits per second
("bps") to 56,000 kbps for dial-up, and literally megabytes per
second on cable.
61
Microsoft founder Bill Gates did not even expect this rapid
increase.
62
These improvements are illustrated in the following table:
*20 These hardware changes have for the most
part out-paced software development.
73
While software development has also advanced rapidly, software
manufacturers have had difficulty keeping pace with hardware's
rapid improvements.
*21 Clarke's prediction was most accurate as to
memory storage. His computer, HAL, had a memory as extensive as
human memory with massive arrays of data at instant disposition.
74
Clarke was also correct about a computer's ability to process
natural language. On the other hand, his expectation that
computing would still focus on isolated, non-networked massive
mainframes was inaccurate. Basically, Clarke correctly predicted
that massive changes would occur, but was incorrect as to his
specific predictions for what those changes would be.
*22 Thus, Clarke's prediction of a self-aware,
non-trivial artificial intelligence program (HAL 9000)
75
was inaccurate. This inaccurate prediction, however, is not
because the task of parsing natural language is impossible.
Rather, the problem exists because attempts to achieve sentience
lack commercial application and are politically unacceptable. In
the 1960s, creating an artificial intelligence agent to meet
Turing's Test was seen as at least an interesting research goal.
However, such projects have not proven profitable. Attempts to
emulate parts of intelligence via expert systems have been the
recent focus of research and applications in artificial
intelligence.
76
*23 Some efforts to approach the problem of
simulating human intelligence using parallel processing, i.e.,
distributed computing,
77
do exist. This may actually be the better way to emulate
sentience. One major problem with an intelligent human-computer
interface is simply determining how to parse speech. Although
parsing speech may be computationally complex due to the fact
that the program must take context into account, it is not
impossible. By distributing the problem-solving mechanisms via
the Internet, parallel processing presents the possibility of
generating a reasonable simulation of human intelligence. The
goal of making computers self-aware, however, raises two
questions: what is "self," and what is "awareness?"
Philosophers, since Descartes' discussion of solipsism
78
in
Meditations on First Philosophy, have tried,
unsuccessfully, to answer such questions. Simulating
intelligence is not impossible, but, given these lingering
questions and the present state of technology, computational
sentience is the stuff of science fiction for now.
*24 Having discussed some of the background of
machine and human intelligence and the standards and measures of
computation, I would like to focus now on how this background
information can be applied to creating a modest computer program
to formally model the law. The computer program which
accompanies this paper seeks to determine whether jurisdiction
exists in the United States for a claim under either the Alien
Tort Claims Act ("ATCA") or Torture Victim Protection Act. If
jurisdiction does exist, it then considers procedural defences.
If no procedural defences exist, it then determines whether a
substantive violation exists. Finally, it generates a report. To
make these determinations, the program must prompt the user to
supply a series of facts. The program does not, however, examine
each element of a tort.
79
*25 The ATCA program accompanying this paper
essentially deduces its conclusions based on pre-programmed
rules and the information supplied by the user. It does not
learn new rules of production or modify its existing rules of
production by deriving new rules from existing ones. The program
reasons deductively, not analogically. After reaching its
conclusions, it generates a report listing the reasons for its
decisions.
*26 Though this program uses deductive
reasoning to reach its conclusions, the common law generally
reasons inductively.
80
To be exact, the common law uses inductive reasoning when
arguing analogically, or by ampliation from existing cases. It
uses deductive reasoning when arguing from statutes. Reasoning
by analogy, i.e., inductive inference, is a very different
operation than inductive ampliation.
*27 Some authors believe that analogical
reasoning is impossible for a computer to model.
81
Such a position may be the result of confounding inductive
ampliation and analogical reasoning. Analogical reasoning and
inductive ampliation are not equivalent algorithms.
82
Analogical reasoning is reasoning from one case to another
similar case.
83
Ampliative induction involves examining and comparing several
known cases in order to derive a new general rule, and then
applying that rule to new cases.
84
Although these processes are similar, they are discrete.
Unfortunately, due to a lack of rigor, clarity, or intellectual
discipline, common law lawyers sometimes ignore this
distinction. Understanding this nuance is one key to
understanding why, and how, inductive reasoning can be modelled
by computer.
*28 Developing an analogical case base would
require more complex algorithms than a deductive rule base.
However difficult the task may be, it is not impossible. A goal
of future research is to develop an inductive solution to this
problem.
*29 Computer programs that allow case-based
analogical reasoning from an existing (static) case base to be
applied to a new case exist today.
85
The next step will be to induce new rules of production from a
dynamic case base that can evolve based on user input. Such a
program would be at least three times more complex than the one
presented here and, theoretically, would be able to model any
area of law. The inquiry in this paper, however, is limited to
an existing, well-defined area of law. Future research will
develop ampliation from dynamic case bases to reason inductively
as well as deductively.
*30 Neurologically, the distinction between
inductive, case-based reasoning and deductive, rule-based
reasoning may be a reflection of the polar differentiation in
the human brain. This differentiation is not found in current
CPU architecture.
86
The specialisation of different hemispheres, one handling
logical, linear tasks such as computation, and the other
handling holistic, creative tasks such as language, along with
the integration of these two aspects, are what distinguish human
reasoning from machine reasoning. Thus, rather than defining a
test for artificial intelligence as Turing does,
87
this author would suggest that the best test of artificial
intelligence would be whether the artificial intelligence can
emulate both creative tasks (i.e., analogy and inductive
ampliation) and linear, computational tasks (i.e., deduction),
and then integrate these two processes to allow the program to
solve new problems which are similar to, but not the same as,
existing, solved problems. The ability to input tasks or output
task results with the use of a natural language such as English
would be practical, but actually would be a secondary measure of
machine intelligence.
*31 The formalization of legal rules has
pedagogical value because it forces legal scholars to think
rigorously and systematically about the law. Formalizing the law
to accurately reflect its results forces one to reconsider
reductionist theories of law. Unfortunately, the majority of
contemporary legal theories seem to be reductionist theories.
88
For example, legal realists argue,
simplicitur, that the
law is merely a rationalization of power.
89
Legal economists contend that the law is a function of
economics.
90
Legal process is based on law as a function of public policy.
91
Each of these theories may be accurate, but each is also only
partially accurate. The process of formalizing law makes evident
to the legal scholar the limitations of each of these theories.
Further, parsing statutes and cases in a disciplined fashion
reveals some of the finer shortcomings of each position. The
realists' critique of formalism seems somewhat weaker if one
closely studies legal doctrines such as
expressio unius.
While the realists
92
are right that some legal maxims clearly contradict each other,
not all do.
93
Similarly, no one economic or policy argument can adequately
explain all of the law. Legal economists who argue that
"economic efficiency" determines the law ignore the problems of
information costs, externalities, and non-fungible goods; these
theorists instead downplay public goods, likely with political
motivations.
94
While these weaknesses in legal economic theories of law may not
be explicitly clear when one is parsing a statute, such
weaknesses do reveal themselves implicitly when one confronts a
mass of apparently conflicting rules, principles, policies,
maxims, and economies. Ideally, legal programming will help
scholars see the weakness in blanket generalizations.
Formalization can thus encourage creative scholars to adduce
theories to explain these shortcomings. The formal structure of
the law is not purely or uniquely rationalization,
implementation of public policy, or an economic balance.
*32 While formalization does implicitly reveal
the theoretical shortcomings of contemporary legal theory, one
cannot predict what new theory or theories would emerge. This is
because the formalization underlying the computer program itself
reflects meta-theoretical assumptions. In this regard, human
intelligence is, at present, clearly superior to machine
intelligence. Although humans, unlike computers, often forget
facts or make mistakes, they are capable of synthesizing
theories that creatively go beyond existing formal rules of
production to generate new ones. They are able to apply
abductive, intuitive, and aleatory operations that are not part
of the basic instruction set of a CPU. This explains some of the
shortcomings of artificial intelligence.
*33 Just as legal scholars may gain insights
about law-making from studying and writing computer programs
about law, law professors may also find programming beneficial.
Instead of seeing law as an amorphous mass of ambiguous and
competing cases, statutes, and constitutions, the professor is
forced to put legal chaos into some sort of order and synthesize
the law into a coherent whole faster than her students.
*34 Such attempts at legal ordering are
necessarily Sisyphean: with each day new cases, and with each
year new statutes, appear in legal discourse. Anti-formalists
will even argue that there is no legal order, that it really is
just chaos.
95
Others will argue, perhaps more creatively, that the legal order
is self-referential, and recursivity and autopoeisis are what
define the law.
96
However, the classical enlightenment legal scholar,
97
and even her pre-enlightenment natural law counterpart, will
argue that law is living logic, that it is necessarily ordered
and that a chaotic law is no law at all.
98
To develop a good program, the legal scholar must put himself in
the shoes of classical or enlightenment rationalism. In order to
understand the program's limits, the jurist must at least be
aware of the radical critiques of legalism and formalism when
using programs as diagnostics or predictions.
*35 The content of the program itself is really
best exposed by use of the program. Try running with the
program, play with it, try to "break" it (and please report any
bugs to the author if you do). The program will run through some
straightforward legal tests and, hopefully, reach correct
conclusions in its summary report.
*36 Although computer applications in law have
expanded from simple word processing to electronic research and
animated trial aids, there have been relatively few applications
of artificial intelligence to law. This is partly because AI is
still a developing technology due to neurological asymmetries
described in this paper. Expert systems generally perform
limited tasks reasonably well, but AI general systems have not
yielded much success. Further, AI, unlike other areas of
programming, has not yet yielded profits.
99
However, AI algorithms do increasingly figure in commercial
programs such as speech recognition and machine translation.
100
AI can be useful not only as a tool to teach legal reasoning to
law students but also as a checklist for legal practitioners.
Future research will hopefully yield new types of
microprocessors that will be developed for AI applications.