14.2 Can We Trust It?
293
course, a third factor, namely outside heat, which causes both, increased ice cream
consumption and forest fires.
Finally, there may be no causal relationship between A and B at all. The bigger
a data set, the more patterns will be found just by coincidence, and this could be
wrongly interpreted as meaningful or, as some people would say, as a signal rather
than noise.
31 In fact, spurious patterns and correlations are quite frequent.
32
Nevertheless, it is, of course, possible to run a society based on correlations. The
application of predictive policing may be seen as example. However, the question
is, whether this would really serve society well. I don’t think so. Correlations are
frequent, while causal relationships are not. Therefore, using correlations as basis of
certain kinds of actions unnecessarily restrains our freedom (effectively introducing
new laws through code).
14.2.3 Trustable AI
There has been the dream that Big Data is the “new oil” and Artificial Intelligence
something like a “digital motor” running on it. So, if it is difficult for humans to
make sense of Big Data, AI might be able to handle it “better than us”. Would AI be
able to automate Big Data analytics? The answer is, partly.
In recent years, it was discovered that AI systems would often discriminate against
women, non-white people, or minorities. This is because these systems are typically
trained with data of the past. That is problematic, since learning from the past may
stabilize old societal paradigms we should actually better replace by something else,
given that today’s world is not sustainable.
Lack of explanation is another important issue. For example, you may get into
the situation that your application for a loan or life insurance is turned down, but
nobody can explain you why. The salespeople would just be able to tell you that
their AI system has recommended them to do so. The reason may be that two of
your neighbors had difficulties paying back their loans. But this is again messing
up correlations and causal relationships. Why should you suffer from this? Hence,
experts have recently pushed for explainable results under labels such as “trustable
AI”. So far, however, one may say that we are still living in a “blackbox society”.
33
31 Silver [7].
32 Vigen [8].
33 Pasquale [9], The Dark Secret at the Heart of AI, MIT Technology Review (April 11, 2017)
https://www.technologyreview.com/s/604087/the-dark-secret-at-the-heart-of-ai/.
293
course, a third factor, namely outside heat, which causes both, increased ice cream
consumption and forest fires.
Finally, there may be no causal relationship between A and B at all. The bigger
a data set, the more patterns will be found just by coincidence, and this could be
wrongly interpreted as meaningful or, as some people would say, as a signal rather
than noise.
31 In fact, spurious patterns and correlations are quite frequent.
32
Nevertheless, it is, of course, possible to run a society based on correlations. The
application of predictive policing may be seen as example. However, the question
is, whether this would really serve society well. I don’t think so. Correlations are
frequent, while causal relationships are not. Therefore, using correlations as basis of
certain kinds of actions unnecessarily restrains our freedom (effectively introducing
new laws through code).
14.2.3 Trustable AI
There has been the dream that Big Data is the “new oil” and Artificial Intelligence
something like a “digital motor” running on it. So, if it is difficult for humans to
make sense of Big Data, AI might be able to handle it “better than us”. Would AI be
able to automate Big Data analytics? The answer is, partly.
In recent years, it was discovered that AI systems would often discriminate against
women, non-white people, or minorities. This is because these systems are typically
trained with data of the past. That is problematic, since learning from the past may
stabilize old societal paradigms we should actually better replace by something else,
given that today’s world is not sustainable.
Lack of explanation is another important issue. For example, you may get into
the situation that your application for a loan or life insurance is turned down, but
nobody can explain you why. The salespeople would just be able to tell you that
their AI system has recommended them to do so. The reason may be that two of
your neighbors had difficulties paying back their loans. But this is again messing
up correlations and causal relationships. Why should you suffer from this? Hence,
experts have recently pushed for explainable results under labels such as “trustable
AI”. So far, however, one may say that we are still living in a “blackbox society”.
33
31 Silver [7].
32 Vigen [8].
33 Pasquale [9], The Dark Secret at the Heart of AI, MIT Technology Review (April 11, 2017)
https://www.technologyreview.com/s/604087/the-dark-secret-at-the-heart-of-ai/.