Are Computer Scientists Building Racist/Sexist AI?
When we mention artificial intelligence, most people imagine an anthropomorphic robot with a computer instead of a brain. And computers are capable of making data-driven decisions based on cold, hard facts and information.
Does this mean that these machines can’t be biased, impartial, sexist, or racist?
Unfortunately, AI is nothing more than the simulation of human cognitive processes executed by machines, and as such, it’s prone to the same preconceived notions as the data it’s fed.
In other words, the source of AI’s potential bias lies in the fact that it’s based on human behavior, which is inherently discriminatory in different ways.
The question is rather, “is it possible to make AI fair and non-discriminatory?”
The Root of the Problem
Elon Musk. Bill Gates. David Hanson. Ashok Goel.
You’re already familiar with the first two names, and the other two belong to an engineer who invented Sophia the Robot and a professor of computer science and cognitive science at the Georgia Institute of Technology, respectively.
My point being?
If we scratch the surface, we’ll spot inequality in the AI industry itself. Most leading tech leaders, inventors, and computer scientists are male.
As a matter of fact, only 12% of machine learning researchers are women. And we can’t deny that this disproportionate ratio can have an influence, maybe even subconsciously, on how AI-powered technology operates.
It is disheartening to think that so few women participate in creating such a game-changing and revolutionary technology, one that will shape our future society.
How can we expect man-made technology to be fair and non-discriminatory when women are a minority in the industry behind it?
AI is data-driven, and this is what makes it so powerful. For example, conversational chatbots are capable of interacting with customers in a human-like manner and offer them relevant and personalized answers and solutions thanks to the fact that they can collect and process a massive amount of customer information, and learn from it.
However, it seems that all eyes are on scientists who are training increasingly sophisticated AI algorithms, but somehow the general public fails to pay attention to the actual data these algorithms use for their functioning.
And that’s another big issue responsible for AI’s sexism, racism, and other forms of discrimination.
These data sets are obtained by scraping Google Images or Google News or aggregating easily accessible information from websites like Wikipedia. Then follows the annotation process that’s usually outsourced to graduate students or through a crowdsourcing platform such as Amazon Mechanical Turk.
This way, racial, gender, and cultural biases can be unintentionally incorporated into collected and annotated data.
The thing is that machines need our help when it comes to interpreting and making sense of all that data. For example, facial recognition isn’t something that comes naturally to AI-powered algorithms – they need an explanation as to how to tell people apart and what to pay attention to.
A striking example of the way AI deals with face recognition has been observed by female computer scientist Joy Boulamwini. She noticed that AI facial recognition systems couldn’t recognize her face, but had no problem identifying her lighter-skinned friends. To make things worse, the system would recognize her as a human when she wore a white mask.
Boulamwini decided to explore this bias problem and fed facial-recognition algorithms with a set of images representing people with a number of different skin colors.
There’s another peculiarity that she observed – these systems delivered more accurate results and performed better on males than on females. Needless to say, these facial recognition systems displayed the worst performance when the subjects were dark-skinned women.
One more thing: Sophia the Robot – she’s white too.
AI-Algorithms Mimic Real Life
Another factor responsible for this bias is a combination of two things. One is a feature of machine-learning algorithms whose role is maximizing overall prediction accuracy for the data used for training.
Now, if a certain group of people such as those with lighter skin tones appear more frequently in the training data, the machine-learning program will optimize prediction accuracy for that particular group.
In other words, if you run a search for the term “grandma,” Google’s algorithm will mostly show you images of elderly women who are light-skinned.
But, is that how all people in the world would imagine a grandma?
Of course, it’s not yet possible or necessary to take everybody’s idea of this concept into consideration, but the trouble is that Google’s AI-powered algorithms show almost no diversity or inclusiveness. And these unapologetically white grannies show the state of the world we live in.
And AI is only its reflection.
Not to blame Google for everything, but the search engine giant has yet another biased algorithm.
Google Translate, for example, will, by default, use masculine pronouns. We can thank the ratio of masculine to feminine pronouns in English corpora being 2:1 for this peculiarity, but every time this program uses a “he” by default, the relative frequency of masculine pronouns on the internet is increased.
Artificial intelligence is undoubtedly the technology of tomorrow, but if we want it to serve us well and help us create a better world, we have to make it fair today. Additional efforts should be invested in collecting and processing the training data so that it’s more inclusive and non-discriminatory.
Jennifer Wilson is a writer at Qeedle.com She knows business processes and operations management inside out. As she understands all the challenges of running a small business firsthand, it’s her mission to tackle the topics that are most relevant to entrepreneurs and offer viable solutions.