Algorithmic Classification in Ontario’s Jails

Introduction

By: Lucie Kapustynski

Since 2021, Ontario’s correctional facilities have implemented SAFER (Security Assessment for Evaluating Risk), which the Ontario Chief Coroner’s Expert Panel characterizes as “an innovative tool… that supports staff in anticipating and mitigating improper behaviour” (2023). SAFER is an automated predictive tool that evaluates incarcerated people’s risk of misconduct and, based on criminal-legal information such as prior arrests, disciplinary records, and charges, each prisoner is given a score that ranges from 0- 100 and determines where they will serve their sentence: minimum, medium, or maximum security detention (Cole, 2026). Similar to artificial intelligence (AI) resume screening systems, both analyze historical information and patterns to make predictions about people; however, SAFER uses these predictions to classify incarcerated people according to their perceived risk, which then affects the conditions in which incarcerated people live. As a result, Ontario’s use of the SAFER system brings into question how it operates despite Charter Rights, its effectiveness and reliability, and unenforced legislative requirements.

Racial Inequality in Risk Classification

SAFER relies on data produced by a justice system with documented disparities and reproduces these inequalities by disproportionately assigning higher-risk classifications for Black people. In February 2025, concerns regarding SAFER and the system’s impact on Black inmates led to legal action against the Ontario government. A proposed class action, Newell v. Ontario, was filed on behalf of Black people incarcerated in Ontario correctional institutions who were assigned “medium and maximum-security classifications at significantly higher rates than non-Black inmates” through SAFER (Goldblatt Partners, 2026). On April 28, 2027, the Ontario Superior Court of Justice will hear the proposed class action, and the plaintiffs seek Charter damages; monetary compensation for individuals whose Charter rights are found to be violated.

The consequences of SAFER’s classifications extend beyond the risk score assigned to an incarcerated person and reflect broader consequences of allowing an algorithmic assessment to determine an individual’s daily life, autonomy, and conditions of confinement. Although SAFER may not explicitly classify individuals based on race, its application produces unequal outcomes for Black inmates, raising concerns about racial discrimination through the system’s effects. Higher-risk classification can result in placement in more restrictive facilities, where inmates may experience harsher conditions and greater limitations on their daily lives, and ultimately, violate rights and liberties protected by the Charter.

The class action lawsuit claims that SAFER violates Black prisoners’ Charter rights by denying them equal protection and benefit of the law under section 15(1), which establishes that every individual has the right to equal protection and benefit of the law without discrimination, including discrimination based on race (Government of Canada, 2025). The plaintiffs point to measures that Canada uses for Indigenous inmates and women inmates, such as Security Reclassification Scales, which are research-based tools used in federal corrections to help determine an inmate’s appropriate security level, to show that Ontario could have adopted comparable measures to address SAFER’s disproportionate impact on Black inmates (Cole, 2026), (Canada, 2026). In failing to do so, the class action argues that Ontario allowed racial disparities in security classifications to persist, subjected Black inmates to more restrictive conditions of confinement, limited their autonomy, and violated their equality rights (Cole, 2026).

Historical Data and High-Stake Decisions

The class action reveals systemic issues within Ontario’s correctional system, particularly when algorithmic tools like SAFER are used to make consequential decisions. According to Ministry documents reviewed by The Breach, SAFER was developed using ten years of historical prisoner data to generalize factors associated with violent or frequent misconduct (2026); however, relying on historical data can be problematic when it is treated as a direct indicator of individual risk without consideration of the social circumstances behind it. As Canada’s historical justice-system data may already reflect racial disparities in policing, charging, and disciplinary decisions, this approach to developing SAFER creates a potential feedback loop: if disparities are present in the data used to develop the algorithm, SAFER learns these patterns and reproduces them when assessing inmates. As a result, SAFER may not be the “innovative tool” it is labeled as it risks carrying existing inequalities under the appearance of data-driven neutrality (2023). 

While Ontario’s inmate classification system has relied on historical data to make decisions on an inmate’s placement, it also blended together clinical reports, institutional behaviour, and Level of Service Inventory; a process that is a risk-and-needs assessment tool used to assess inmates based on factors related to their correctional needs (Ontario, 2021). Evidently, SAFER did not introduce the use of historical justice-system information but loses this personally informed-decision making process. A more individualized approach that mirrors Ontario’s Level of Service Inventory and factors in an individual’s current level of risk, may reduce the disparities embedded in past data and prove to be more effective. Ultimately, if Canada’s goal is to practise equitable means of ensuring justice, the priority should be to ensure that no minority groups are excluded from this ideal. Implementing better background systems to better understand systemic injustices will help to uphold Canada’s democracy, as stated by the Charter Rights.

SAFER will survive the lawsuit. Look out for a rebrand

By: Daniel Park

As discussed in the previous section, the SAFER algorithm is facing a class action over racial bias in how it sorts prisoners into security levels. Unfortunately, examples abroad hint that the lawsuit won’t kill SAFER entirely. Three earlier systems faced the same accusation, and in each case the tool continues to run under a new name.

Friends of SAFER

In 2013, a Wisconsin judge told defendant Eric Loomis he’d been “identified, through the COMPAS assessment, as an individual who is at high risk to the community,” then sentenced him to eight and a half years. (COMPAS is Wisconsin’s predecessor of Ontario’s SAFER, for context.) ProPublica’s 2016 investigation found the tool flagged Black defendants as high risk at roughly double the rate of white defendants among people who went on to not reoffend. Surprisingly (or unsurprisingly), two Dartmouth researchers found COMPAS performed no better than untrained volunteers recruited online. Northpointe renamed itself Equivant, started calling itself a company that builds “software for justice,” and kept selling updated versions of the same model underneath. The same story holds in England, with its OASys system. Active since 2001, it currently holds over seven million risk scores, processes more than 1,300 assessments a day, and no outside researcher has ever been allowed to audit it to this day. Its “replacement,” a system called ARNS, is being built by the same Ministry of Justice team, with the same contractor that’s supported OASys all along. Similarly, SAFER was designed by Grant Duwe, who also built Minnesota’s MnSTARR and its 2016 automated version, simply copying and pasting the same biased algorithm from the United States to Canada. Black people make up roughly 5% of Ontario’s population but represented close to 27% of prisoners held in maximum security between 2022 and 2025, a gap that seems to mimic that of ProPublica’s COMPAS from a decade ago.

Finding an Effective Middle-Ground

Ontario’s Ministry of the Solicitor General has announced a plan to review the pilots for potential expansion. An expansion, not a shutdown.

If a full reversal isn’t realistic, it’s more resourceful and effective to advocate for checks and audits, ensuring the reform doesn’t stop at renaming. As impractical it is to entirely halt the AI development cycle for its safety concerns, it’s better to advocate for the right safety and audit mechanisms instead. The technical means to deploy such software audits already exist; a team at the University of Chicago released an open-source auditing tool called Aequitas back in 2018. In fact, the policy exists too: California mandates an audit of pretrial risk algorithms every three years, and New York City’s public agencies operate under an Algorithmic Impact Assessment requirement that forces disclosure of when and how an algorithm gets used. The tooling and legal template are there, but Ontario (and Minnesota and England) is missing the rule to enforce them. Notably, Partnership of AI points out that a model can seem unbiased averaged across a whole population while still failing badly for one group inside it. What this means: in a set of hundreds of numbers, one “17” won’t have a noticeable impact on the average of the set if all other numbers are either “1” or “2.” (The writer of this article has done the calculation and got 1.66 as the mean value.) The audit only works if the subgroups are checked specifically. In this industry, None of the measures mentioned so far has ever taken a risk assessment tool offline. Overall, it could change who gets to assess the tool, making it less of a marketing move and more of an obligatory report. Given the American and British precedents, this is the realistic move the public can and should advocate for, not a complete abandonment.  

Eight years unproclaimed: the law that would have governed SAFER

By: Dhruv Gadgil

The Act That Was Never Proclaimed

In May 2018 the Legislature passed the Correctional Services and Reintegration Act, intended to replace the Ministry of Correctional Services Act as the governing law for provincial jails. Section 41 requires that on admission a superintendent “assess the inmate using an evidence-based security classification tool” before assigning minimum, medium or maximum security. Subsection 41(2) provides that the superintendent “shall not discriminate against any inmate because of any ground of discrimination prohibited by the Human Rights Code” in making that assignment, and s. 41(4) requires written reasons for every classification and reclassification. Section 29 separately obliges anyone administering the Act to consider systemic and individual circumstances when making a decision that limits the liberty of a First Nations, Inuit or Métis person.

The Act received Royal Assent on 7 May 2018 and has never been proclaimed into force. Ontario’s own ministry filings still list it as not proclaimed (Ministry of the Solicitor General, 2025). Provincial jails continue to operate under the Ministry of Correctional Services Act and Regulation 778, which say nothing about how a classification tool must be built, validated, or explained to the person it classifies. SAFER entered that silence in 2021.

What Federal Law Requires

The absence matters because federal law already contains a duty of this kind. In Ewert v Canada (2018), the Supreme Court held by a majority of seven to two that the Correctional Service of Canada had breached s. 24(1) of the Corrections and Conditional Release Act, which requires the Service to take all reasonable steps to ensure that information it uses about an offender is as accurate as possible. CSC had relied for years on actuarial risk instruments never validated for Indigenous offenders despite knowing the concern. The Court dismissed Ewert’s Charter arguments and granted relief on the statute alone, declaring that CSC must determine whether its tools are subject to cross-cultural variance before continuing to rely on them. That statute governs federal custody only. Ontario has no proclaimed equivalent, which is why Newell v. Ontario proceeds under s. 15 of the Charter rather than as a claim that the province failed a duty to validate.

Two other routes offer little at present. The Enhancing Digital Security and Trust Act, enacted through Bill 194 and identified in this journal as Ontario’s response to the halting of federal AI legislation, imposes requirements on public sector use of artificial intelligence only in circumstances prescribed by regulation. The federal Directive on Automated Decision-Making, which mandates algorithmic impact assessments, binds federal institutions and does not extend to provincial corrections.

The practical consequence appears in the complaints themselves. Ontario’s Ombudsman recorded 126 complaints from inmates about SAFER in 2025-2026, including reports that scores were unfair and could not be changed even after program participation. Section 41(4) would have entitled each of those people to written reasons for their classification, and the common law duty of fairness recognised in Cardinal v Director of Kent Institution may already do some of that work, though no Ontario decision has applied it to an automated classification score.

Why Proclamation Alone Would Not Settle It

Bill 116, an opposition private member’s bill introduced in May 2026, would require the Minister to table a plan with timelines for bringing the 2018 schedules into force. It sits at first reading and is unlikely to advance. The more difficult point is that proclamation alone would not settle the question, because s. 41 operates through regulations that were never drafted and its language predates automated classification entirely. Ontario’s legal framework does not currently ask whether a scoring tool is valid for the people it scores, and the closest thing the province ever wrote would not clearly have asked either.

Conclusion

Altogether, we explore the same failure from three different angles: a tool trained on data containing the predisposed biases of the justice system, track record of similar tools across borders, and a law that was passed but never brought into force. These problems are a reality. SAFER faces a Charter challenge; hundreds of complaints have been logged, and the statute Ontario passed in 2018 to govern inmate classification has never taken effect.

Ontario needs the political drive to proclaim the classification standard it wrote for itself, to write the regulations that standard depends on, and an audit requirement of the kind California and New York City already impose. Without it, SAFER’s future is already laid out along paths Minnesota, Wisconsin and England have been down before: lawsuits and reviews that leave the tool running much as it was.