accident for the near future in the same extent it has decreased over the past
30 years owing to safety improvements. This huge effect of Fukushima Daiichi in
revising the global estimation of a core meltdown can be interpreted as evidence
that besides the design, the location and the operating of reactors the probability of
an accident also depends on institutional factors like the strength and ability of
nuclear safety authorities, a factor which is not taken into account in probabilistic
assessments. In fact, like in Japan there are a lot of countries wherein nuclear safety
authorities are captured by operators and fail to enforce safety standards.
As a conclusion, uncertainties prevail. There is no overarching probability of
nuclear accident to use to make a rational decision for society to invest in or to
phase out nuclear power generation, to determine the right level of nuclear safety
expenditures, not to say to identify the economically optimal level of nuclear safety.
Unlike transport crashes no means can be inferred from observed frequencies.
Moreover, the probability of a nuclear accident differs according to the design and
the location of reactors but also according to institutional characteristics (independent regulator, liability rules, experience of operators, etc.). We do not know the
probability distribution of nuclear accidents, even for a given reactor design and
location. Last but not least, one has always to keep in mind that probabilistic
analysis requires knowing all the states of the world. A probability cannot be
assigned to an unknown event, or to put it another ways to black swans and
unknown unknowns.
2.3 Perception of Probabilities
Utility function and human behavior
It is well known that many people are risk-averse: they would rather, for instance, a
certain gain of 100 to an expected gain of 110. Since Bernoulli [5], this psychological trait is represented by a concave utility function. The Swiss mathematician
opened the way for progress towards decision theory
6 through a back-and-forth
between economic modeling and psychological experimentation. The latter would,
for instance, pick up an anomaly—in a particular instance people’s behavior did not
conform to what theory predicted—and the former would repair it, altering the
mathematical properties of the utility function or the weighting of probabilities. The
works by Allais and Ellsberg were two key moments in this achievement.
Following an experiment showing that people with good knowledge of the theory
of probability were violating an axiom of expected utility theory, Allais [7] proposed to weight probabilities depending on their value, with high coefficient for low
probabilities, and vice versa. Putting it another way, preferences assigned to
probabilities are not linear. This is more than just a technical response. It makes
6
For a comprehensive panorama on decision theory under uncertainties see [6].
84
R. Bizet and F. Lévêque
30 years owing to safety improvements. This huge effect of Fukushima Daiichi in
revising the global estimation of a core meltdown can be interpreted as evidence
that besides the design, the location and the operating of reactors the probability of
an accident also depends on institutional factors like the strength and ability of
nuclear safety authorities, a factor which is not taken into account in probabilistic
assessments. In fact, like in Japan there are a lot of countries wherein nuclear safety
authorities are captured by operators and fail to enforce safety standards.
As a conclusion, uncertainties prevail. There is no overarching probability of
nuclear accident to use to make a rational decision for society to invest in or to
phase out nuclear power generation, to determine the right level of nuclear safety
expenditures, not to say to identify the economically optimal level of nuclear safety.
Unlike transport crashes no means can be inferred from observed frequencies.
Moreover, the probability of a nuclear accident differs according to the design and
the location of reactors but also according to institutional characteristics (independent regulator, liability rules, experience of operators, etc.). We do not know the
probability distribution of nuclear accidents, even for a given reactor design and
location. Last but not least, one has always to keep in mind that probabilistic
analysis requires knowing all the states of the world. A probability cannot be
assigned to an unknown event, or to put it another ways to black swans and
unknown unknowns.
2.3 Perception of Probabilities
Utility function and human behavior
It is well known that many people are risk-averse: they would rather, for instance, a
certain gain of 100 to an expected gain of 110. Since Bernoulli [5], this psychological trait is represented by a concave utility function. The Swiss mathematician
opened the way for progress towards decision theory
6 through a back-and-forth
between economic modeling and psychological experimentation. The latter would,
for instance, pick up an anomaly—in a particular instance people’s behavior did not
conform to what theory predicted—and the former would repair it, altering the
mathematical properties of the utility function or the weighting of probabilities. The
works by Allais and Ellsberg were two key moments in this achievement.
Following an experiment showing that people with good knowledge of the theory
of probability were violating an axiom of expected utility theory, Allais [7] proposed to weight probabilities depending on their value, with high coefficient for low
probabilities, and vice versa. Putting it another way, preferences assigned to
probabilities are not linear. This is more than just a technical response. It makes
6
For a comprehensive panorama on decision theory under uncertainties see [6].
84
R. Bizet and F. Lévêque