Tribune. Three months after the launch of the Apple credit card in August 2019 in the United States, the co-founder of Apple Steve Wozniack accuses the service of discriminating against women by assigning credit lines ten times lower than men! The reason ? The algorithm automatically determines the credit line. Yes, ok, but what is the explanation? Steve Wozniack and his wife are still waiting for the answer …
More and more examples in justice, health, education and finance, show that artificial intelligence (AI) tools can not be deployed without control in security systems or access to resources essential to the risk of generalizing bias, potentially discriminatory, difficult to interpret and for which no explanation is provided to users.
The conclusion is increasingly clear: AI must integrate ethics from the design of the algorithms (” ethics by design »). The ethical performance of the algorithm (no discrimination, respect of individuals …) must be included in the performance criteria, as well as the accuracy of the predictions.
Laws, morality and culture
But integrating ethics into algorithms raises daunting challenges for five reasons.
First, ethical and legal standards are often unclear and do not lend themselves to mathematical formulation.
Secondly, ethics is not universal. Laws and moral values are associated with culture: what is considered morally acceptable in France is not in Japan, for example. This is crucial because it implies that the codification of ethical rules or automatically learned by the algorithms will have to be adapted to the legal and cultural system of a group of people, a sector, a country in which it is deployed.
Third, ethics is political. The creation of a fair and non-discriminatory system requires compromise and the balancing of conflicting interests. On the one hand, it is impossible to build a universal algorithm to simultaneously prevent all forms of discrimination (gender, ethnicity, etc.).
On the other hand, it is impossible to guarantee both individual equity – for example, equal individuals should be treated equally regardless of their ethnic origin or gender – and group equity – individuals in essence gender, for example, must be treated unequally to restore social equity. The algorithm should therefore be guided by political choices.