IEEE Technology and Society Magazine - December 2021 - 90

Soft Law Governance of Artificial Intelligence
algorithmic bias, although there are other concerns,
such as the ability to fool such systems into thinking
one person is another.
Privacy
People concerned with the privacy implications
of FRT worry that it will expose accurate personal
data to unknown or undesired parties. For example,
using facial recognition on a collection of social
media posts could identify a group of people that
frequently hang out together. Recognizing the same
face on multiple cameras across a city downtown
could be used to virtually follow an individual, showing
where they went and what path they took. Sentiment
analysis (using FRT to judge the emotional
response of an individual) could reveal information
about someone's shopping preferences in a grocery
store. Other critics of the technology argue that the
government's use of it will discourage individuals
from participating in peaceful protests for fear of
state surveillance [30].
As mentioned in the introduction, your face, at
least when in public areas or on other's property, is
not exactly private. In theory, FRT-related privacy
concerns could be replicated by a human observer.
Someone could review social media posts to identify
a group of friends. An investigator could tail a person
across town or look at security camera footage to do
the same. A waiter or clerk might recognize a customer,
identify their mood, and accommodate them
accordingly. The difference, and what concerns critics
of the technology, is that FRT potentially allows
what would be very labor-intensive work of observing
faces and matching them to individual identities
to be done rapidly and at relatively low cost-and
therefore much more frequently.
Bias
Many are worried that FRT can and is biasing
decision-making and creating discriminatory effects
[31]-[33]. Facial identification algorithms can produce
both false negatives (failing to find a match
to an individual in the database even though they
are included in the database) and false positives
(finding an incorrect individual as a match). Inaccurate
FRT can result in incorrect and harmful
decisions: in Detroit, a mistaken identification from
FRT led police to arrest an innocent man [34]. If
the level of inaccuracy is inconsistent across different
demographic groups, the algorithm is biased
90
and could result in worse effects for some groups.
Indeed, research has demonstrated that some FRT
systems are less accurate at identifying black faces
[35], [36]. Other recent studies of government FRT
databases have shown troubling problems when it
comes to misidentifying Asians, African Americans,
Native Americans, and Pacific Islanders. Because of
how many of these algorithms are trained, they may
be less accurate when applied to subgroups of the
population that are not well represented in the training
data [37]. Similarly, FRT presents issues of bias
for people with disabilities. Data sets to train FRT
applications often do not include people in wheelchairs
or people with Down syndrome, for instance,
and can disproportionately misidentify people with
disabilities or not recognize them at all [38]. While
other computational applications that rely on large
datasets raise similar concerns about bias, the FRT
concerns are particularly poignant because they
are evaluating photos, and, in a very literal way,
judging people by their skin color or other physical
attributes.
Bias is but one concerning form of inaccuracy.
For example, there is evidence that as photographic
databases grow larger, the number of false positives
can rise because there are many people who look
like others. While general accuracy of FRT systems is
a vital parameter (depending on the use), there are
strong commercial incentives to improve accuracy
while similar incentives may not hold for correcting
various types of bias.
Consequences of early hard law
approaches to commercial FRT
Motivated by these concerns, a handful of states
have enacted laws to regulate commercial uses of
FRT. While the evidence is not conclusive, subsequent
developments suggest that such laws have
had unintended consequences for consumers and
innovation.
Existing technology-neutral hard law, if enforced,
could potentially mitigate many of the most feared
effects of commercial FRT. Federal and state laws
like the Fair Credit Reporting Act, equal opportunity
laws, and state and federal consumer protection
and privacy laws already govern the collection of
data and prohibit discrimination against protected
classes of people [39].
In addition to these technology-neutral laws,
three states-Illinois, Washington, and Texas-have
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
