As humankind builds machines in our own image, it’s become
apparent that this likeness is both a gift and a curse. It’s still in its early
stages, but AI
is advancing rapidly. Unfortunately, it’s already reproducing many of the
biases we mortals suffered under for hundreds and even thousands of years.
What Is Artificial Intelligence Bias?
IBM predicts the number of biased AI systems will only increase in the next
five years. What does this mean? The computer sciences have long been
guided by the principle, “Garbage In, Garbage Out.” If you begin with bad data
or a flawed method, your end-product will be equally imperfect.
Training artificial intelligence involves feeding it with
as much data as you can. Sometimes, this information is text-based databases,
and other times it’s photographs. Then, the AI uses a process called deep
learning to identify patterns. As it learns, it’s able to make useful
predictions and identifications later on, based on unseen data.
The trouble is, AI systems are only as objective and fair
as the information they receive. According to IBM, researchers have identified
and classified more than 180 distinct human biases, any of which can creep into
our algorithms if we don’t take care.
Examples of Artificial Intelligence Bias
One of the most instructive examples of artificial
intelligence bias so far went viral almost immediately. Researchers Trevor
Paglen and Kate Crawford wanted to know what would happen when they used
ImageNet — a collection
of 14 million public images uploaded and labeled by ordinary people
— to train an AI.
Their technology, which they call “ImageNet Roulette,”
used these existing graphics to learn how to tag photos of people that it had
never seen before. The results are still available throughout the Twitterverse.
When people began uploading pictures of their own into ImageNet, the AI
delivered results that ranged from comical to disturbing.
In one example, the AI indicated that a member of the band
Set It Off was a beekeeper — he isn’t. It labeled another band member a traitor
and treasonist. In other instances, the AI slapped labels on people
containing racial or gender slurs too crude and cruel to reproduce here.
Remember — the AI learned its methodology by learning from photos people
already uploaded to the internet and tagged with their own identifiers.
Since then, the creators of ImageNet have made attempts to
de-bias the results. One way to do this is to remove photos and tags that
describe non-visual attributes. In other words, they’re attempting to help the
AI judge fewer books by their covers.
Another example of AI bias involves the world of finance.
A company called AppZen, which does business with Amazon and WeWork, has an AI
finance platform that seeks to expedite the screening, approval or rejection of
loan applicants. The system can sift through the relevant data much faster than
a human, but it’s already demonstrated bias-related issues.
Kunal Verma, co-founder of AppZen, says the system may deliver an unfavorable verdict
if the borrower lives in an area where others have defaulted on their loans
already. He adds: “It may also happen that the area may have a lot of people of
certain minorities or other characteristics that could lead to a bias in the
algorithm.”
Why Is AI Bias a Problem for All of Us?
You might have come across the idea that AI can help us feed the
world, which may well end up being true. However, the whole point of turning AI
toward such a task is to eliminate racial, geopolitical and economic iniquities
humanity already suffers. They don’t have to be, but racism and inequality are
facts of human life right now, and, so far, it appears our machines take after
us in this way.
The idea of a biased AI determining who’s creditworthy
enough for a bank loan based on the community we live in or the color of our
skin should be enough to worry any of us. Plus, there’s the matter of
including AI in recruiting and screening job applicants. This use of AI is
under active development by several tech companies.
One business, HireVue, boasts that its AI video platform can determine which candidates are most likely to
succeed in a certain position. Given what we know about bias in this
technology, this is a dubious proposition. Researchers at the University of
Virginia found that their image-recognition AI held extremely parochial and old-fashioned ideas about which tasks and jobs
are masculine and which are feminine. Sporting paraphernalia was associated
with men. Pots and pans? Women.
Other AI projects demonstrate a similar tendency to sell
women short. After training an AI with text gathered from news stories, researchers
asked the technology to complete the following sentence. “Man is to computer
programmer as woman is to X.” It answered, “homemaker.”
The Future of Artificial Intelligence and Bias
Human beings need to trust each other and work together —
our survival depends on it. It’s also becoming apparent that AI can help us do
many things more efficiently — such as producing and distributing resources.
Just as bias erodes trust and harmony between humans, it does the same between
humans and machines.
Until scientists learn better methods for de-biasing AI,
we would be wise not to put too much of our waking lives on autopilot.
Recent Stories
Follow Us On
Get the latest tech stories and news in seconds!
Sign up for our newsletter below to receive updates about technology trends