Can robots inherit human bias? Yes. Now, the harm has a face.

The process used two AI algorithms. The first was trained to look at images of people and tell them apart from other images. The second generates images to try to fool the first algorithm into thinking that its generated image belongs to a group of real people I photographed in my studio. This process iterates and the second algorithm continues to improve until it consistently fools the first algorithm.

The website thispersondoesnotexist.com used this type of algorithm to create stunningly realistic images of people who, as the name of the website makes clear, don’t exist. What I did differently was to photograph my original, real subjects using a technique called “light painting.” During a 20-minute exposure, I used a flashlight to illuminate each person’s face unevenly while the subjects were moving, creating images of the subjects with parts of their faces distorted or missing. The images created by the algorithm are, in turn, distorted. If you were creating a representation of a human and you didn’t have all the information to put it together, you would end up with distorted ones such as these.

When a mortgage company, a recruiting service or a crime-prediction software develops a distorted version of people, it is an invisible kind of harm. These photographs make the pain visible by applying the process to a human face.

What can we do to prevent bias in AI, and the harm it causes?

One important aspect of good data is that it needs to have breadth and depth. For example, looking at data on a large number of customers, and deeper data on each customer. This enables models to handle situations better and more predictably and help reduce bias; in fact, it was this lack of breadth in data that Amazon had to deal with in its recruiting software. AI researchers are defining more ways to improve fairness for groups and individuals.

Some solutions that seem promising, though, don’t actually work. Turns out, removing protected attributes, such as gender, race, religion or disability, from the training data before modeling does nothing to address bias and even possibly concedes it. That is because this “fairness through unawareness,” as it has been called, ignores redundant encodings — ways of inferring a protected attribute, such as race or ethnicity, from unprotected features, such as, say, a ZIP code in a highly segregated city or an Hispanic surname. To address this, we remove the attributes that are highly correlated with the protected attribute. The algorithm can also be checked for false positives and false negatives early on.

In 2014, Stephen Hawking speculated aloud on a future many Hollywood films have depicted: “The development of full artificial intelligence could spell the end of the human race.”

This disturbing quote is often thought to refer to phenomena such as self-conscious, AI-enabled robots that could eventually take over the world. While the AI in current use is far too narrow to bring about the end of humans, it has already created disturbing problems.

What many are not aware of is what Hawking said next: “I am an optimist, and I believe that we can create AI for the good of the world. That it can work in harmony with us. We simply need to be aware of the dangers, identify them, employ the best possible practice and management, and prepare for its consequences well in advance.”

Making AI fair is not just a nice idea, but a social imperative. If we study and question the technology from many angles, AI has the potential to improve the quality of life of everyone on the planet, raising our earning potentials, and helping us to live longer and healthier lives.