Tübingen – The sense of righteousness is deeply embedded in a human from an early age. It is an essential part of the moral understanding of a person, of the standards and the values an individual lives by. The right as well as the trust in being treated fair in any aspect of life forms the basis of a peaceful coexistence in a society. But while our society undergoes a massive change in times of digitalization – affecting all areas of life – the necessity arises for self-learning machines to be fair when supporting humans in making decisions. Only when Artificial Intelligence interprets and applies fairness in the same way as people do, will they accept a computer´s decision-making process. Self-learning algorithms, which may act upon racist or sexist features, won´t find acceptance in a society. Hence transparency is key in a world where Artificial Intelligence becomes more relevant in every day decision making.
“Back in the day when intelligent, self-learning machines were applied only in industrial settings, no one thought about whether a computer was fair. No one made it a point to set moral or ethical standards in Machine Learning”, Niki Kilbertus explains. He is a researcher at the Max Planck Institute for Intelligent Systems in Tübingen. “But now that the same algorithms are applied in a social context, one has to ask: do these self-learning machines act just or do they discriminate, regardless of whether it is intentional or not.” So instead of forming unbiased decisions, sometimes Artificial Intelligence does the exact opposite: it discriminates based upon the data, which it is trained on.
“Avoiding Discrimination through Causal Reasoning“ is the title of a paper Kilbertus published together with Mateo Rojas-Carulla, Giambattista Parascandolo, Moritz Hardt, Dominik Janzing und Bernhard Schölkopf, Director of the Empirical Inference Department at the Max Planck Institute for Intelligent Systems. The publication aims to raise awareness: when Machine Learning is applied, the people programming the algorithms should look twice at the variables on which the calculation is based upon. Could they potentially lead to discriminatory outcomes? Kilbertus and his colleagues call upon the programmers to first scrutinize data, asking the question why.
Discrimination can be based on attributes such as gender, race, religion or sexual orientation. But even through something like a zip code potential debtor has to provide when applying for a loan, a self-learning algorithm could discriminate against a person or a demographic group. A bank could for instance withhold loans to people with a certain postal code, just because the self-learning algorithm calculated that the default rate is higher in this area than elsewhere. If we would let the algorithm decide, the debtor wouldn´t receive the loan – even though she or he is in fact creditworthy”, Kilbertus explains.
Kilbertus and his colleagues look for a fair solution in a case like that. They search for the causal reason why debtors default on their risk in this particular area. “It could be that the default rate is just higher, which the bank then uses as a criterion to give no one with this postal code a loan – just to be on the safe side. If at the same time a certain minority lives in this area, then an implicit discrimination is taking place, the minority faces a disadvantage.” They suggest to question how exactly the data was generated. After all, how is Artificial Intelligence supposed to influence our everyday life, if its decisions cause contortions?
The problem is the formal definition of a fair decision: Fairness is subjective, there is no one definition applicable to all cultures. Hence Kilbertus´ algorithm, a mathematical formula with many variables, exists only on paper. Like most research at the various Max Planck Institutes, it is fundamental research. “At this point it is all theory. We didn´t do any experiments – the idea is just conceptual. No company in the world uses a fair algorithm as we envision it. The problem is the causal reasoning. It would take up a lot of time and generate costs”, says Kilbertus. “What we know is that the current methods aren´t enough to entirely reflect our intuition of fairness.” Nevertheless, Kilbertus´ and his colleagues´ intention is for people to rethink the way Machine Learning and Artificial Intelligence influences the decision-making process in a society. “If machines are to absorb our values, then one mustn´t turn a blind eye on the data. We have to ask ourselves how the data came about. We must question the context in which it were sourced. We must first understand this linkage, before Machine Learning is applied in a social context and before we can accept it as just and fair.”