IEEE Technology and Society Magazine - December 2021 - 58

Soft Law Governance of Artificial Intelligence
only possible if they act from a position of scientific
authority, commercial neutrality, normative surety,
and delegated authority. These are considerable
assumptions as the conditions which made the establishment
of other review board do not pertain to the
situation of AI. As yet, there has been no " Nuremburg
Trial " to motivate public concern about unethical
AI research, nor has there been a watershed magazine
investigation into AI data sets, such as was the
case which spurred the Animal Welfare Act of 1974
[17]. Instead, AI review boards borrow the language
of their historical homologues and inherit the problems
known to bedevil the review board infrastructure
for research, such as arguments that they are
mechanisms for institutional control or protection,
not control of unethical or unsafe research or protection
of humans or animals from risks associated
with research participation [18], [19]. Furthermore,
AI review boards act without a source of legitimate
legal authority to conduct reviews.
What chance is there that AI review boards might
emerge as a mechanism to govern this technology?
There is a glimmer of hope that these bodies might
be supported by the vast " gray " structures of soft law,
such as through standards or even through organizations
whose position in a technical sphere lead to
widespread acceptance and later authority due to
convenience, as is the case with ICANN [20]. If soft
law is to be a source of legitimacy and authority for
AI review boards, how might this work?
Concepts
Soft law might be the most fruitful area for examining
sources of authority for creating AI review
boards. As each of these concepts-AI, review
boards, and soft law-are contested, I briefly define
them here. Then, in subsequent sections, I show how
the lack of clear definition creates additional challenges
for soft law governance in this space.
Artificial intelligence
The goal of AI is to answer questions, both simple
and complex, through regular procedures implemented
by neutral, nonhuman, and error-limited
systems. Knitting together logic, mathematics, statistics,
and symbolic reasoning, AI is expected to
answer questions ranging from " what is that smell? "
to " how can we organize all of human knowledge
into interconnected nodes that provide answers
to any question? " [21]. Hardware, software, and
58
Review boards
When specifically designed to mitigate risks to
human or animal participants, review boards are
institutions that apply ethical principles to specified
research problems to produce actionable recommendations
that improve a project's risk-benefit profile
[28], [29]. Advocates for AI review boards often
invoke this model because in addition to abiding
by regulation, they ensure that researchers abide by
IEEE Technology and Society Magazine
analytical methods are avenues of basic research,
applied research, and customer-facing industry products.
In its research modes, AI researchers study the
performance of different forms of algorithm vis-a-vis
one another or when applied to new contexts, as in
transfer learning. In its applied modes, AI is understood
as the use of heuristic reasoning systems, knowledge
engineering systems, unsupervised clustering
methods, or recurrent neural networks to answer
computationally complex questions that build a product
satisfying a specific domain need. In its customer
facing modes, AI is a component of systems that satisfy
consumer demands, whether that is for efficient
and low-cost electricity or for recommendations for
the best K-pop song to listen when running.
An important characteristic of all AI is that it is
beset by more " failure " than success. These systems
may fail to attain the state-of-the-art performance
benchmark or achieve an intended task at all. In
the production environment, it is estimated that
over three-quarters of proposed AI systems are not
forwarded to the users' environment because they
did not achieve adequate performance or acceptable
safety [22]. While failure of AI in the research
environment causes consternation for researchers,
funders, and, increasingly, environmentalists, it is
AI in the production environment that raises the
most significant questions posited for review board
consideration. Lack of success in these scenarios is
more concerning than not achieving a benchmark
score on conventional tests. This is because production-level
AI fails when it does not meet social or
ethical normative expectations, such as equality, justice,
and autonomy. The goal of AI review boards is
to ensure that these events are limited. Alas, because
these are dynamic systems that operate on a scale of
breadth, depth, and speed of calculation that defy
request for human readable explanation in many
cases, it is difficult to say what a review board must
do to control AI failures [23]-[27].
IEEE Technology and Society Magazine - December 2021

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