Criminal Risk Assessment and the Character Trap

Over the past three decades, successive birth cohorts in the United States have come of age in very different worlds of crime and its control. These shifting contexts shape people’s life chances in ways that challenge the belief in stable, individual propensities to commit crime and in timeless rules for predicting risk. Focusing on the life course of different birth cohorts — on when we are rather than who we are — reveals the power of the birth lottery of history.

This matters because common risk-assessment practices pervade the criminal justice system and extend well beyond it. Formal risk instruments are used to inform pretrial release and probation decisions, while criminal history information is used in sentencing, employment screening, tenant screening and occupational licensing. With the emergence of AI tools and large-scale databases, predictive risk assessment is accelerating.

But prognostications like these rest on assumptions of an individual’s stable criminal propensity or character. New research exposes the perils of this approach, revealing how rapidly changing times challenge common notions about prediction and enduring propensities to commit crime.

To isolate social change in life experiences — or when, not who, we are — requires a departure from the usual developmental approach of tracking one group of children born at the same time. To meet this challenge, my research team studied multiple cohorts of over 1,000 Chicago children coming of age over the past three decades.

Consider Darnell Jackson and Andre Lewis (both pseudonyms), one born in the mid‑1990s and the other in the early 1980s. They shared the same race, the same city and the same economic, family and neighborhood backgrounds. Yet children born in the 1980s, like Andre, held different tickets in the birth lottery of history and came of age in a very different social world than those born just over a decade later, like Darnell.

Andre’s cohort witnessed the rise of drug‑related violence, became teens during the murder epidemic of the early 1990s and lived through rising incarceration and aggressive drug policing. By contrast, when children born in the mid-1990s arrived, violence was plummeting and drug arrests in Chicago had fallen by 90% while disorder arrests dropped nearly 100% between 1995 and 2020. Children in Darnell’s cohort were lucky in other ways too. Lead exposure, known to impair child development, declined sharply, and when they turned 25, incarceration had dropped to its lowest point in over 25 years.

These are just two individuals, of course, but they reflect the broader picture. By the time they were 20, children of the mid-1990s were arrested at half the rate of kids born a short time earlier, in the early to mid‑1980s. The younger cohorts were also less likely to witness gun violence and use guns as adults. 

You might think these differences are due to usual suspects like demographic composition, poverty or early life conditions that varied across cohorts. But that’s not the case. Even after I accounted for factors like race, family socioeconomic status, parental criminality and childhood neighborhood disadvantage, large gaps in life trajectories persisted.

Here’s what’s also surprising: The decline in arrest rates between cohorts was strongest for the most disadvantaged — particularly poor Black youth. So among the poorest children especially, things got much better, which is not what we might expect from today’s popular narrative of decline.

These findings challenge the growing risk assessment paradigm that cuts across the political aisle. The Senate’s bipartisan criminal justice reform bill of 2018 called for the development of new risk assessment and prediction tools. Yet in my data, risk assessment instruments trained on kids born in the 1980s using factors like family poverty, first-generation immigrant status, low self-control, male sex and having a single parent overpredicted the arrest probability of children born in the mid‑90s by more than 50%. This large overprediction is due to both changing crime and the shifting relationship of predictive factors with arrest.

History is baked into criminal records, in other words — which, in this case, leads to a systematic gap between actual and predicted arrest patterns across birth cohorts. This bias exists within all racial groups and among those arrested as juveniles, so it is a unique form of bias. It also exists for absolute as well as relative rankings. This matters because even a prediction algorithm that rank-ordered properly — the usual practice — might successfully identify the highest-risk individuals even if they don’t pose a high risk in absolute terms. As in medicine, absolute risk matters most.

The implications for justice are troubling. Inflated risk predictions may lead individuals to be denied bail, sentenced more severely, denied a job or rejected for an apartment. These mechanisms can trigger further criminalization, leading to what I call the “character trap,” a self-reinforcing cycle that labels and punishes individuals based on assumptions about their stable criminal character. This can backfire, severing social connections and blocking opportunities for those criminally labeled, leading them to conform to the very same negative predictions.

And when certain individuals or groups repeatedly commit or are arrested for crimes, they are typically deemed to possess a criminal character. This idea manifests in terms like “chronic offenders” or “superpredators.”

To be clear, mine is not a call to abandon individual responsibility, punishment or even predictive tools. Humans are biased too, of course, so machines are not the enemy. But recognizing the challenge of social change is a first step to pragmatic solutions. 

At a minimum, updates to criminal information systems are needed. To decide who gets released before trial, New York City used risk assessment instruments that were unchanged for nearly two decades. We can imagine how outdated the prediction models are in less-resourced departments. 

Yet updating alone cannot solve the problem. We also need to incorporate into our prediction models the changing way common risk factors forecast future behavior across successive birth cohorts. This will not be easy. We need to invest in the science of studying societal and organizational changes, both to improve predictive performance and to advance knowledge about how the changing world around us influences human behavior.

Perhaps the deepest challenge is cognitive bias in the face of a changing world. Perceptions of past crime waves, with images of superpredators, rampant immigrant criminality and chronic offenders looming large, still drive policy. The Executive Order from the White House in 2025, for example, can be read as unleashing anew the superpredator logic and prison‑filling mandate of prior decades, despite clear evidence of falling crime. That’s fighting the last war on crime.

Despite good news on declines in arrest, incarceration, violence and gun use, we continue to look for the Andres of the world when there are more Darnells. We should focus less on predicting individual criminality or a supposed criminal character and more on cultivating the institutional and environmental conditions that prevent crime and help us meet the unexpected challenges that will come. Put a bit differently, perhaps we need to be thinking more about the morality of societies than the morality of individuals.