292
14 Summary: What’s Wrong with AI?
14.2 Can We Trust It?
14.2.1 Big Data Analytics
In an article of the “Wired” magazine 2008, Chris Anderson (*1961) claimed that
we would soon see the end of theory, and the data deluge would make the scientific
method obsolete.
28 If one just had enough data, certain people started to believe, data
quantity could be turned into data quality, and, therefore, Big Data would be able
to reveal the truth by itself. In the past few years, however, this paradigm has been
seriously questioned. As data volume increases much faster than processing power,
there is the phenomenon of “dark data”, which will never be processed and, hence, it
will take scientists to decide which data should be processed and how.
29 In fact, it is
not trivial at all to distill raw data into useful information, knowledge and wisdom.
In the following, I will describe some of the problems related to the question “how
to connect the dots”.
It is frequently assumed that more data and more model parameters are better to
get an accurate picture of the world. However, it often happens that people “can’t see
the forest for the trees”. “Overfitting”, where one happens to fit models to random
fluctuations or otherwise meaningless data, can easily happen. “Sensitivity”, where
outcomes of data analyses change significantly, when some data points are added or
subtracted, or another algorithm or computer hardware is used, is another problem.
A third problem are errors of first and second kind, i.e. “false positives” (“false
alarms”) and cases, where alarms should go off, but fail to do so. A typical example
is “predictive policing”, where false positives are overwhelming (often above 99
percent), and dozens of people are needed to clean the suspect lists.
30 And these are
by far not all the problems …
14.2.2 Correlation Versus Causality
In Big Data, it is easy to find patterns and correlations. But what do they actually
mean? Say, one finds a correlation between two variables A and B. Then, does A
cause B or B cause A? Or is there a third factor C, which causes A and B? For
example, consider the correlation between the number of ice-cream-eating children
and the number of forest fires. Forbidding children to eat ice cream will obviously
not reduce the number of forest fires at all—despite the strong correlation. It is, of
28 The End of Theory: The Data Deluge Makes the Scientific Method Obsolete, Wired (June 24,
2008) https://www.wired.com/2008/06/pb-theory/.
29 Das Digital-Manifest: Digitale Demokratie statt Datendiktatur, Spektrum der Wissenschaft
(November 12, 2015) https://www.spektrum.de/thema/das-digital-manifest/1375924.
30 BKA: Überwachung von Flugpassagieren liefert Fehler über Fehler, Süddeutsche Zeitung (April
24, 2019) https://www.sueddeutsche.de/digital/fluggastdaten-bka-falschtreffer-1.4419760.
14 Summary: What’s Wrong with AI?
14.2 Can We Trust It?
14.2.1 Big Data Analytics
In an article of the “Wired” magazine 2008, Chris Anderson (*1961) claimed that
we would soon see the end of theory, and the data deluge would make the scientific
method obsolete.
28 If one just had enough data, certain people started to believe, data
quantity could be turned into data quality, and, therefore, Big Data would be able
to reveal the truth by itself. In the past few years, however, this paradigm has been
seriously questioned. As data volume increases much faster than processing power,
there is the phenomenon of “dark data”, which will never be processed and, hence, it
will take scientists to decide which data should be processed and how.
29 In fact, it is
not trivial at all to distill raw data into useful information, knowledge and wisdom.
In the following, I will describe some of the problems related to the question “how
to connect the dots”.
It is frequently assumed that more data and more model parameters are better to
get an accurate picture of the world. However, it often happens that people “can’t see
the forest for the trees”. “Overfitting”, where one happens to fit models to random
fluctuations or otherwise meaningless data, can easily happen. “Sensitivity”, where
outcomes of data analyses change significantly, when some data points are added or
subtracted, or another algorithm or computer hardware is used, is another problem.
A third problem are errors of first and second kind, i.e. “false positives” (“false
alarms”) and cases, where alarms should go off, but fail to do so. A typical example
is “predictive policing”, where false positives are overwhelming (often above 99
percent), and dozens of people are needed to clean the suspect lists.
30 And these are
by far not all the problems …
14.2.2 Correlation Versus Causality
In Big Data, it is easy to find patterns and correlations. But what do they actually
mean? Say, one finds a correlation between two variables A and B. Then, does A
cause B or B cause A? Or is there a third factor C, which causes A and B? For
example, consider the correlation between the number of ice-cream-eating children
and the number of forest fires. Forbidding children to eat ice cream will obviously
not reduce the number of forest fires at all—despite the strong correlation. It is, of
28 The End of Theory: The Data Deluge Makes the Scientific Method Obsolete, Wired (June 24,
2008) https://www.wired.com/2008/06/pb-theory/.
29 Das Digital-Manifest: Digitale Demokratie statt Datendiktatur, Spektrum der Wissenschaft
(November 12, 2015) https://www.spektrum.de/thema/das-digital-manifest/1375924.
30 BKA: Überwachung von Flugpassagieren liefert Fehler über Fehler, Süddeutsche Zeitung (April
24, 2019) https://www.sueddeutsche.de/digital/fluggastdaten-bka-falschtreffer-1.4419760.