This semester (Spring 2023), I got an opportunity to assist Dr. Charmaine Royal with her course Race, Genomics, and Society. We had our last class in the past week and Dr. Royal asked us to take the learnings from the class to real-world. In other words, moving from bench side to curbside when it comes to discourse on race. My posts this week are an attempt to reflect on how AI impacts the existing racial inequities in a positive or negative way.
Can we disassociate race from AI?
As ML advances and becomes a big part of our society, it becomes important to think about how it interacts with social strata like race, gender, sexual orientation, etc. I see a lot of emphasis is laid on increasing model accuracy or creating the most state-of-the-art deep learning models but there is a little conversation around how higher accuracy doesn’t necessarily lead to fair, just, and equitable outcomes.
There are some very important works being done by researchers in highlighting cases of biased AI algorithms, for example, facial recognition failing to work with dark-skinned people and crime prediction algorithms unfairly targeting Black and Latino people for crimes they did not commit.
Incidents and cases like these demand introspective conversations on whether we can disassociate race from AI or not.
Are algorithms really fair?
Computers are fair.
Algorithms don’t differentiate.
We all expect technology to be unbiased.
But how true is that?
“Although these automated systems of technology are purported to be more efficient, precise, and objective than the human, it has become now widely known that the technologies are masking the reproduction of inequalities and social histories of sociopolitical violence; further indicating that maybe the human bar is not enough,” says Dixon-Román, Professor of Critical Race, Media, and Educational Studies and Director of the Gordon Institute.
How does this happen?
The data on which these ML algorithms are trained are biased and the outcomes returned are also biased. Take for example statistics from a paper by Hutson et al., 98% of oncology-related data being from the white population. When trained on this data, the ML model starts treating them as noise and results in higher chances of misclassifying with a possibility of cancer in people of color.
To avoid this, the first thing we can do is to ensure we start focussing on data quality and use a balanced dataset for treating AI models for life-threatening conditions like cancer. Next, we can focus on improving our data quality standards that help developers with clear guidelines and expectations.
Race, healthcare, and AI in the mix
If we don’t address the importance of having a balanced dataset, then we will end up perpetuating racial disparities that exist in healthcare.
A team at Emory University made use of large-scale medical imaging datasets from both public and private sources, datasets with thousands of chest X-rays, chest CT scans, mammograms, hand X-rays, and spinal X-rays from racially diverse patient populations.
They found that standard deep learning models — computer models developed to help speed the task of reading and detecting things like fractures in bones and pneumonia in lungs — could predict with startling accuracy the self-reported race of a patient from a radiologic image, despite the image having no patient information associated with it.
What’s even more surprising is that the AI models could determine race more accurately than complex statistical analyses developed specifically to predict race based on age, sex, gender, body mass, and even disease diagnoses.
This leads to discriminatory access to healthcare and deepens the already existing inequities. We need to start harnessing the capability of AI models to detect race with such accuracy to provide targeted support, instead of letting the AI models determine patient outcomes based on race.
Do AI developers have ethical responsibilities when it comes to race?
Yes, the ethical responsibilities of AI developers become even more crucial. It is not enough to simply create functional and efficient systems. Developers must also consider the potential impact of their systems on society as a whole, including the impact on different racial groups and marginalized communities.
To fulfill this ethical responsibility, developers must actively seek out potential biases in their data and work to mitigate them. They should also consider the potential impact of their systems on different communities and work to ensure that their systems are not perpetuating systemic inequalities. For example, an AI system used in hiring should be designed to avoid bias against certain groups, such as women or people of color. Similarly, an AI system used in criminal justice should be designed to avoid perpetuating the disproportionate impact of the criminal justice system on people of color.
In addition, AI developers should consider the long-term implications of their systems. They should be aware that the impact of their systems may change over time and work to ensure that their systems remain ethical and socially responsible as circumstances change. This may require ongoing updates and revisions to their systems, as well as ongoing monitoring and evaluation to ensure that their systems are meeting their ethical responsibilities.
The ethical responsibility of AI developers is to create systems that are not only functional and efficient but also socially responsible and ethical.
What can we do as an AI/ML community to make fair, just, and equitable AI systems?
The first step is to acknowledge and have conversations around racial biases existing in AI systems. The second step is to demand more transparent systems aka enhancing explainability of the algorithms. AI systems should be designed in such a way that their decision-making process is transparent and understandable. Moreover, AI systems should be auditable and accountable. The AI/ML community can work towards developing standards for transparency, accountability, and audibility of AI systems.
The next step is to foster collaboration between different stakeholders. The development of AI systems requires input from people from different backgrounds, including computer scientists, social scientists, policymakers, and community members. Bringing together these different perspectives can help identify potential biases and ensure that AI systems are designed to serve the intended population. The AI/ML community can work towards creating collaborative platforms that bring together different stakeholders to design and deploy fair and equitable AI systems.
By developing diverse and representative datasets, creating transparent systems, and fostering collaboration, we can create AI systems that serve everyone equitably.
Let’s do our bit.