EAI Researchers Reveal How COVID-19 Lifted the Fog on Racial Disparities in Prison System

By: Tyler Wells Lynch

The United States incarcerates a higher percentage of its population than any other country on Earth, and the makeup of that population is extremely unequal: Black and latino people are vastly overrepresented. 

But in the first year of the COVID-19 pandemic, penitentiaries saw dramatic population declines, as a spate of early releases and stalled admissions decongested prisons nationwide. Between March 2020 and July 2021, the overall U.S. prison population fell by a stunning 17%, or about 200,000 people. However, this historic dip was not matched by a proportional reduction in racial disparity. In fact, quite the opposite. The real story of the pandemic was that it lifted the fog on inequities already present in the U.S. prison system.

Hidden Injustice

In a new paper published in the journal Nature, researchers from the Institute for Experiential AI (EAI)—including Sam Scarpino and Tina Eliassi-Rad—as well as representatives from several leading universities reveal how racial disparities in prisons deepened during the pandemic, disproportionately benefiting white people. While the percentage of incarcerated black individuals rose by 0.9 percentage points to 39.8%, for white individuals it fell by nearly half a percentage point.

Why? Researchers first looked at admissions and release data to explain the discrepancy, theorizing that courts and parole boards were favoring white inmates. They also thought societal imbalances in who was able to shelter in place or who was encountering the criminal legal system might explain the differences. Sam Scarpino, Director of AI + Life Science at EAI and one of the study’s authors, specifically cited releases: “What happened during the pandemic was this huge shock to the system and the courts closed. So for a period of time, there were almost no admissions into jail and prison systems.”

But as researchers looked closer at the data, they found release and admission data alone could not explain the discrepancies. In 18 states, for example, the rate of prison admissions among black individuals was lower than it was for white people during the early months of the pandemic. Similarly, release rates in many states were actually lower relative to pre-pandemic values. So why the uptick in racial disproportion? In a word: sentencing.

The Legacy of Long Sentences

To look at racial disparities in the prison system over the past decade, you could be forgiven for thinking things were moving in the right direction. Despite being greatly overrepresented, the proportion of incarcerated black people was on a steady decline, falling from 41.6% in 2013 to 38.9% in 2020. One positive outcome of the pandemic was that states became more generous about releasing their incarceration data, largely to reflect what they perceived as the good news of shrinking populations. However, the truth behind the data was much more insidious, revealing a legacy of inequity going back decades. 

“Black individuals, Latino individuals, Native Americans—anyone who is non-white in this country—are sentenced for longer periods of time for the same crime than white individuals,” Scarpino said. “So what we were seeing is that it wasn't who was being released, it wasn't who was being admitted—it was who was left behind.”

COVID-19 showed how social injustice persists in ways that datasets sometimes elide. Researchers have known for some time that prison populations are racially disproportionate, but precisely what’s driving those disparities is harder to explain.

“Is it prosecutorial discretion in terms of who gets charged and who gets taken to sentencing versus pretrial diversions?” Scarpino asked. Or are other factors at play, such as the frequency of police encounters in certain neighborhoods, prohibitively expensive cash bail systems, mandatory minimum sentences? While the answer is certainly “all of the above,” observational data like these don’t lend themselves to the kinds of causal relationships that data scientists seek to identify.

Unequal Inputs = Unequal Outputs

It also doesn’t help that courts are using these data in all kinds of irresponsible ways. The most notorious example is the use of algorithms to gauge the risk of recidivism for parolees, which researchers have shown perpetuates discrimination. 

“If you have a system that is inequitable then the AI will learn about that inequity and will amplify it, and no amount of data can remediate that,” Scarpino said. “One of the things that we need and that we've learned in the criminal legal system is that you have to have humans in the loop.”

Only that way, Scarpino explained, can we ensure that human societal values, norms, and laws are recognized and enforced by AI systems. This logic is not limited to the legal system. Healthcare systems, too, depend on vast amounts of modeled data, much of it carrying similarly unequal or biased inputs that can impact care at the patient level.

“These same problems that we're learning about are going to translate directly in a negative way from the mass incarceration, criminal legal space into the healthcare space, the life sciences, education—everything that AI touches.”

At EAI, researchers like Sam Scarpino are developing a mutually reinforcing approach to artificial intelligence. By placing a human “in the loop,” they claim organizations stand the best chance of harnessing the best of both worlds—the cutting-edge predictive powers of AI with the common sense values of human operators. As more and more organizations adopt AI, the need to ground applications in a responsible framework becomes both a reputation and ethical imperative.

Learn more about Scarpino’s research.