Page 147 - ITU Journal - ICT Discoveries - Volume 1, No. 2, December 2018 - Second special issue on Data for Good

,78 -2851$/ ,&7 'LVFRYHULHV 9RO 'HFHPEHU

individuals and institutions have a rationale goal of any treatment? Soon, values come into play.
incentive for this to be perpetuated, as it works to Opponents of hitting children or of the death
their benefits. In many cases, if citizens-customers penalty, or torture, also argue along moral lines,
were fully informed and aware of these individuals’ irrespective of outcomes and efficiency. And it
motives and actions, they would neither give them would seem opinions trump facts; that culture cards
their money nor their vote. make the best of hands.

The way to achieve this is not to bombard people Yet we argue there is still room and a need for
with facts and tell them on social media they are rational outcome-based arguments in many of these
stupid or evil if they think otherwise. Why not? debates. No study has concluded unambiguously
Because we know this does not work. For such a that the death penalty has deterred crime. What is
system to work, there ought to be something more, known for a hard fact is that innocents will be killed
a “connective tissue” that allows learning to happen, in the process. Torture has been shown to “work” in
information to flow, facts to be heard and matter. few cases, and to lead to bad information in many,
Key ingredients for this seem to include greater while a society that uses torture will probably not
trust, empathy, or “rational compassion”, as “work” for the majority over time. Female genital
discussed below, or shared experiences and mutilation will have no place in a human AI, because
mingling, among and between individuals and it leads to horrific results. Perspectives on social
groups. By design, a human AI would require and justice put forth by Rawls, Sen, and Nussbaum come
foster the kinds of societal characteristics and civic into play [15]–[17] in ways that would take much
processes (especially with respect to social more space and time to give justice to. But
interactions) that would work best for itself. It is not fundamentally, as suggested above, a human AI
entirely clear what those are and will be system is also one where what a good, desirable
(community discussions? elections?), but it seems result is, is discussed and determined on the basis
reasonable to think that the answer are not of facts, to allow for gradual adjustments and
Facebook battles and Twitter storms. improvements.

Another basic challenge is knowing what actually Another key challenge is access to data, particularly
“works” and how, when and where it “works”. The to the kinds of data that would be necessary for a
best economists in the world including Nobel Prize human AI to start functioning. This sensitive data
winners (or to be factually accurate, recipients of holds the most keys to figuring out and advocating
the Sveriges Riksbank Prize in Economic Sciences in convincingly for what works. For instance,
Memory of Alfred Nobel) still disagree about what assessing whether a new transportation system
policies foster inclusive economic growth. There are may result or has resulted in increased economic
broad areas of agreement, but no consensus on the opportunities and lower criminality would be
right sets of policies. This applies to almost all significantly improved by having access to fine
domains of social life because assigning causality or grained mobility data from cell phones. Most “AI
credit is difficult in complex systems where when so data” are collected and stored by private companies
many variables (and values) interplay both as that legally act as data controllers. There have been
inputs and outputs. Most politicians claim that their many examples of and discussions about data
actions should be given credit for rising gross sharing projects and agreements, but to date there
domestic product (GDP) and falling unemployment, are no systematic standards and norms for
or blame the business cycle for opposite outcomes; accessing these “AI data” ethically, and safely at
when in reality assigning credit or blame is hard in scale to power a human AI.
all cases, especially with few data points.
Last, there is the privacy imperative, as a
Another major core challenge is agreeing on what fundamental human right. The vision of a human AI
the “good” end result ought to be. In most AI is not an Orwellian one. It is not about looking into
systems, the end result is a given (as in the “cat vs. individual records or about targeting specific
dog” example) but this is not the case in a human AI. individuals or groups. First, because this would not
Should societies aim for perfect income equality? work: recent societal reactions and legal trends
Should economic policy aim to raise GDP, with all its suggest that while people’s attitudes towards
limitations? Some say, on balance, yes. Others say, privacy are changing, we are not seeing the
on balance, no. Should prolonging life be the end destruction of privacy as a marker and driver of

‹ ,QWHUQDWLRQDO 7HOHFRPPXQLFDWLRQ 8QLRQ