Discrimination is often portrayed as a defect of the human mind that needs to be corrected. Egalitarian social theories assert that social equality should prevail, and to that end, we should fight any form of discrimination vigorously, as it is invariably irrational, caused by baseless hostility and prejudice.
We all discriminate in our daily lives, for instance by favoring our relatives when helping others (a phenomenon explained by evolutionary game theory). It doesn’t mean that it is always fair and that opposing some forms of discrimination is a bad policy. But there are many instances in which it helps and can even save lives. It is in fact an essential evolutionary mechanism that helps us survive and evolve. Fighting against any form of discrimination on principle is irrational and dangerous. By understanding what it is, gathering data, testing its validity on a case-by-case basis and outlining risks and root causes, false hypotheses can be exposed more efficiently.
As this is a highly politicized subject, it’s good to put aside passionate convictions and think of humans as self-learning organisms that have optimized, over countless generations, algorithms for maximizing their chance of survival. Through this lens, discrimination is the result of solving a pattern-recognition problem called classification, as part of a natural reinforcement learning process. In psychology, this is known as “discrimination learning.”
A caveman who escapes a bear attack interprets the encounter as negative reward and immediately looks for patterns to avoid similar events in the future. He starts by identifying the most discriminative characteristics of the attacker and concludes that all similar animals, in this case bears, are potentially dangerous to him. It doesn’t mean they are bad in an absolute sense. This is a rational, acceptable behavior, despite the possibility that some bears might be inoffensive. Because of the high risk involved, the best strategy is to generalize early. If another bear were to “sue” the caveman for discrimination, the human could defend himself by invoking what is known in legal terms as “objective justification”: achieving a legitimate aim. One caveman could be entirely wrong, but as more of them reach the same conclusion, the probability that they are all wrong in avoiding bears is decreasing. This is how we’ve learned to avoid predators and other dangerous situations, to discern what is edible, and so on.
The same approach is used by many computer programs. Email spam filters receive as negative feedback spam reports; they analyze those emails and learn what are the most discriminative characteristics, called features. In the context of emails, the words are important features. This is how spam filters use machine learning as they learn to classify emails containing “win,” “money,” and “casino” as being spam, which in most (but not all) cases is accurate.
To discriminate means to make a distinction based on a group property. People say that “the Dutch are straightforward” and “old folks are more risk averse.” Usually, nobody objects to such statements, and some even take pride in the positive traits linked to their group. These are properties of the average; it doesn’t mean that everyone in the group shares them. Saying that a young driver is more likely to cause an accident is a form of discrimination, but it is not an incorrect statement, and insurance companies are using age as a feature for pricing. In economics, this is known as statistical discrimination: estimating an individual’s performance based on his group identity and the average behavior of that group.
While some features are allowed for classification, others are banned by contemporary social norms. Analyzing behavior based on ethnic or religious features is immediately labeled as racist, bigoted, or xenophobic and sometimes even forbidden by law. But could it be that no such correlations exist? Is it plausible that hundreds or thousands of years of shared history never form a group identity, with more prevalent behaviors and goals, that can be incompatible or conflicting with the goals and behaviors of others? Is it unthinkable that these incompatibilities and conflicts can’t always be solved through imposed integration?
A well known ethnic minority in Romania are the Roma people, also known as gypsies, which represent around 3% of the population. The name’s similarity with Romanians is a confusing coincidence: the Roma are a distinct ethnic group that migrated from India, arriving in Europe 900 years ago. For Romanians, the distinct lifestyle and high incidence of antisocial behavior in Roma communities is a well-known, observable fact. The EU used to criticize the Romanian state for not “fighting discrimination” enough. When Romania joined the EU, and many of its citizens travelled freely to other European countries, suddenly the same officials started to complain about improvised camps and high crime rates of “minorities of Romanian or Bulgarian origin,” while trying hard to avoid ethnic distinctions. But is such censorship in any way useful, and why then is mentioning the nationality allowed? Does it help to ban a few hand-picked features, such as ethnicity and religion, and claim that it is irrational to use them for classification? Wouldn’t it be better for everyone to acknowledge reality and design realistic policies, with cultural and behavioral particularities in mind, while raising awareness about a history of slavery and persecution? This policy of censoring observations can turn against the very groups it tries to protect.
There are many problems with classification. First, it can be based on false data or data presented in a misleading way. Even if the data is correct, classification models are rarely perfect. They can be oversimplified, a problem called underfitting. Even if a behavior is observed more often in one group than in others, there can still be many individuals, possibly the majority, in that group who do not exhibit it (false positives). If it is a negative trait, these individuals may be treated unfairly, and punitive measures can perpetuate the very behaviors they try to correct. If it is a positive one, classification can lead to unfounded favoritism of false positives. There might be other much better models and features, and the perception of behavior as being negative, neutral, or positive can be subjective, and sometimes it can be simply wrong.
If a feature is found to predict behavior perceived as negative in some context, the next question is how to act upon this information. In the extreme, history shows this can lead to persecution and even mass murder. That is why today, many choose to turn a blind eye, simply ignoring correlations and data. Picking the best course of action is an optimization problem that can’t be addressed without moral and ethical judgements, evaluating risks, benefits, short-term and long-term effects. But it is important to understand these tradeoffs and the rationality and caveats of discrimination, to stay objective and not make quick, binary, dogmatic judgements guided only by ideological convictions rather than logic and facts. Censorship never works indefinitely, and if mainstream media and leaders don’t acknowledge, explain, and address reality, then tabloids and fringe politicians hijack the public’s attention and promote hatred based on flawed, oversimplified models of the world for petty financial or political gains. This polarization weakens the social fabric, makes countries vulnerable to both external and internal risks, and can lead to dark places.