IEEE Technology and Society Magazine - December 2021 - 82

Soft Law Governance of Artificial Intelligence
governments justify the continued abuse of AI systems
to imprison statistically greater proportions
of brown and black people, even children, while
corporations implicitly claim that people have consented
to what arguably amounts to data slavery.
AI contextualization of law
There are few vanishing laws in any jurisdiction
in the United States that address AI bias and discrimination
directly [70] and none enacted at the federal
level [71]. Absent specific laws, then, the principal difficulty
in harnessing existing antidiscrimination laws
to require unbiased and nondiscriminatory uses of AI
is one of cross-disciplinary interpretation and contextualization.
Some exemplifying questions follow.
How is the requisite element of intentionality
to be determined where the discrimination occurs
or results from black-boxed AI computations? Will
the linkages between intention and discrimination
be rendered even more untraceable through
homomorphic encryption? What are reasonable
AI practices required under the unfairness doctrine
as one of the central consumer protection
tenets? How are big data-consuming computationally
intelligent machines qualified as expert
witnesses under Daubert [72] and Federal Rules
of Evidence 701 and 702 [73] where AI explainability
is still a nascent design goal and defendants
counter discovery requests by asserting trade
secret protections? How is privacy protected
where the not old, but vastly outdated notion of
reasonable linkability is a poor fiction where corporate
interests with scant nonidentity-revealing
data points can pinpoint the associated individual
with stunning accuracy? How may machine learning
systems be accountably verified as unbiased
and nondiscriminatory where their computational
power and terabytes of data enable them to learn
models that create opaque race or other suspect
class proxies through correlations that are impenetrable
to human cognition?
These myriad interpretive questions and more
demand answers on an accelerated time scale. The
rise of AI in daily life is not like that of the Internet.
In the latter case, interpretative actions seemed to
have occurred much earlier in its adoption. AI has
been at large in the world for decades, and some still
say self-regulation is the appropriate path. No need
to test algorithms before turning them upon unsuspecting
people and communities, some say. That
82
sounds more like self-interested nonregulation and
the rest be damned. Do not misconceive the view,
however. Powerful, innovative, and economically
productive tech companies are deeply important,
provided, however, they are responsibly run vis-àvis
society and beyond the primacy of shareholders.
The demonstrated risk and persistent and extended
nature of algorithmic harms argues, strongly, against
the more traditional laissez faire approach to market
regulation that has characterized U.S. policymaking
for decades [74].
Output from one AI system used in a cascading
series of other such systems propagate the harms,
but likely also compound them throughout the ecosystem.
In consequence, data contamination propagates
with such rapidity and at such scale that it may
be appropriately considered a digital contagion, an
unspeakable and pandemic harm engine. Together,
these mean that algorithmic harms from biased and
discriminatory implementations and uses of AI very
likely may be irreparable, unable to be remediated
and harming people, their families and communities,
and society as a whole.
What cannot stand is that biased and discriminatory
AI uses be permitted with impunity and with
no thought to the algorithm harms that accrue to
people, real people who want or have jobs, families,
homes, educational opportunities, and their freedom.
Unlike many other harms, AI-mediated harms
can become, like racism, more pervasive and permanent
within the information ecosystem. Furthermore,
and unlike other harms that may be siloed as
to their areas of impact, AI-mediated harms are propagated
throughout that ecosystem [74].
In summary, tHere appears to be virtually universal
agreement in soft law or its progenitors that
AI systems and processes should be unbiased and
nondiscriminatory. Current efforts, however, have
yet to produce actionable guidance on nondiscrimination
and antibias standards for AI. These efforts,
once mature, will serve as powerful, if not legally
binding, impeti for unbiased and nondiscriminatory
AI design and development. As to unbiased
and nondiscriminatory implementations and uses,
however, these soft law standards alone will not
suffice to achieve those aims. That said, proven
models exist for expediting the formation of soft AI
governance and for incorporating soft law rapidly
into positive law so as to make compliance and
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
