AI for Reentry and Criminal Justice Organizations

Successful reentry requires coordinated support across multiple domains, and the research on what actually reduces recidivism is clear. Education dramatically changes outcomes: according to the Prison Policy Initiative, someone who completes high school while incarcerated reduces their recidivism likelihood from roughly 70-80% to around 50%. An associate degree drops it to 13.7%, a bachelor's degree to 5.6%. Employment is similarly powerful: formerly incarcerated people face unemployment rates over 27%, and stable employment after release is one of the strongest predictors of successful reintegration.

This is where AI can genuinely help. The challenge for reentry organizations isn't usually a lack of knowledge about what works. It's capacity: case managers carrying large caseloads who struggle to provide individualized attention to everyone who needs it. AI tools that help identify who needs immediate attention, that match individuals to jobs with genuine precision, or that track multiple dimensions of progress simultaneously can meaningfully extend what an organization can accomplish.

The critical distinction is between AI as decision support and AI as decision-making. In reentry work, this distinction isn't merely technical. It determines whether technology serves the person being helped, or becomes another system that processes them. Every use of AI in this sector should be designed around augmenting human judgment, not replacing it.

Algorithmic risk assessment tools have been used in the criminal justice system for decades, and their record is deeply mixed. These tools attempt to predict the likelihood that an individual will be rearrested, reconvicted, or otherwise not successfully reintegrate, using factors like prior arrest history, employment status, residential stability, and substance use patterns. In principle, such tools could help organizations prioritize scarce intensive services for those who need them most.

In practice, the most widely used tool, COMPAS (Correctional Offender Management Profiling for Alternative Sanctions), has been extensively documented to produce racially disparate results. An investigation by ProPublica found that Black defendants were significantly more likely to be incorrectly classified as higher risk for future crime, while white defendants were more likely to be incorrectly classified as lower risk. These disparities arise partly because the tools use historical criminal justice data, which itself reflects decades of racially unequal policing and sentencing.

This is what researchers call the "ratchet effect": algorithms trained on biased historical data encode those biases into future predictions, which then generate new biased data, perpetuating a cycle. For nonprofit organizations, this creates a profound ethical tension. Using an algorithm that produces racially disparate risk scores to allocate services or supervision intensity means the organization itself becomes a mechanism for amplifying existing injustice, regardless of its intentions.

This doesn't mean risk assessment tools have no place in reentry work. It means their use requires extraordinary care, robust human oversight, and a commitment to ongoing equity auditing. A risk score should function as one input into a multifaceted professional assessment, never as a determining factor. Case managers need to understand how the tool works, what its documented limitations are, and how to override its recommendations when their professional judgment suggests it is wrong.

Reentry organizations occupy a unique ethical position in the AI conversation. The people they serve have often already been subject to the harms of algorithmic decision-making in the criminal justice system. They have experienced firsthand how data-driven systems can compound disadvantage, produce unjust outcomes, and remove human judgment from decisions that profoundly affect people's lives. Introducing AI into the services meant to help them recover requires extraordinary intentionality about ensuring these tools function differently.

The most important principle is consistent: people, not algorithms, make decisions. This means every consequential decision about a client's case plan, program placement, service intensity, or referrals must be made by a human case manager who has reviewed the full picture of that individual's circumstances and goals. AI outputs are inputs to that decision-making process, not substitutes for it.

Centering the voices of formerly incarcerated people in decisions about AI adoption is not merely good practice; it's ethically required. Organizations that deploy AI tools affecting their clients without meaningful input from those clients are repeating a pattern of top-down decision-making that has historically failed justice-involved communities. This means creating genuine feedback mechanisms, not token consultations, and being willing to abandon tools that the community finds harmful regardless of their apparent efficiency gains.