IEEE Technology and Society Magazine - March 2022 - 71

underlying mechanism that is not interpretable by
people. Often even the programmers who created
a neural network do not understand how it works,
nor why it reached the decision that it did. In sum,
it is possible to have a very effective machine-learning
model-in the sense that it makes very accurate,
useful decisions on complex tasks-but whose inner
workings are comparatively difficult, if not impossible,
for people to understand and interrogate.
Such inscrutable technological models may be
particularly problematic within law. Basic legal principles
require that legal decision makers be able to
explain why they came to the decisions that they did.
Articulated rationale is a central tenet of legal decision
making, particularly when decisions involve
the deprivation of liberty or property. However, to
the extent that legal officials are assisted by artificial
intelligence systems that have core interpretability
limitations, such articulable rationales may not be
possible, undermining central legal norms. Moreover,
as previously described, such uninterpretable
mathematical models may further mask underlying,
and undesirable, biases that may be difficult to
detect by human inspection.
However, while it is generally desirable to make
machine-based legal decisions more interpretable,
it is important to acknowledge that such technical
changes will not solve underlying structural-legal
problems where they exist. In some cases, it may
make them worse if not done carefully and in good
faith. As discussed earlier, machine decisions often
reflect a superficial veneer of value neutrality, when
in fact the technology can contain subtle, embedded
value preferences. The concern is that improving
a technical aspect of a legal AI system may
leave the false impression that a deeper, underlying
legal-societal issue has been resolved and no longer
requires attention.
Formalizing law and rules-based systems
A different class of artificial intelligence techniques,
grounded in computer logic and rules, provides
distinct embedded values issues within law.
This rules-based approach typically involves the
formal representation of laws on a computer system
using computer logic [13]. In such an approach,
computer programmers, lawyers, and others examine
particular laws and attempt to translate them
into a set of comparable rules that a computer can
follow, while attempting to preserve the underlying
March 2022
logic and meaning of the laws. The translation of
law and legal relationships into computer structures,
such as rules and ontologies, is sometimes
called formalization.
Many such systems aim to aid in legal decision
making or to allow the assessment of legal outcomes.
For example, in the United States, software
systems, such as Turbotax, assist citizens in filing
their personal income taxes. These software systems
attempt to formally model a portion of U.S.
personal income tax laws; they aim to replicate the
underlying logic of the laws in computer processable
form and allow for computation of tax liability
under the law [26].
A major issue is that the very act of " translating "
laws into computer rules actually masks a series
of subjective and contestable decisions about the
meaning and scope of the law. Many applications
of law involve unpredictability-uncertainty about
which laws do, or do not, apply to a given situation,
how these laws are to be interpreted, and once interpreted
what the outcome will be when applied to
particular facts. In a typical legal scenario, there are
multiple plausible interpretations of a given law,
each with slightly different meanings and scope, and
all which are reasonable. One of the major roles of
legal officials-such as judges-is to resolve these
uncertainties in applying laws in particular circumstances,
taking into account considerations such
as the text, meaning, and purpose of the law, and
competing public policies. Society often does not
often know a definitive answer about such legal
uncertainty until a legal official makes a binding,
final determination, electing one set of possible
arguments and interpretations over others.
By contrast, the process of " translating " a law
for a computer system necessarily involves a series
of judgment calls about the meaning, content, and
applicability of the law (and other legally uncertain
issues) on the part of the translator. Thus, to formalize
a law in as a series of computer rules is to commit
to one particular legal interpretation over others.
This process involves implicitly choosing one set
of contestable arguments about the meaning and
scope of the law over other plausible readings. Problematically,
it may not be obvious, even to the programmers
who engage in such formalization, that
translating a law into comparable computer code
actually involves an implicit set of subjective, and
perhaps value-laden, interpretive choices.
71
IEEE Technology and Society Magazine - March 2022

Table of Contents for the Digital Edition of IEEE Technology and Society Magazine - March 2022
