This dissertation provided a descriptive analysis of the phenomenon of discrimination. We first dissected discrimination by means of decision theory. In so doing, we started with a broad definition of discrimination and then identified more and more distinctive manifestations of it. First of all, we separated social from non-social discrimination. Then, within the concept of social discrimination, we further differentiated statistical from taste-based discrimination. Finally, we investigated to what behaviour the combination of different types of discrimination can lead. During the whole dissection, the decision-maker’s state of knowledge was an essential aspect. Here, we distinguished two states: decision-making under certainty and decision-making under uncertainty. The main difference between these two is given by the fact that while certainty is objectively given, in case of uncertainty probabilities are subjectively formed. Due to that statistical discrimination is only possible if a decision situation underlies uncertainty. Here, a statistical discriminator uses the group memberships of the people involved in a decision situation as proxies in order to assess scenarios’ subjective probabilities. In contrast, taste-based discrimination is possible in both kinds of decision-making and involves that the decision-maker has a taste for certain people/groups. Moreover, so as to have such tastes, she needs agent-relative preferences. In turn, there is no taste-based discrimination if the decision-maker has agent-neutral preferences.

Subsequently, we investigated taste-based discrimination. One of the most intruding question regarding taste-based discrimination is as follows: Where do we draw the line between those we treat prosocially and those we treat neutrally or even antisocially? Social identity theory provided an answer to this question. We have a taste for our ingroup and/or a distaste for our outgroups. Yet, the precise definition of the ingroup and outgroup is changeable and depends on the situation. Here, self-categorisation theory helped us to determine which of the many possible group constellations becomes salient. Next, we have investigated whether ingroup love and/or outgroup derogation gives rise to ingroup favouritism and found that the former is stronger than the latter. Additionally, we demonstrated that not all tastes have to stem from an ingroup-outgroup context, yet when looking more closely, such tastes often still appear to be intertwined with social identity. Then, we discussed the question whether taste-based discrimination is actually always statistical discrimination with ingroup favouring beliefs. We found that such beliefs certainly are of importance in regard to ingroup favouritism. Nevertheless, they seem not to be able to explain all ingroup favouritism that we observe in experiments. Thus, taste-based discrimination appears to actually exist, which requires that people have agent-relative social preferences. There are multiple explanatory approaches for such preferences. The most promising one provide kin altruism, reciprocal altruism, indirect reciprocity, and costly signalling in combination with parochial altruism, cultural group selection, and gene-culture coevolution.

In the next part of the dissertation we focused on how we get beliefs and what has to be fulfilled in order that they are rational, leading to rational statistical discrimination. Subjective expected utility theory has few requirements in order that a belief is labelled as rational. It only has to be consistent with the other beliefs and updated by use of Bayes’ law. As a consequence, it does not provide a theory of prior belief generation. This led us to Bayesianism and how people deviate from it. First, we analysed whether there are inherent prior beliefs. Such beliefs would not have been learned individually but collectively over the course of evolution. Here, we found that people appear to belief in the superiority of familiar alternatives. The existence of such a belief can be explained via error management theory. Additionally, there seem to be prior beliefs about the ingroup and outgroup as well. Next, we looked at how people update their beliefs and thereby whether they stick to Bayes’ law. We found four apparent deviations: (1) People are not good at handling probabilities but rather deduce the probability of an event from its availability. (2) People incorrectly remember their prior probabilities after having them updated. (3) People gather and process confirming evidence differently than disconfirming evidence and are less critical in regard to their own beliefs than those of others. (4) Social identity can affect our belief formation process in such a way that it leads to beliefs that tend to flatter the ingroup and decry the outgroup. Finally, we examined the role and characteristics of a decision-maker’s learning environment. We showed that our Western world is shaped by historical (and partly still ongoing) oppressions of certain groups. Today, a decision-maker’s learning environment still inheres these circumstances to some degree, which as a consequence find expression in her beliefs. So, the beliefs of an agent-neutral person can reflect agent-relative convictions if the environment she learns in was co-shaped by agent-relative people.

The final part of the dissertation reassembled discrimination. We first put the major aspects of the previous chapters in a descriptive model of discrimination. On one hand, we distinguished whether the formation of our beliefs is irrelevant due to correctly recognised certainty, adheres to objective Bayesianism (or equivalent), or adheres to subjective Bayesianism or all other forms of belief formation. On the other hand, we separated decision-makers with agent-neutral preferences from those with agent-relative preferences. The combination of these two dimensions of distinction led to six interactions. From this descriptive model of discrimination, we then derived five aspects that a normative theory of discrimination should consider. They involve the approach to discrimination, the omnipresence of uncertainty and as a consequence the virtually inevitable usage of group specific beliefs, our belief formation process, the manifold manifestations of agent-relative preferences, and the importance of someone’s learning environment.

The goal of this dissertation was to provide a nuanced perspective on discrimination that is free from judgments of legitimacy and illegitimacy. This is what we have done. So, what is the scientific novelty value of this dissertation? For the first time, decision theory was neatly employed on the phenomenon of discrimination. In this way, we derived the two forms of social discrimination that have already been mentioned in the literature, namely taste-based and statistical discrimination. Ingroup favouritism was then integrated into and thereby explained within the decision theoretical framework. This is the first time this has been done in such a comprehensive way. Next, this dissertation provides an in-depth analysis of human biases that directly or indirectly relate to groups and reveals how they interfere with objective and subjective Bayesianism. In so doing, we bundled various biases that seem to be manifestations of the same mechanism and examined their universality as well as ultimate explanation. This has not been done before in such a thorough way. Finally, this dissertation provides a new descriptive model of discrimination that builds on the previous findings and lists five implications for a normative theory of discrimination. Considering these implications, it can be inferred that decision theory itself seems to be insufficient so as to define legitimate and illegitimate discrimination.

In a next step, these descriptive insights into discrimination and their implications can be applied on the normative discourse on discrimination. At this, the decision-theoretical language we introduced so as to define different forms of discrimination can particularly help to clarify what kind of discrimination one actually talks about and eventually condemns. The mathematical language used in this dissertation provides a precise mutual definitional basis which differing normative theories of discrimination can refer to. Hereby, hardened normative fronts regarding discrimination might hopefully loosen up a bit because misunderstandings about the property of discrimination should become less likely.

This dissertation’s descriptive analysis of discrimination has limitations. First of all, as we have discussed several times before, we face epistemological problems when we want to detect the accurate type(s) of discrimination from empirical observations. Although we can to some degree deduce it/them if there exists a basis of comparison which ideally is as large as possible, there can never be complete certainty (cf. Kant, 2011[1786]). This circumstance ultimately underlies all empirical studies that we discussed. Second, the subjective expected utility theory that we used assumes that while we do not know the probabilities of scenarios, we know all their characteristics. In other words, there are no unknown unknowns. Yet, situations also exist where we neither know the probabilities of scenarios nor the characteristics of all possible scenarios. Our dissection of discrimination has omitted such conditions. Third, our distinction of social and non-social discrimination is more complicated in real life because agent-relative preferences can also influence our preferences for things. Fourth, many fields of research that we introduced still have open questions. Most strikingly, the ultimate explanations for why we have inherent prior beliefs and do not update our beliefs according to Bayes’ law need more evidence. Similarly, the puzzle of the evolution of agent-relative social preferences is also not yet conclusively solved. Fifth, our analysis of discrimination mainly considered psychological as well as evolutionary explanations for different kinds of discrimination and only briefly discussed sociological influences and implications. Finally, although the very goal of this dissertation was to provide a descriptive analysis of discrimination, some normative judgments were inevitable. For example, this involves how we defined discrimination, which theories we used so as to explain discrimination, or which dimensions we chose for the descriptive model of discrimination as well as how we defined these dimensions.

At the very end of this dissertation, let’s go back to the two examples we used in the introduction: the sly vixens campaign and the applicant screening algorithm. What can we tell about them after our dissection of discrimination? We start with the example of the sly vixens campaign. Here, the national railway company of Switzerland (SBB) exclusively looked for women who, while wearing fox ears and a fox tail, would make morning commuters aware of extra trains. Moreover, the SBB advertised this job on an online platform of two universities and thereby probably excluded non-academics. As we said in the introduction, this is a case of discrimination because some groups are systematically treated differently than others. Yet, what type of discrimination is it? Of course, we cannot know that for sure but it seems that the SBB were mainly (biased) statistical and not taste-based discriminators in this case. The SBB officially replied that the reason why they particularly addressed women was that the sly foxes and vixens have to wear a hairband (on which the fox ears are mounted) and they thought that women can wear these better (Iseli, 2017; Heininger & Hartmann, 2017). So, the SBB seem to have based their decision on a statistic about which gender sits a hairband better on. And although they do not state that explicitly, they might also have applied a statistic which says that women are more likely to do and/or more accepted when they do such assistant jobs than men.Footnote 1 Particularly the beliefs of the last sentence would in all likelihood have had to stem from an environment that was co-shaped by taste-based discriminators. Finally, the SBB might have particularly addressed students because they are statistically more likely to do little side jobs than the average citizen or other groups.

The case of the applicant screening algorithm is a bit more complicated. First of all, we exclude the possibility that the algorithm is a taste-based discriminator. Now, the goal of the algorithm is to find the applicant that suits the firm best. Thereby, it is forbidden to use the category “skin colour”. In so doing, it finds a negative correlation between how far away someone lives from her workplace and how long that person stays at the firm. This leads to statistical discrimination: Those who live close to the workplace are ceteris paribus more likely to be employed than those who live further away from the workplace. Consequently, the categorisation of individuals into groups is defined by the distance between their home and workplace. Skin colour in and of itself is irrelevant for this categorisation (as prescribed). However, there is a correlation between skin colour and the distance to workplace. So, does the algorithm ultimately still statistically discriminate between people of different skin colour? Following this dissertation’s definition of discrimination, this is not the case because the algorithm is blind for skin colour. It does not know this category which is why it can also not use it for any kind of discrimination. In contrast, the circumstance that black people tend to live further away from their potential workplace than others is in all likelihood due to taste-based discriminators who co-shaped the momentary environment.

These two examples reveal how crucial the learning environment of statistical discriminators is and how (past) taste-based discriminators can influence the beliefs of agent-neutral decision-makers. The current rise of algorithms will further demonstrate this. Meanwhile, nationalism, antisemitism, sexism, homophobia, xenophobia, anti-westernism, anti-islamism, or simply taste-based discrimination still exists and partly even increases. So, discrimination remains a hot topic. When discussing it, we should not forget that the actual ability to discriminate is a precious facility that we need in everyday life. Thus, it appears not to be expedient to generally condemn discrimination. But where to draw the line between legitimate and illegitimate discrimination is a difficult question. This descriptive analysis of discrimination can provide the language but not the answer for it.