Episode 4: In discussion with Marie Geny - Health Sensors Under the M(A)icroscope

And I think it's a very well explanation to all of these subjects' choices.

Now we will just move on to the IT perspective, because it's important to understand the technical aspects of the tools you studied in order to then understand the legal consequences you analyzed.

Regarding AI and discrimination based on physical appearance, you studied on your paper, you explained that it is possible in several sectors for persons to be discriminated by administrative or judicial decisions, especially due to their physical appearances.

So could you please illustrate a bit more the occurrence of biases? So how does it technologically happen?

So talking about the bias and so how the technical bias happen, I'd say there are two types, you know, two different types of biases.

The first type would be the fact that there are bias that are introduced by the developers during the design phase.

So the developer that are going to choose the variables of the algorithm and they will process like inputs and these influences, they will result in the outputs.

So if the variables are biased at the beginning with regard to gender or ethnicity, then the algorithm will reproduce and even amplify those biases.

So this would be the first type, I would say.

And the second one would be the fact that there is a concern with representativeness.

So this would mean that if there is a lack of diversity in the training data set, This will result in the automatic repetition of stereotypes.

So an automatic repetition of stereotypes is leading to discrimination.

So discrimination here, it refers to the fact that you are treating people who are in similar situations differently.

Or it could also mean that you're treating people in different situations similarly.

So this is from the International Court of Human Rights and also all the work that has been done.

To illustrate with an example, there is the fact that the Amazon AI, the recruitment AI, was known to favour or discovered to favour male candidates over female candidates.

And it was lowering the score of a female profile because when the AI encounters feminized job titles, such as a word where it is stated woman or word, for example, not in this context, but if it says stewardess instead of steward, then this may increase.

have been due either to the male bias criteria that was chosen in the algorithm, or it was due to the fact that it was like under-representation of women in the training data set.

And systematically, it ended up disadvantaging female applicants.

So these are the two different types of bias that I would present in this aspect.

Okay, I understand.