Transparency in Data Mining: From Theory to Practice - Discrimination and Privacy in the Information Society - page 308

by the analysis; with such data in hand they can have a better understanding of the
connection between their actions, the government's findings and their
implications. It also empowers data subjects who understand how their data was
used (although this is an overall weak argument). Thus, many of the theories are
aligned at this juncture, and lead to the conclusion that transparency at this
juncture is crucial, and must be attended to with vigor.
Segment (C) has generated the greatest interest in the context of governmental
data mining and proper disclosure. It also raises several interesting related issues.
First, we examine the notion of disclosure of the actual patterns used. The
arguments for transparency are strong; these are matters within the public interest,
and both shaming and political forces will be in place. Autonomy interests will be
in place as well - especially in terms of those affected by the process
(crowdsourcing arguments, however, are relatively weak). Yet at this juncture, an
ounce of realism is called for. In this context, the arguments regarding opacity are
of greatest strength; revealing the actual factors used will allow individuals to
circumvent the governmental objective. While taking into account the existing
legal rules and governmental sentiment, calling for transparency in this element
has no chance - and probably with good reason.
Yet this should not be the ending point of a discussion of transparency at this
juncture. Transparency could be reflected in other aspects of this segment. One is
interpretability - whether we must require that all relevant processes will be
understandable to humans even if the process is not disclosed to the entire public.
I believe such a duty is crucial. Furthermore, a requirement to set it in place could
be derived from transparency justifications.
Applying the various transparency theories to this specific issue easily leads to
the conclusion set forth above. Interpretability could promote effectiveness, via
fear of shaming. While the information revealed will not be shared with the
public, if really ridiculous factors are applied, such information has the risk of
leaking (and thus launching a shaming dynamic). Thus, the government will think
twice before using problematic correlations. To some extent, this requirement will
enhance the autonomy of those affected by the process. Individuals might not be
privy to “the logic” behind the decision, but will at least know someone is looking
into the matter, and has additional tools to do so. I would thus recommend that all
processes be interpretable, even at the cost of lowering overall efficiency.
The same arguments cannot hold, however, when examining a call for
causation on the basis of transparency-related considerations. On its face,
transparency might call for developing causation theories prior to using predictive
proxies in the field. A causation requirement could be derived from the
transparency theoretical framework. A causation requirement will promote
effectiveness as an important check on governmental actions. Causation will also
generate public interest. Developing such models, even internally, will enhance
autonomy, as an additional element to assure the process is not arbitrary. These
studies might also be built into a “crowdsourcing” dynamic (even when only
shared minimally). Experts will examine the strength of the causation theory, or
try and come up with an alternative one.