IEEE Technology and Society Magazine - September 2021 - 82

Feature
as a way to account for its existence and hegemonic
societal impact.
This raises questions on the role of reflective judgment
as a fundamental act of interpretation for problem
formulation. The exercise of reflective judgment
[1] is understood here as a cognitive expression of
an " enlarged mentality " that accounts for particulars
without necessarily subsuming them under
collective propositions which act as universal rules
or principles. The premise for exercising reflective
judgment is found in the view which defends the
permanence of problems in ML. To put it differently,
problems exist as ontological realities if and only if
an autonomous reflective capacity is exercised and
inscribed temporally at the level of the infrastructure
of the ML workflow. Rosales states, " In industry, you
never run out of problems that there is no answer
for; or that the answers are not that great when you
try them. This creates an influx of problems that you
wish you had more time to solve " [7]. This means
that judgments are not only instrumentally useful
to provide solutions to existing problems, but most
importantly reflective judgments also establish the
autonomy of problem-making as an activity worthy
in its own right. We need to distinguish here between
problem-making and problem-solving where the
making of a problem is logically and ontologically
prior to problem-solving. The point here is not only
that the activity of problem-making is a permanent
fixture of ML ecosystems, but that problems are also
epistemologically and ethically irreducible. The reason
is due to the essentially contingent and ambiguous
nature of ML problems dependent on multiple
and shifting parameters of analysis.
Paths forward
In what follows, we will chart a few paths toward
responsible problem formulation in ML. We propose
two questions in this respect. How should problem
formulation operate as a practice of evaluation and
interpretation of algorithmic design and dataset
construction? That is, what are the basic existential
assumptions which shape problem formulation? The
immediate response is that problem formulation
involves context-dependent assumptions associated
with particular data science projects. One general
assumption that we can point to, however, is related
to what we showed above as a solutionist bias toward
action and an " apologize later " attitude. Here, we
have many examples to unveil this attitude such as
82
" teachers being fired by an algorithm, " " essay grading
software, " or biased health-care algorithms [8].
To this generic assumption, we want to oppose
an ontological one. One such ontological assumption
in conceptualizing problem formulation lies
in the recognition that we need to grant epistemological
and ethical dignity to ML problems via the
interpretive role of problem formulation. While ML
problems do not simply exist in the pure realm of
thought as Aristotle's god and should not be treated
for their own sake, we need to be able to accept the
irreducibility of problem formulation in ML both
as an ethical practice and a value. This means that
problems are not simply useful and should not be
dealt with on an instrumental basis. The reason is
due to the fact that the approach of technically useful
problem-solving is not necessarily construed as
good for all, but can and is indeed harmful to many.
Also, solutions are never simply that which resolves
and closes a problem. Instead, within socio-technical
systems, solutions are in fact generating both preventable
and unforeseen problems. That is why the
position that we referenced at the beginning of this
article which dismisses problems and focuses on
narrow action-driven solutions is simply untenable
and inconsistent with public norms of responsible
creation of technical objects.
Instead, we should focus on creating a space-
time for problem formulation in the ML workflow,
that is, an interpretive space for moments of reflective
judgments in the constitution of ML problem formulation.
These reflective judgments, we showed,
give shape to an epistemology of ethics. Moving forward,
our computationally defined epistemological
problems should be mediated by three modes of
ethical rationalization: practices of unveiling historical
reasons for technical artifacts, strategic pluralism,
and contestability.
First, in the context of problem formulation, we
need to develop methodologies to historicize metrics,
that is, document their constitution and performance
by analyzing their previous effects and results
on given populations or constituencies, and by placing,
evaluating, and comparing them in defined contexts
of use.
Second, we need to reflexively create conditions
for a pluralist ethical commitment in our relation
to metrics. In themselves, pluralist ethical commitments
should be aligned to social good values
such as meaningfulness and inclusivity with the
IEEE Technology and Society Magazine
IEEE Technology and Society Magazine - September 2021

Table of Contents for the Digital Edition of IEEE Technology and Society Magazine - September 2021
