Evolving to a more equitable AI
The pandemic that has raged across the globe over the past year has shone a cold, hard light on many things—the varied levels of preparedness to respond; collective attitudes toward health, technology, and science; and vast financial and social inequities. As the world continues to navigate the covid-19 health crisis, and some places even begin a gradual return to work, school, travel, and recreation, it’s critical to resolve the competing priorities of protecting the public’s health equitably while ensuring privacy.
The extended crisis has led to rapid change in work and social behavior, as well as an increased reliance on technology. It’s now more critical than ever that companies, governments, and society exercise caution in applying technology and handling personal information. The expanded and rapid adoption of artificial intelligence (AI) demonstrates how adaptive technologies are prone to intersect with humans and social institutions in potentially risky or inequitable ways.
“Our relationship with technology as a whole will have shifted dramatically post-pandemic,” says Yoav Schlesinger, principal of the ethical AI practice at Salesforce. “There will be a negotiation process between people, businesses, government, and technology; how their data flows between all of those parties will get renegotiated in a new social data contract.”
AI in action
As the covid-19 crisis began to unfold in early 2020, scientists looked to AI to support a variety of medical uses, such as identifying potential drug candidates for vaccines or treatment, helping detect potential covid-19 symptoms, and allocating scarce resources like intensive-care-unit beds and ventilators. Specifically, they leaned on the analytical power of AI-augmented systems to develop cutting-edge vaccines and treatments.
While advanced data analytics tools can help extract insights from a massive amount of data, the result has not always been more equitable outcomes. In fact, AI-driven tools and the data sets they work with can perpetuate inherent bias or systemic inequity. Throughout the pandemic, agencies like the Centers for Disease Control and Prevention and the World Health Organization have gathered tremendous amounts of data, but the data doesn’t necessarily accurately represent populations that have been disproportionately and negatively affected—including black, brown, and indigenous people—nor do some of the diagnostic advances they’ve made, says Schlesinger.
For example, biometric wearables like Fitbit or Apple Watch demonstrate promise in their ability to detect potential covid-19 symptoms, such as changes in temperature or oxygen saturation. Yet those analyses rely on often flawed or limited data sets and can introduce bias or unfairness that disproportionately affect vulnerable people and communities.
“There is some research that shows the green LED light has a more difficult time reading pulse and oxygen saturation on darker skin tones,” says Schlesinger, referring to the semiconductor light source. “So it might not do an equally good job at catching covid symptoms for those with black and brown skin.”
AI has shown greater efficacy in helping analyze enormous data sets. A team at the Viterbi School of Engineering at the University of Southern California developed an AI framework to help analyze covid-19 vaccine candidates. After identifying 26 potential candidates, it narrowed the field to 11 that were most likely to succeed. The data source for the analysis was the Immune Epitope Database, which includes more than 600,000 contagion determinants arising from more than 3,600 species.
Other researchers from Viterbi are applying AI to decipher cultural codes more accurately and better understand the social norms that guide ethnic and racial group behavior. That can have a significant impact on how a certain population fares during a crisis like the pandemic, owing to religious ceremonies, traditions, and other social mores that can facilitate viral spread.
Lead scientists Kristina Lerman and Fred Morstatter have based their research on Moral Foundations Theory, which describes the “intuitive ethics” that form a culture’s moral constructs, such as caring, fairness, loyalty, and authority, helping inform individual and group behavior.
“Our goal is to develop a framework that allows us to understand the dynamics that drive the decision-making process of a culture at a deeper level,” says Morstatter in a report released by USC. “And by doing so, we generate more culturally informed forecasts.”
The research also examines how to deploy AI in an ethical and fair way. “Most people, but not all, are interested in making the world a better place,” says Schlesinger. “Now we have to go to the next level—what goals do we want to achieve, and what outcomes would we like to see? How will we measure success, and what will it look like?”
Assuaging ethical concerns
It’s critical to interrogate the assumptions about collected data and AI processes, Schlesinger says. “We talk about achieving fairness through awareness. At every step of the process, you’re making value judgments or assumptions that will weight your outcomes in a particular direction,” he says. “That is the fundamental challenge of building ethical AI, which is to look at all the places where humans are biased.”
Part of that challenge is performing a critical examination of the data sets that inform AI systems. It’s essential to understand the data sources and the composition of the data, and to answer such questions as: How is the data made up? Does it encompass a diverse array of stakeholders? What is the best way to deploy that data into a model to minimize bias and maximize fairness?
As people go back to work, employers may now be using sensing technologies with AI built in, including thermal cameras to detect high temperatures; audio sensors to detect coughs or raised voices, which contribute to the spread of respiratory droplets; and video streams to monitor hand-washing procedures, physical distancing regulations, and mask requirements.
Such monitoring and analysis systems not only have technical-accuracy challenges but pose core risks to human rights, privacy, security, and trust. The impetus for increased surveillance has been a troubling side effect of the pandemic. Government agencies have used surveillance-camera footage, smartphone location data, credit card purchase records, and even passive temperature scans in crowded public areas like airports to help trace movements of people who may have contracted or been exposed to covid-19 and establish virus transmission chains.
“The first question that needs to be answered is not just can we do this—but should we?” says Schlesinger. “Scanning individuals for their biometric data without their consent raises ethical concerns, even if it’s positioned as a benefit for the greater good. We should have a robust conversation as a society about whether there is good reason to implement these technologies in the first place.”
What the future looks like
As society returns to something approaching normal, it’s time to fundamentally re-evaluate the relationship with data and establish new norms for collecting data, as well as the appropriate use—and potential misuse—of data. When building and deploying AI, technologists will continue to make those necessary assumptions about data and the processes, but the underpinnings of that data should be questioned. Is the data legitimately sourced? Who assembled it? What assumptions is it based on? Is it accurately presented? How can citizens’ and consumers’ privacy be preserved?
As AI is more widely deployed, it’s essential to consider how to also engender trust. Using AI to augment human decision-making, and not entirely replace human input, is one approach.
“There will be more questions about the role AI should play in society, its relationship with human beings, and what are appropriate tasks for humans and what are appropriate tasks for an AI,” says Schlesinger. “There are certain areas where AI’s capabilities and its ability to augment human capabilities will accelerate our trust and reliance. In places where AI doesn’t replace humans, but augments their efforts, that is the next horizon.”
There will always be situations in which a human needs to be involved in the decision-making. “In regulated industries, for example, like health care, banking, and finance, there needs to be a human in the loop in order to maintain compliance,” says Schlesinger. “You can’t just deploy AI to make care decisions without a clinician’s input. As much as we would love to believe AI is capable of doing that, AI doesn’t have empathy yet, and probably never will.”
It’s critical for data collected and created by AI to not exacerbate but minimize inequity. There must be a balance between finding ways for AI to help accelerate human and social progress, promoting equitable actions and responses, and simply recognizing that certain problems will require human solutions.
This content was produced by Insights, the custom content arm of MIT Technology Review. It was not written by MIT Technology Review’s editorial staff.