IEEE Technology and Society Magazine - December 2021 - 32

Soft Law Governance of Artificial Intelligence
There has also been extensive work in new communities,
rightly focused not just on what AI can
do, but what it should do-and what it should not
do. These expanding multistakeholder groups are
advancing new ideas, issues, concepts, and needs,
and bring new stakeholders to the issues. AlgorithmWatch
catalogues over 150 different frameworks and
guidelines for ethical AI design [61]. Jobin et al. [62]
analyzed 80 such documents, and a team from Harvard's
Berkman Klein [63] examined a subset of 36
frameworks. Building on these and other analyses,
Gutierrez and Marchant [64] ran a comprehensive
study of 634 AI soft law programs. Even with significant
variations among approaches, these AI soft law
initiatives reflect convergence on several subjects
that they aim to mandate or encourage, including
transparency and explainability, fairness and nondiscrimination,
safety and security, as well as responsibility
and privacy [61]-[63]. Important ideas will
continue to compete for adoption-for example,
whether the foundation for these frameworks and
developer behavior is best guided by ethical principles,
private standards, guidelines, among others
[64]-[66]. These debates are valuable and need to
play out. The subject matter in discussion, though,
has converged enough that now is, indeed, the time
for multistakeholder dialogue on more specific rules
of the road for AI.
There is wide recognition that soft law on its own
will not be enough [67]. For AI soft law to gain credibility,
it needs a " bottom-up " specificity to allow
real assessment of whether organizations or projects
conform to those rules [29], [46], [68]. To begin, just
as cloud services adoption was supported by the system
of certification and assurance mechanisms to
provide confidence that services met expectations,
credibility of AI soft law will require guidance that
is sufficiently specific for us to hold someone to it-
whether through certification or other assurance or
through contractual or hard law means [29], [46].
High-level principles need to be supported by specific
requirements that AI services and developers
can be evaluated against. Those requirements need
to be accessible and meaningful to a wide range of
implementers and understandable by stakeholders
and beneficiaries. Striking an effective balance
in developing these specific requirements and the
measures to evaluate them is not a trivial task-the
array of AI methods and applications pairs with the
rapid development of these technologies to make
32
generalizing difficult, especially as applications
become increasingly specific to industry segments
and narrow use cases. The techniques for developing
soft law initiatives and advancing their credibility
will need to be tailored to this evolving mix
of horizontal and vertical needs [68]. As it has in
other areas, the international standards system can
provide a key starting point for this, with early initiatives
already well under way at ISO/IEC and IEEE
[69], [70].
Notwithstanding the aspirational tone seen
in many developing AI guidelines shown in the
Algorithm Watch catalogue, they clearly intend
to do more specific work than simply coordinating
the behavior of industry players. These ethical
guidelines and frameworks explicitly intend to
help both developers and society avoid some of
the many negative externalities and dangers that
AI technologies and implementations could create.
Though they may seek to drive behavior as a
matter of conscience rather than regulation, on the
spectrum of soft law types, these are decidedly regulative
efforts. They are asking developers to take
steps or forbear from them in ways that require
specific decisions, controls, and tradeoffs. These
initiatives demand much more thoughtful work on
credibility and adoption than coordinative efforts
alone or implementation guidance under a government
mandate [36]. Moreover, most of these
efforts intend to apply rules to actors beyond those
participating in their creation, setting a heightened
need for effective work on credibility and adoption
across these initiatives.
The race to develop the rules, hard and soft, for
AI has already begun. Nonstate, soft law programs
are an inevitable part of the complex and evolving
policy ecosystem. Less assured is which initiatives
will gain adoption and earn respect. Successful initiatives
will not " go it alone " -they will depend on
the ecosystem they are embedded in, and will have
close, effective links across that ecosystem. They will
take advantage of guidance and input in a continuing
dialogue with stakeholders. They may benefit
from architectures, opportunities, and credibility
signals communicated by established orchestrators
(e.g., G20 or OECD). They will need the support of
key intermediaries such as the systems that provide
all aspects of assurance that a commitment has been
delivered-for example, those who can provide
accreditation, audit, certification, and other forms
IEEE Technology and Society Magazine
IEEE Technology and Society Magazine - December 2021

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