Is it Possible to Create Artificial Intelligence Without Bias?
It is also an almost essential technology for many people who browse the internet using screen readers. All this is thanks to artificial intelligence.
Of course, the machines are not perfect; sometimes, they fail. But lately, we are discovering something worrying: sometimes computers fail as humans fail.
For a woman and a man with the same expression, AI systems may tend to believe that she is happy and he is in a bad mood. We call this type of error bias, including racist, sexist, and ableist tendencies that can end up harming people.
The Face as a Mirror of the Soul
To study these biases, we will focus on a specific application, the automatic recognition of emotions in photographs.
First, we need to clarify what we mean by “emotion.” The most used classification is based on six basic emotions: fear, sadness, happiness, anger, disgust, and surprise. Psychologist Paul Ekman proposed this classification in the 1970s.
These emotions are more or less universal and recognized by everyone. However, it has also been shown that they recognize each other somewhat better between people of the same social group, gender, age… We do not all express ourselves precisely the same, nor do we read the expressions of the rest in the same way. Even without realizing it, we are biased.
These differences are seen in many contexts and sometimes turn into stereotypes and prejudices. For example, we expect women to be more happy than angry, and the opposite is true for men. And this is reflected on the internet, where the photos include, above all, smiling women.
On the other hand, for an artificial intelligence system to learn to distinguish these emotions, we also need to think about how people understand them.
The face is only one part of a very complex puzzle. Gestures, posture, and our words also contribute… Although work is being done to resolve all these modalities with artificial intelligence, the most popular and versatile form is recognition based on face photos.
How Does an Artificial Intelligence Learn?
Creating bias-free artificial intelligence is quite a challenge. And it all starts with how we make this technology ‘learn.’ We call the field of artificial intelligence dedicated to this learning machine learning. Although there are many different forms of knowledge, the most common is supervised learning.
The idea is simple: we learn from examples. And artificial intelligence needs to know what we want to obtain for each example. To learn to recognize emotions, we need many photos of faces with different emotions: happy, sad, etc. The key is that we must know what emotion appears in each image.
Later, we passed the photos and their associated emotions to artificial intelligence. Through a learning algorithm, the system will learn “only” to relate the images with the emotions that appear. Picture by picture, we ask you to predict an emotion: if it’s right, we move on, and if it’s wrong, we adjust the model to correct this case.
Little by little, you will learn and fail fewer and fewer examples. If we think about it, it’s not that different from how we humans learn.
As you can see, examples are essential in this process. Although advances allow us to learn with few examples or errors, a large and well-labeled set of examples is vital to achieving good artificial intelligence.
Unfortunately, in practice, it is common to have examples with errors. In our case, they would be from faces labeled with the wrong emotion to photos without faces or animal faces. But there are other problems, sometimes more subtle and worrying: racism, sexism, ableism…
When the Algorithms Go Wrong
As you can imagine, if our examples are biased, the machine will learn and reproduce these biases. Sometimes it will even multiply the effect of the biases. For example, artificial intelligence will likely confuse skin color with mood if we only have angry dark-skinned people and happy light-skinned people in our photos. He tends to predict anger whenever he sees dark-skinned people.
Unfortunately, this is not just a theory. It has already been shown, for example, that facial analysis systems to recognize gender fail more for black women than for white men and that they regularly make mistakes with people who are trans or non-normative in appearance.
One of the most notorious examples was when in 2018, an artificial intelligence system mistakenly identified 28 US members of Congress as criminals. Of the politicians identified, 40% were people of color, even though they only represented 20% of Congress.
All this was because the system had been trained primarily on white people and confused people of color.
Detecting and reducing these biases is a very active field of research with a tremendous social impact. Many of them are subtly related to several demographic factors simultaneously, making it a complex analysis.
In addition, all phases of learning must be reviewed, from data collection and its measurements to the final application. And usually, it is not the same people who work in each phase.
One Database to Include Them All
Let’s go back to emotion recognition. On the internet, there are many databases of emotions already labeled. Unfortunately, the most extensive databases also have strong sex/gender, race, and age biases.
We must develop diverse and balanced databases on which to work little by little. We need to include all kinds of people in our databases. In addition, all of them must be well represented in each emotion.
Finally, if we want to collect data without bias, we must think about the whole process. All phases, from data collection to the final tests of artificial intelligence, must be carried out in a careful and accessible manner. And it is necessary to involve people who can recognize and point out possible biases in all of them.
And All This For What?
This whole business of recognizing emotions may sound abstract, but it already has critical applications. The most common is assistive technology, such as automatic photo descriptions for the visually impaired. It is also already used in domestic robots.
I can even apply it in medicine, where it has been possible to automatically recognize pain in newborns who sometimes do not express it through crying.
In any case, the study of biases in artificial intelligence goes beyond emotions. The technologies we develop have a significant impact on people’s lives. We have a moral duty to make sure they are fair and that their effect on the world is positive.
We want to build artificial intelligence that we can trust that makes us smile.