Ontario jails have quietly deployed an artificial intelligence system since 2021 to sort prisoners into security levels. The tool, known as SAFER, promises to forecast misconduct. But a class-action lawsuit and fresh reporting show it funnels Black inmates into harsher maximum-security conditions far more often than others. The pattern revives old questions about algorithms trained on flawed data.
The Security Assessment for Evaluating Risk program feeds in details like past arrests, charges and jail discipline records. It spits out a score from 0 to 100. That number decides whether someone lands in minimum, medium or maximum security. Higher classifications bring tighter cells, fewer visits, limited programs and reduced movement. Most Ontario prisoners sit on remand, legally innocent while awaiting trial. Yet the system treats them with the weight of full criminal histories.
Breach Media uncovered the disparities through government data analyzed by University of Toronto criminologist Scot Wortley. Black Ontarians form just 5.4 percent of the population. They account for nearly 27 percent of those held in maximum security between 2022 and 2025. White residents, at 63.3 percent of the province, make up only about 41 percent of maximum-security prisoners. Black women fared even worse. Over a 16-month stretch ending in 2025, more than twice the proportion of Black women received maximum-security scores compared with white women.
The numbers hit hard. And they echo patterns seen south of the border. A landmark 2016 investigation by ProPublica dissected Northpointe’s COMPAS recidivism tool. Black defendants who did not reoffend were nearly twice as likely as white counterparts to be flagged as high risk, 45 percent versus 23 percent. White defendants who did reoffend were mislabeled low risk almost twice as often. The same data problems surface in Canada. Arrest and charging records reflect documented anti-Black bias in policing and courts. Plug that information into an algorithm. The output hardens into institutional fact.
Ontario’s Ministry of the Solicitor General knew this risk. Internal training documents reviewed by Breach Media state that “Indigenous and racialized individuals face systemic discrimination in our justice system. As a result, assessments like SAFER would likely contribute to the overrepresentation of Indigenous inmates in maximum security.” The province responded with safeguards for Indigenous prisoners. Staff must review low-end maximum scores for them and can override downward. Native Inmate Liaison Officers factor in colonialism’s effects. No equivalent steps were taken for Black inmates. “We are continuously evaluating to determine if it is necessary to make similar adjustments for other groups,” the documents note. Human rights lawyer Nana Yanful called the choice telling. “They are knowingly contributing to a program that will further exacerbate the conditions of confinement for Black folks. They just don’t care that this is going to impact them more negatively than non-Black incarcerated people.”
The class action, filed in February 2025 by Koskie Minsky LLP, targets these gaps. It alleges violations of Charter rights to equal protection and benefit under the law. Lawyers argue Ontario could have copied its Indigenous mitigations for Black prisoners but chose not to. Caitlin Leach, a Koskie Minsky attorney on the case, points to upstream bias. “A police officer’s decision to charge someone with a particular offense that becomes part of their criminal record. It might be a Black person who’s charged with an offense, and a white person in the same circumstances might not be charged with that offense.” The lawsuit focuses on outcomes. It does not need to crack open the algorithm’s code.
Details about SAFER remain sparse. The ministry describes it as an automated predictive tool built by external researcher Dr. Grant Duwe of the Minnesota Department of Corrections. He analyzed 10 years of historical Ontario prisoner data to spot factors tied to violent or frequent misconduct. The ministry claims the tool predicts such incidents with equal accuracy across racial groups. Neither the ministry nor Duwe responded to Breach Media’s repeated requests for evidence or methodological explanation. An Ontario Ombudsman report tallied 126 prisoner complaints about SAFER. Some said staff promised score reductions through programming but offered no clear guidance on which programs counted. The report also flagged “concerns about the disproportionate impact of this tool on Black and Indigenous inmates.” A ministry review is underway. The program, however, already operates in most provincial jails.
This story fits a larger pattern. The Law Commission of Ontario released a detailed paper in April 2025 examining AI risk tools for bail, sentencing and recidivism predictions. It warns that historical data often embeds past discrimination. “Feeding biased data into AI tools will inevitably result in the tools generating biased outputs,” the report states. It highlights COMPAS as a cautionary tale of algorithmic racism. Canadian sentencing principles, shaped by Supreme Court rulings like Ipeelee for Indigenous offenders and Morris for Black ones, require judges to consider and ameliorate systemic racism. Group-based predictions clash with the demand for individualized justice. The paper questions whether such tools can ever square with Charter guarantees of equality and procedural fairness when liberty hangs in the balance.
Overrides exist. Jail staff can adjust SAFER scores. Limited data show Black and Indigenous prisoners make up the bulk of overrides, 36 percent and 39 percent respectively. Yet records do not reveal whether those changes raised or lowered security levels. Yanful expressed skepticism about one common override trigger: suspected gang affiliation. “I’ve had clients before who say, ‘I literally am not part of this gang that the police have alleged I’m part of, I just live in this neighborhood where this group of people is alleged to be operating.’ Discretion is often used against Black folks to their detriment.”
Critics like Lindsay Jennings of the Tracking (In)Justice Project stress the human element remains. Discipline reports, a SAFER input, depend on subjective calls by guards. Mood, familiarity or outside advocacy can sway outcomes. Four out of five Ontario prisoners sit in remand. Many face charges later withdrawn or unproven. SAFER appears to weigh those charges anyway. Ontario Human Rights Commission research on Toronto police from 2013 to 2017 found Black people charged at higher rates yet overrepresented in withdrawn cases and less likely to be convicted than white people.
The broader debate stretches back years. Wisconsin’s use of COMPAS in sentencing drew a 2017 challenge from defendant Eric Loomis. The state’s supreme court allowed the tool but required judges to understand its limits. Canadian courts have shown caution. In Ewert v. Canada, the Supreme Court ruled certain risk tools could not be applied to Indigenous inmates without proven validity for that population. The Law Commission paper urges proactive governance. It calls for transparency mandates, validation standards, and policies that prevent tools from laundering historical bias into future decisions. Without such guardrails, it argues, public trust erodes and inequality grows.
Recent academic work reinforces the concern. A 2025 study in Race and Justice on federal Canadian corrections found significant racial differences in risk scores. Black and Indigenous individuals received higher assessments than white counterparts even after controls. Parole and housing decisions sometimes reflected apparent bias. Another 2025 paper in RSF: The Russell Sage Foundation Journal of the Social Sciences discussed how administrative data used in predictive tools can widen racial disparities rather than shrink them. These findings arrive as more jurisdictions eye AI for efficiency. But efficiency at what cost?
Shauna, whose name was changed for fear of retribution, learned about SAFER while trying to visit her Black incarcerated partner. His score dictated whether visits would be in-person or video only. He requested his score for weeks without success. Stories like hers multiply across facilities. Prisoners report frustration with an opaque system that shapes their daily existence yet offers little recourse to challenge it.
The class action could take years to resolve. In the meantime SAFER continues. A ministry review may yield changes. Advocates push for full disclosure of the model’s inputs, weights and validation studies. They want mitigations extended to Black prisoners and independent audits. Until then the tool stands as a concrete example of how yesterday’s biases can be coded into tomorrow’s decisions. The scores feel objective. The outcomes tell another story.
But the technology’s defenders argue properly designed systems could reduce human prejudice. The evidence so far suggests otherwise. Data reflects society. Society contains deep inequities. Feed one into the other without correction and the inequities harden. Ontario’s experience with SAFER offers a live case study. It deserves close watching, rigorous testing and honest reckoning. Prisoners’ liberty and dignity depend on it.
WebProNews is an iEntry Publication