72
4 Google as God?
4.7 Errors or First and Second Kind: Doing Good Isn’t
Easy
First of all, what’s the goal function one should optimize? Money? Happiness?
Health? Security? Innovation? Peace? Or sustainability? Surprisingly, there is no
scientific method that can give a definite answer, but if we chose the wrong goal,
it could end in disaster.
35 Therefore, choosing and implementing a single solution
everywhere is highly dangerous.
36 But even if the goal function was clear, distinguishing “good” from “bad” solutions is often difficult, for statistical reasons. For
example, good and bad risks or behaviors can hardly ever be perfectly separated,
as all measurements are noisy. Consequently, there is the problem of false positive
classifications (false alarms, so-called type I errors) and of false negatives (type II
errors, where the alarm is not triggered when it should be).
Imagine a population of 500 million people, in which there are 500 terrorists. Let’s
assume that we can identify terrorists with an extremely impressive 99% accuracy.
In this scenario, there are 1% false negatives (type II error), which means that 5
terrorists are not detected, while 495 will be discovered. Newspapers reported that
about 50 terrorist acts in Western countries were prevented over the past 12 years
or so, while the attack on the Boston marathon and on the Charlie Hebdo editorial
office in Paris could not be prevented, even though the perpetrators were listed in
databases of terrorist suspects (in other words, they turned out to be false negatives).
How many false positives (false alarms) could we expect? If the type I error is
just 1 in 10,000, there will be 50,000 innocent suspects, while if it is 1 in 1000 there
will be 500,000 innocent suspects. If it is 1% (which is entirely plausible), there will
be 5 million innocent suspects! Indeed, it has been reported that there are between 1
and 8 million people on the various lists of suspects in the USA.
37 If these figures are
correct, this means that for every genuine terrorist, around 10,000 innocent citizens
will be wrongly categorized as potential terrorists, perhaps even more. Since the 9/11
attacks, about 40,000 suspects have had to undergo special questioning and screening
procedures at international airports. In 99% of these cases, it was later concluded
35 For a number of years, using asbestos in buildings has been the preferred choice for fire protection
and other means. However, in the end it turned out to cause lung cancer, and its widespread use was
very costly. The use of Contergan, which caused the abnormal development of childrens’ bodies, is
another well-known disaster of this kind. The reasons for the widespread increase of allergies and
obesity are not yet known for sure. However, the underlying problem is again that a huge number
of people has been simultaneously exposed to new substances. The same applies to the massive
deaths of bees.
36 That’s probably why social, ecological and economic systems happen to “explore” a large diversity
of goal functions, where some competitive or selection mechanism helps good solutions to spread.
This speaks for pluralism. In fact, the reduction of diversity is one of the greatest dangers. As we
will see later on, socio-economic diversity is as important as biodiversity. Many species (including
many plants) multiply themselves in a way that increases diversity. If this is not done, handicaps or
low IQ might result. Therefore, laws even forbid that brothers and sisters marry.
37 It recently got known that the police of Bavaria (a federal state of Germany with 12.7 million in
habitants) stores about 1 million people in a list of suspects, see http://www.sueddeutsche.de/bay
ern/datenbank-der-polizei-verdaechtig-fuers-leben-1.2549799.
4 Google as God?
4.7 Errors or First and Second Kind: Doing Good Isn’t
Easy
First of all, what’s the goal function one should optimize? Money? Happiness?
Health? Security? Innovation? Peace? Or sustainability? Surprisingly, there is no
scientific method that can give a definite answer, but if we chose the wrong goal,
it could end in disaster.
35 Therefore, choosing and implementing a single solution
everywhere is highly dangerous.
36 But even if the goal function was clear, distinguishing “good” from “bad” solutions is often difficult, for statistical reasons. For
example, good and bad risks or behaviors can hardly ever be perfectly separated,
as all measurements are noisy. Consequently, there is the problem of false positive
classifications (false alarms, so-called type I errors) and of false negatives (type II
errors, where the alarm is not triggered when it should be).
Imagine a population of 500 million people, in which there are 500 terrorists. Let’s
assume that we can identify terrorists with an extremely impressive 99% accuracy.
In this scenario, there are 1% false negatives (type II error), which means that 5
terrorists are not detected, while 495 will be discovered. Newspapers reported that
about 50 terrorist acts in Western countries were prevented over the past 12 years
or so, while the attack on the Boston marathon and on the Charlie Hebdo editorial
office in Paris could not be prevented, even though the perpetrators were listed in
databases of terrorist suspects (in other words, they turned out to be false negatives).
How many false positives (false alarms) could we expect? If the type I error is
just 1 in 10,000, there will be 50,000 innocent suspects, while if it is 1 in 1000 there
will be 500,000 innocent suspects. If it is 1% (which is entirely plausible), there will
be 5 million innocent suspects! Indeed, it has been reported that there are between 1
and 8 million people on the various lists of suspects in the USA.
37 If these figures are
correct, this means that for every genuine terrorist, around 10,000 innocent citizens
will be wrongly categorized as potential terrorists, perhaps even more. Since the 9/11
attacks, about 40,000 suspects have had to undergo special questioning and screening
procedures at international airports. In 99% of these cases, it was later concluded
35 For a number of years, using asbestos in buildings has been the preferred choice for fire protection
and other means. However, in the end it turned out to cause lung cancer, and its widespread use was
very costly. The use of Contergan, which caused the abnormal development of childrens’ bodies, is
another well-known disaster of this kind. The reasons for the widespread increase of allergies and
obesity are not yet known for sure. However, the underlying problem is again that a huge number
of people has been simultaneously exposed to new substances. The same applies to the massive
deaths of bees.
36 That’s probably why social, ecological and economic systems happen to “explore” a large diversity
of goal functions, where some competitive or selection mechanism helps good solutions to spread.
This speaks for pluralism. In fact, the reduction of diversity is one of the greatest dangers. As we
will see later on, socio-economic diversity is as important as biodiversity. Many species (including
many plants) multiply themselves in a way that increases diversity. If this is not done, handicaps or
low IQ might result. Therefore, laws even forbid that brothers and sisters marry.
37 It recently got known that the police of Bavaria (a federal state of Germany with 12.7 million in
habitants) stores about 1 million people in a list of suspects, see http://www.sueddeutsche.de/bay
ern/datenbank-der-polizei-verdaechtig-fuers-leben-1.2549799.