The Verdict Written in Code

On a cold morning in February 2013, Eric Loomis sat in a courtroom in Waukesha County, Wisconsin, and waited to learn his fate. He was twenty-four years old, unemployed, and had been arrested for driving a car that had been involved in a shooting. The charges themselves were relatively minor—operating a vehicle without the owner's consent, fleeing from police, and reckless endangerment—but the consequences that followed would reshape the landscape of American jurisprudence in ways that no one in that courtroom could have anticipated. When Judge John Daley pronounced sentence, he cited a factor that had never before appeared in an American sentencing hearing: a numerical risk score generated by a proprietary computer algorithm. "I'm not exactly sure how it's weighed," the judge admitted, "but I think it's safe to say that based on the risk assessment, that the defendant is a high risk to the community." Eric Loomis received six years in prison, the maximum allowed under the plea agreement. He would spend the next several years fighting to understand what that number meant and how it had come to determine the trajectory of his life. He would become, in the process, the unlikely center of a legal revolution that would force judges, lawyers, scholars, and citizens to confront a question that the framers of the Constitution never imagined: What happens to justice when the scales are weighed by algorithms?

The algorithm in question was called COMPAS—Correctional Offender Management Profiling for Alternative Sanctions—and it had been developed by Northpointe, a company based in Florida that specialized in providing risk assessment tools to the criminal justice system. COMPAS operated by analyzing answers to a questionnaire that inmates completed shortly after their arrest, asking about their education, employment history, family circumstances, substance use, and prior encounters with the law. The system then generated two scores: one measuring the likelihood that the defendant would reoffend, and another measuring the likelihood of violent reoffending. These scores placed individuals into categories—low, medium, or high risk—and these categories, in turn, influenced decisions throughout the criminal justice system, from bail hearings to sentencing to parole determinations. For Eric Loomis, COMPAS had classified him as high risk, a designation that would haunt his legal battles for years to come.

The Loomis case would eventually reach the Wisconsin Supreme Court, which in 2016 became the first high court in the United States to consider whether the use of algorithmic risk scores in sentencing violated defendants' rights to due process. The court's ruling was nuanced but ultimately permissive: while it upheld the constitutionality of using COMPAS in sentencing, it imposed certain requirements for disclosure and transparency that would have far-reaching implications. The court determined that defendants had a right to know the general nature of the factors being considered, even if the specific proprietary formula remained protected as a trade secret. More significantly, the court acknowledged that the use of algorithmic predictions in criminal sentencing raised profound questions about fairness, accuracy, and the very meaning of justice in an age of machine learning. The Loomis case, the court wrote, represented "a relatively new area of jurisprudence" that would require continued scrutiny as technology evolved. What the court could not have foreseen was just how rapidly that evolution would proceed, or how deeply algorithmic governance would come to penetrate every level of American criminal justice.

To understand the true significance of the Loomis case, we must travel back nearly two centuries, to an era when the idea of predicting human behavior mathematically was itself a radical proposition. The story of actuarial justice begins not with computers or algorithms, but with the actuarial tables developed by the insurance industry in the nineteenth century. Insurers, seeking to calculate risk and set premiums for life, health, and property insurance, discovered that certain demographic characteristics—age, occupation, gender, geographic location—could be used to predict the likelihood of future events. These statistical regularities, they reasoned, could be quantified and applied to individuals based on their membership in particular groups. The mathematics of probability, originally developed for games of chance, was being repurposed to govern the vagaries of human existence. And once the principle had been established that the future could be predicted from the past, that individual fate could be reduced to statistical aggregates, the logic would eventually be extended to the most consequential decisions of all: who should be freed, who should be confined, and for how long.

The convergence of actuarial science with criminal justice found its most disturbing expression in the eugenics movement that swept through America and Europe in the late nineteenth and early twentieth centuries. Eugenics—the pseudoscientific belief that human breeding could be directed to improve the species—provided a veneer of scientific legitimacy to centuries of prejudice and discrimination. Proponents of eugenics argued that criminality, poverty, feeblemindedness, and other social problems were hereditary traits that could be identified through careful observation and measurement. Once identified, the logical response was to prevent the "unfit" from reproducing: through forced sterilization, immigration restriction, and institutionalization. The famous Buck v. Bell Supreme Court decision of 1927, in which Justice Oliver Wendell Holmes infamously declared that "three generations of imbeciles are enough" to justify forced sterilization, represented the legal apotheosis of this movement. The machinery of state, guided by the pseudo-science of the day, had been deployed to predict and prevent future "defectives" from ever coming into existence. That eight generations would later be sterilized under North Carolina's eugenics program—many of them poor, many of them Black, nearly all of them innocent of any crime—was a testament to the dangers of conflating prediction with causation, correlation with destiny.

It would be a mistake, however, to view the history of actuarial justice solely through the lens of its most egregious abuses. Alongside the eugenics movement, more nuanced traditions of statistical reasoning were developing that would eventually inform modern risk assessment. The work of Adolphe Quetelet, the Belgian mathematician who in the 1830s pioneered the use of social statistics to understand patterns of crime and behavior, represented a more sophisticated attempt to use numbers to understand human affairs. Quetelet famously observed that the number of crimes committed each year remained remarkably stable across European nations, a regularity that suggested social forces beyond individual choice were at work. His concept of the "average man"—l'homme moyen—proposed that statistical regularities could reveal underlying laws of human behavior that transcended individual idiosyncrasies. Whether Quetelet's ideas were used to justify oppression or to argue for social reform depended largely on the political commitments of those who invoked his name, but his fundamental insight—that human behavior could be studied scientifically and that aggregate patterns might inform policy—remained influential.

The modern era of algorithmic risk assessment in criminal justice traces its roots to a pivotal experiment conducted in 1928 at Sing Sing prison in New York. A young psychiatrist named Walter R. Packard devised a system for predicting which prisoners would commit crimes after their release, based on answers to a questionnaire covering their personal history, family background, and psychological characteristics. Packard's system achieved a remarkable accuracy rate of 80 percent—a