In addition, her scientific evolution and career, focused on the design of Artificial Intelligence (AI) tools to address problems in different sectors, make her the ideal person to ask about what’s behind Machine Learning, why there are biases in this type of technology, and what this means for society.

Can you tell us a little about yourself? Who is Itziar Landa?

I’m a person with a technical profile, although I’ve evolved a lot. Now I work in management, but it took me a while to get here, because I felt most at home on the technical side of things.

We’re always trying to work to blur the barriers between the Sciences and Letters. When managing people and when working in more technical fields, do you think you have that open, holistic perspective and that you incorporate it into your daily routine?

Totally. I believe that professionals with Machine Learning profiles tend to come from Physics and Mathematics, almost more than the “natural” evolution that is expected – from Computer Science and Engineering. This is very connected to the mathematical level of abstraction; it’s all tied together.

For example, in philosophy and in reasoning, you have to have a higher level of abstraction to defend other positions and other ideas. You have to go beyond your own way of thinking, to an abstract sphere, to be able to understand it. This happens a lot in Mathematics because, when they say that it’s deterministic (that is, 1+1=2), this is not the case.

If you want to make progress in Algebra, in Physics, even in Telecommunications, a level of abstraction is needed that goes beyond memory and logical reasoning alone.

That ability to think things through, to go beyond your barriers and take it further – that is obtained through many degrees falling under Letters, and in many degrees in the Sciences, especially when you are doing a PhD program. In the end, it is about taking the topic on which you want to “expand” your knowledge to another level. And just as this can attract some people, it can also deter others.

What is the process you need to succumb to in order to reach that level of abstraction?

To get to that point in the conversation, you have to warm up and prepare your mind. They say that research is very lonely, but that’s not true: when you share those ideas with someone, it’s magical. I still feel that the formation of ideas and problem-solving, on a team or with someone who stimulates you in this regard, is vital.

In our daily lives we see how voice assistants, chatbots, and even facial recognition programs reproduce intersectional discriminatory biases. Could you explain to us the process for designing algorithms and data training in Machine Learning?

We shouldn’t be surprised that this is happening. The examples in music, virtual assistants, and even some tests that have been done to create robot judges with AI, feature biases because the algorithms have a part that is training, and another part that is programming.

When we accuse the algorithm, we must remember that while its design is very important, and the programmer has a hand in everything, a very large percentage of the algorithm’s proper functioning depends on the data provided. The data is nothing more than a reflection of reality. An algorithm will learn from what you teach it, and the reality is that the world has biases: it’s sexist, homophobic, racist… this is a reality. And these data sets contain those biases.

Can you give us an example with everyday services or products?

Yes, a very controversial case was an Apple Pay line of credit in which dramatic gender biases were detected. Everyone blamed the algorithm, and that’s something that can also be analyzed. But, on the other hand, what data did they train it with? For example, a simple algorithm is used to classify “yes” or “no.” If you train an algorithm with a million people, giving it data about them and whether or not they were given credit, it learns through all the examples which combination of features is the best and how to separate “yes” and “no.” On the technical side, you have to manage to separate the data dichotomously: “yes” and “no.” That is what is programmed for… the different parameters to streamline that classification. A good algorithm is one that leaves everything well classified.

Now, if you have fed in data in which loans were granted to men 99% of the times, the algorithm will learn from that reality. Therefore, when the algorithm is used to analyze a new case, it will have a bias.

In this process, how could these types of biases be counteracted? Where does the human’s role come in?

As a woman, what can I say? The ideal thing would be for the world to change. Technology has to be objective and, right now, that algorithm is objective. What it’s doing is working with subjective data. We are blaming technology when the most crucial point (and I’m not saying it’s the only one) is the data it uses.

We can either modify that data, or we can modify the algorithm. And we can also put mechanisms in place to monitor these algorithms to detect their impartiality.

These three options would translate, technically speaking, into there being a branch of technology that actually exists, and is advancing: “Explainable AI.” It’s not a question of black box models. We have heard a lot of “We don’t know what Google’s algorithm bases its decisions on.” But we do know.

There are models that are explainable. What we do is draw a decision tree that says something along the lines of: “Can a loan be given to this individual?” The first branch of the tree says, “Man or woman?” “Yes or no?” And you move down the branches, answering yes or no questions. “Do they have a stable job or not?” At the end it tells you what the characteristics are and the probability of each final decision. This is an example of how an explainable algorithm could be done.

Why aren’t all models made explainable?

Today explainable models are not those with the best success rate, but they do exist. Technology can advance down that road so that these types of models are used more than the black box type, or neural networks…. Those are much less explainable. However, it has been decided to move towards a technological line that produces better results; one with fewer false positives, but that’s not explainable.

Other measures have also been taken. There are some courts in which the algorithms that are being used are measured to determine their degree of impartiality, but I really see this as complex, because you want to make something objective, and on top of that you’re adding another layer of subjectivity.

Of course, other techniques are being used, such as filtering the data to verify that the datasets introduced into the algorithms are more objective… but there you’re also modifying things.

On the importance of including different perspectives to minimize biases, how should we proceed?

There are areas in which different perspectives can make things easier; for example, to classify failures in industry, when working with animals, when detecting droughts with drones. In these cases, the degree of subjectivity is much less. On the contrary, if you want to try to minimize biases in algorithms meant to replace judges in courts, no… including different perspectives isn’t going to help you there.

You have to think about the aim for which you want to use the algorithm. Some algorithms will have to be used only to support decisions, and others could, perhaps, be automatic. Decision support, for example, could take the form of helping someone to quickly look up similar past judgements or laws. It could process and provide information to help that person make a decision.

However, facial recognition, traffic management, CCTV cameras… all of that could be automated with less risk of biases. An algorithm doesn’t have biases related to the color of the car, its size, etc. Biases are always related to people.

AI can learn about “cause and effect” and “action/reaction,” but it will never understand the actual logic behind that leads one to perpetrate a crime - something that may be considered in a trial. Faced with a situation and a result, two people may do the same thing, but the degree of intention and emotions that each person has cannot be put into neat parameters. Therefore, an algorithm will not be able to discern what the motive behind the action was. It will understand the cause and the effect, but I don’t think that’s what we’re after. There are certain things that will always be left for people to do – and that’s a good thing.

How do you think technology will evolve in the future regarding intersectional biases? Will we be able to counter them? Is there a real interest in doing so, and are efforts being made?

Science is revolving around AI explainability, because that’s really the direct part that can be improved… the rest is a question of data processing.

Progress is also being made with AI ethics committees, although we can’t audit algorithms at private companies yet. We’re trying to work on the ability to undertake audits and on getting private companies themselves to make their algorithms open-source, to prove that they’re transparent and have no biases.

For example, in Medicine, all the algorithms that are being used to prioritize one patient over another on a waiting list, to undertake the early detection of cancer, etc. All of these algorithms in the Public Health field have to be able to be audited. Since we’re working on that, we want all companies to join this open-source initiative. Progress is being made on making the philosophy behind the algorithm transparent; people are already becoming aware that they can be audited… and how.

Finally, we want to give you some time so that you can tell us about your views on this subject matter from a more personal standpoint. Do you want to contribute anything else to the conversation?

Personally, I think’s it relevant to note that the presence of women in Science has fallen again, and this will be reflected in companies in a few years. I think we have to make a big effort to work on biases in education.

By the time students turn 18, they have too many biases in their head when choosing what to study in college. It’s normal to be confused, or to not know what you want, but part of that confusion needs to be tackled by society, since there are biases that dissuade women from starting to study Science, Technology, or Engineering if they’re not 100% sure of her decisions.

Women feel that when they opt to study Engineering, they have to go all the way... they must get exceptional grades. Women have to be very sure of themselves, or a little mad, or have an army of support behind them so that, if they’re only “C” students, they feel confident and start studying Engineering anyway. Men, however, don’t think twice about giving it a try, despite only being “C” students; women don’t do that.

If you don’t see yourself as intelligent, you aren’t going to study for a degree that is considered a challenge... you aren’t going to “expose yourself.” Men, from a very young age, already have a certain “innate” courage (fostered through education and society from the time they’re very young) to take on challenges. When you analyze the average secondary school grades that females had when starting Technology studies, vs. those of men, the differences are huge.

For us girls, it’s a question of going against social norms. “Are you really going to study Telecommunications if you don’t have an ‘A’ average? Who do you think you are?” But men just dive right in. Society has encouraged them to be engineers, and women to be more cautious. Then, of course, the relationship between grades in one’s studies and professional success is another debate.

You have it so instilled that it becomes a reality, because we experience it as such. When do we have to nip this in the bud? Before these things get too far, because even the most self-assured and confident girl has biases instilled in her. So, as I was saying, we must reverse this trend, and the crucial tool to accomplish this is education from childhood.