Page 145 - ITU Journal - ICT Discoveries - Volume 1, No. 2, December 2018 - Second special issue on Data for Good

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

3. APPLYING THE PRINCIPLES OF AI TO at large; that widespread corruption is harmful to
HUMAN SYSTEMS: TOWARDS A the majority; that only providing candies and beer
HUMAN AI in corporate cafeterias would not be a good idea. All
of those things tend to yield bad results. We could
We call such a system a human artificial learn that hitting a child for educational purposes
intelligence; a human AI [7]. What would this be and does not “work”, that it is more likely to yield an
do? What would it not do? unstable, unhappy, and violent adult [8]. We could
learn that there is no conclusive evidence that the
The basic principle is that as with current ‘simple’ death penalty works as a deterrent to major crimes
(or narrow) AIs what “works” to “get it right”, [9]. We could learn that human activity over the
policies, programs, behaviors, actions, would get past two centuries has caused many animal species
rewarded and reinforced. Those that “don’t work” to become extinct, while temperatures rose, and
would be penalized and weakened. This too would oceans became more acidic [10].
be enabled by data fed feedback loops. Over time,
you would have human systems (societies, The vision we sketch here is wider than just using
governments, organizations) with a pretty good narrow AIs; it is one where data would fuel those
sense of what “works”, i.e. the sets of policies, human systems by applying the ‘good magic’ of
programs, behaviors, and actions that yield good current narrow AI systems, the credit assignment
results. In addition to providing the core analogy (of function, by identifying, rewarding and reinforcing
learning and reinforcing what works), AIs would be what yields good results. The core principle is
a central part of this system, generating and learning through feedback; the system’s fuel is data.
crunching data and taking over tasks and helping In short, let’s figure out what “works” best, possibly
decision making under general human oversight. for the majority, reward it, and strive to only or
mostly do what contributes to these ends. Over
A key to this is learning and agreeing through time, what helps yield good results will take over
feedback what yields good versus bad results, and what does not, become the most prevalent; ideally
acting accordingly the next time(s) around. Such turned into norms that need less enforcement.
processes already happen. Attempt to have a Human systems would be better off, say safer, fairer,
barbecue in a crowded subway car, and people will more civil, more sustainable, because the opposite
most probably tell you not to. Why? Because it will results do not “work” for most people.
seem like a dangerous thing to do to most riders.
How do they know? Through past experience or Let’s give some simple examples. If a judge (or
(more likely) through “common sense” based on entire justice system) systematically pronounces
past observations and inference. Sometimes we harsher sentences for similar offenses against
learn by insinuations or through intuitions. For people of color, they should be fired (or reformed).
example, talk nonstop loudly at parties and, at some If the way kids are taught impedes their learning
point, you will stop being invited. Most of us will abilities and lifelong prospects, they should be
soon connect the dots. We also have instincts and taught differently. If a government does a lousy job,
reflexes nurtured through thousands of years of steals money, or kills its citizens, it should be
collective learning. We close or cover our eyes if a changed. This may simply feel like common sense or
projectile gets near them, because that yields better liberal democracy at work, but bad policies, bad
results than keeping them wide open to take a actions, and bad results are pervasive even in the
closer look. most “developed” liberal democracies, for many
reasons.
These are, in many ways, core features and
outcomes of evolutionary processes. White rabbits Critically, this is not simply a call for better data in
tend to have higher survival rates in snowy plains the hands of benevolent “Bismarckian”
than brown ones. After a while, there are only white policymakers who would (finally!) be able to make
rabbits left running around in snowy plains. It is good decisions. We do not believe that some of the
also the gist of culture; societies learn and teach greatest threats and challenges of our time, and
what “works” for them, and turn this learning into those to come, are primarily due to poor
codes and norms. Most societies have learned that information available to the ruling classes. One
not providing basic education to their children is reason is that a fair share of politicians and people
not great, neither for the children nor for the society in positions of power are either uninterested in the

‹ ,QWHUQDWLRQDO 7HOHFRPPXQLFDWLRQ 8QLRQ