Calculating the shortest route on our phones, automatically creating playlists with our favorite songs, or even finding the most relevant result on a search engine: algorithms help you every day. But what would happen if a recruiting algorithm systematically left out women or ethnic minorities? How can we make sure these errors  are acknowledged and corrected?

To answer these questions we interviewed forty experts from various sectors with the aim to offer concrete solutions to  limit potential abuses and increase  public trust in algorithms.

This report proposes a French perspective on algorithmic bias,  essentially viewed today through an American lens. It builds on the study published by Télécom Paris and Fondation Abeona, Algorithms: Bias, Discrimination and Fairness in 2019.

What is an algorithm?

An algorithm is a sequence of operations or instructions towards an objective. In many ways, a recipe is an algorithm. Algorithms exist in many forms. One of these forms, machine learning, has undergone significant developments in recent years.

Machine learning algorithms learn on their own, which means that they themselves create a logic to  reach a conclusion based on access to numerous examples. If asked whether a cat is in a picture, they will formulate the solution such as: "given that there are sharp ears, I am 90% confident that it is a cat".

Why can algorithms be biased?

A bias is the mean deviation between the prediction made and the actual value. Often, bias comes from the data being used to train the algorithm. If, for example, the pictures of cats shown to the algorithm are always taken on a carpet, but not dogs, the algorithm will consider that having a carpet in the background is associated with the presence of a cat. If it is then presented with an image of a dog  on a carpet, the algorithm would conclude that the picture is that of a cat and thus introduce a bias.

Why can algorithms discriminate?

In certain cases, biases could lead to discriminations. A classification  algorithm  could produce a different decision depending, for example, on a person's gender or ethnicity. A recruitment algorithm, for example, could be given a task to identify people with best profiles suited for a particular job. Analyzing historical data, the algorithm will observe that  90% of people who hold these positions are men. The algorithm would then conclude that women are rarely qualified to do this job, create a rule based on this observation and not recommend women to the recruiter.

Designing algorithms to reduce discriminations

The fight against algorithmic bias is therefore, above all, a fight against discriminations that pre-exist algorithms. Algorithms, despite the risk of bias, represent in many ways a progress to fight discriminations as they can ensure objective decisions that could otherwise be biased. The challenge is not only to produce algorithms that are fair, but also to reduce discrimination in society. After discovering a bias in society, we could thus program an algorithm to correct it.

Choosing individual fairness or group fairness?

The main challenge is therefore to create fair algorithms. But first, fairness is not as easy as it seems: is it about judging everyone on the same criterias (individual fairness)? Or is it about ensuring that no group (gender, ethnicity, etc.) is discriminated against? Unfortunately, both are mathematically incompatible, which means that we cannot have a total individual fairness and a total group fairness.
 
Indeed, fairness can be conceived in two different ways:

• Individual fairness, which ensures that individuals with similar profiles are treated in the same way, taking into account the individual characteristics to adapt the decision
   
• Group fairness, which ensures that the decision-making process does not arbitrarily favor a certain group, taking into account the fact that the individual belongs to a group to adapt the decision.

Measuring algorithmic bias

Measuring bias  requires identifying  populations potentially affected by discriminations. For example, to know if an advertising algorithm discriminates against women, we need to know how many women were selected to receive the advertisement and therefore have access to the variable "gender". In France, in certains cases, such as ethnic origin, the data is protected - and rightly so. However, without access to such data we cannot know if discrimination occured.

A performance challenge

Although correcting an algorithm to make it fair is a priority, the correction could also affect its performance. Therefore, there is a cost of reduced performance that needs to be taken into account when combatting algorithmic bias.

An innovation challenge

The challenge of reducing algorithmic bias is also achieving an equilibrium between protecting citizens against discriminations, while  promoting  innovation and supporting the digital economy. Restricting the use of algorithms means curbing the growth of the French digital industry and potentially accepting American and Chinese technological superiority. However, adopting a laissez-faire approach implies  ignoring the potential negative effects on our social fabric.