AI Can Be Racist: Let’s Make Sure It Works For Everyone.
For decades, artificial intelligence (AI) was mainly just a science fiction trope, like flying cars or jetpacks. But almost overnight, it seems the future has arrived.
Today, AI-driven algorithms and systems are becoming standard operating tools in science, education, customer service, criminal justice, finance, traffic control, and almost any other corner of our lives that you can imagine. Meanwhile, sophisticated AI-driven language processing tools like OpenAI’s ChatGPT (funded by Microsoft, my former employer) can now write complex multi-paragraph essays on virtually any topic within moments. In fact, ChatGPT has already passed the bar and several AP exams. In the words of Stanford professor Erik Brynjolfsson, these chatbots will soon become the “calculator for writing” and radically overhaul white-collar work as much as the Internet and mobile phone revolutions did earlier. (Note: I didn’t use ChatGPT to write this essay, but very soon, you may be unable to tell the difference).
It’s hard not to be simultaneously excited and alarmed by the AI upheaval happening around us. Five years ago – an eternity in AI terms – Google CEO Sundar Pichai called AI “one of the most important things humanity is working on…more profound than electricity or fire.” As we think about significant periods of American history transformed by technology, AI may be as big as the Industrial Revolution. Last month, OpenAI CEO Sam Altman deemed AI potentially “the greatest technology humanity has yet developed.”
While this may all seem like hyperbole, it only takes a few minutes of playing with ChatGPT to feel like we’re on the cusp of some significant changes. That said, there is a huge and troubling AI defect that should give everyone pause. As remarkable as these new AI tools are, many are still subtly perpetuating – or worse, enhancing – the biases of their (mostly white male) creators. Rather than transcend our human limitations to create a more equitable future for everyone, AI tools often inadvertently repeat the same mistakes and lapses in judgment, allowing racism and discrimination to fester throughout our society today.
There are many sobering examples of the harm that AI can cause. In 2018, MIT student Joy Buolamwini wrote about her experience discovering that the facial recognition algorithms in her lab — used all over the world– couldn’t detect Black faces. She even had to wear a white mask to get the computer to recognize her as a person. (You can watch her excellent Ted Talk here.) Similarly, users discovered in 2020 that Twitter’s image-cropping tool constantly focused on white faces. Some mistakes are graver than omissions. AI robots trained on billions of images consistently identified women as “homemakers” and people of color as “criminals” or “janitors.”
In a 2019 panel conversation on these issues, Harvard professor Latanya Sweeney talked about, in her son’s words, “the day her computer was racist.” As she was showing some work to a colleague, search engines served up ads for tools to expunge her non-existent jail records, based solely on her African American-sounding name. These engines, she noted, routinely push predatory credit lending on those with Black-sounding names too. “Many of us came into technology in a belief that [it] would build a better tomorrow and right many of the wrongs that society had produced,” she said, “so you can imagine how shocked I was to see those ads.”
These “algorithmic biases” are more than insulting – they have serious real-world implications for people of color. Today, algorithms help decide credit scores and viable candidates for job openings and college admissions. They “predict” crime and help courts determine who deserves bail and how long sentences should be. They help doctors forecast cancer and mortality rates and decide on appropriate medical treatments. They are interwoven into every facet of our lives, and if they’ve learned racism along the way, they will perpetuate it.
So how do we fix this problem? Remember that Artificial Intelligence isn’t designed to be racist – it’s just picking it up from us. Take it from the AI itself. When I asked ChatGPT if it was racist, it told me: “I do not have personal beliefs or feelings, including racism. I was programmed to provide responses based on the input I received and the knowledge and language patterns I have been trained on. My responses are based solely on the information provided.” And that’s the key. Too often, we are training these algorithms with incomplete or imperfect inputs that lead to unintentionally racist outcomes. As Buolamini put it, “if the training sets aren’t really that diverse, any face that deviates from the established norm will be harder to detect.”
“A model’s blind spots reflect the judgements and priorities of its creators,” writes Cathy O’Neil, author of Weapons of Math Destruction. But the converse can also be true. If we can diversify both the researchers who are creating AI systems and the datasets these algorithms use to learn; we can help teach them better habits and ensure more equitable outcomes. “Training sets don’t materialize out of nowhere,” says Buolamwini. “There’s an opportunity to create full-spectrum training sets that reflect a richer portion of humanity.”
After her experiences with racist AIs, Buolamwini founded the Algorithmic Justice League to encourage more diversity among AI coders while expanding the range and quality of training sets systems are trained on. Luis Salazar, a Seattle tech entrepreneur focused on purpose-driven innovations, has been meeting with AI leaders about these issues and launched AI for Social Progress (AI4SP.org) to encourage the use of more diverse training sets that can mitigate bias in AI technologies.
I strongly encourage business leaders and philanthropists to support these and similar efforts going forward and evaluate the outcomes generated by the AI systems they use for gender and racial bias. AI is now rewriting how we live, work, and play; if we get it right, the future looks remarkable. But we must get serious about expunging systemic bias and racism from these platforms before it’s too late. Let’s work to make AI the dawn of an exciting new era for everyone, not an unfortunate tech-enabled pathway to continue repeating our past mistakes.
