Ontario’s jail risk system shows a Black-white gap its data cannot explain

• Among SAFER records with race reported, 39.8% of assessments for Black prisoners resulted in maximum security, compared with 23.2% for white prisoners.
• The 147,089-record dataset cannot identify unique people or explain the gap; race was missing in roughly three in 10 records, and some reported levels include managerial overrides.
• Ministry training predicted that unequal justice-system inputs could reproduce disadvantage, but the tailored scoring and override rules it initially described focused on Indigenous prisoners and women.

Ontario’s automated jail risk system produced a large and persistent Black-white classification gap in provincial records from September 2022 through May 2025. The same court record that reveals the disparity also shows why it cannot yet be treated as proof that the algorithm itself caused harsher confinement.

That distinction matters because a SAFER score is only the first link in a chain. A manager can accept or override the recommendation, and operational limits can determine whether the resulting security level changes where or how a person is held.

The outcome gap is clear; its cause is not

SAFER—Security Assessment for Evaluating Risk—uses information in Ontario’s Offender Tracking Information System to estimate a prisoner’s risk of violent or frequent institutional misconduct. It recommends minimum, medium or maximum security and is automatically rerun at scheduled points during a person’s custody.

University of Toronto criminologist Scot Wortley analyzed provincial data filed in a proposed class action. He was retained by the plaintiff’s lawyers, and the supplied certification record does not include Ontario’s response to his report.

The data contain 147,089 assessment records from September 2022 through May 2025. Among records with race reported, the distribution was:

The maximum-security share for Black records was about 1.7 times the white share. Medium and maximum results together accounted for 82.4% of Black records and 63.4% of white records. In the latest period, February 2024 through May 2025, the maximum shares were 40.9% and 23.8%, respectively.

The comparison with Indigenous records is important. Ontario used higher cutoffs and a specific downward-override process for Indigenous prisoners; their maximum-security share in the data was slightly below the white share. But these outcome tables do not establish that the safeguards caused that result, because the underlying risk factors can differ across groups.

Nor are the records a count of prisoners. SAFER reassesses people after 30, 60, 90, 180 and 365 days and then every six months for men, with a similar schedule for women. The data do not include an identifier that would allow Wortley to count unique people or follow one person’s progression. The proportion of records missing race rose from 28.2% in the first period to 32.3% in the last. His analysis assumes that the likelihood of reporting race did not differ in a way that would materially change the comparison; the available data cannot test that assumption.

Wortley also found a separate ministry dashboard with far fewer records and totals that did not consistently reconcile. It showed the same direction of disparity, but its inconsistencies weaken any claim to precision.

Most decisively, the supplied data do not contain the full inputs needed to test why classifications differed. Wortley identified charges, criminal history, institutional misconduct, escape history and alleged security-threat-group affiliation among the missing explanatory variables. The observed gap is not a validation study and does not establish discriminatory intent or causation.

Ontario predicted that unequal inputs could travel into SAFER

The policy problem does not rest on the outcome data alone. Ministry training obtained through freedom-of-information requests said Indigenous and racialized people faced systemic discrimination and were more likely to be charged, remanded and convicted. It warned that assessments such as SAFER could carry that disadvantage into maximum-security results.

Ontario’s initial mitigation strategy adjusted score cutoffs for Indigenous prisoners and women and created an Indigenous-specific override. The training said comparable supports for other racialized groups would take time and named Black prisoners directly:

“We will evaluate SAFER before making similar adjustments for other groups.”

An investigation drawing on those records made the delayed protection for Black prisoners central to its account of the rollout. The documents themselves, reproduced in the plaintiff’s motion record, support the narrower conclusion: officials anticipated unequal upstream inputs, chose group-specific mitigations for Indigenous people and women, and said evaluation would precede similar adjustments for others.

Ricardo Newell’s statement of claim alleges that applying SAFER without comparable safeguards for Black prisoners violated equality rights under the Charter. Those are allegations. The action has not been certified, and no court has found Ontario liable.

Managerial overrides complicate both sides of the argument. Training allowed a level to be raised or lowered when relevant information was not in the database, including positive recent behaviour, alleged gang ties, victimization of others or a current management issue. Written reasons were required. Of 748 overrides in one dataset covering September 2022 through April 2023, 36% involved Black records and 39% involved Indigenous records. The data do not say whether those overrides moved people up or down, so they cannot show whether discretion reduced the disparity or widened it.

A recorded security level is not the same as a cell placement

Ontario’s 2024 policy says managers must accept or override an initial recommendation and should house people by security level “where possible.” Its reference chart associates maximum security with limited movement, fewer on-unit programs and less access to work and amenities, while lower levels can receive more autonomy and incentives.

But Ontario’s lawyer made an important qualification when producing the latest data. The monthly records reflect a person’s security level at the time of the data pull; where a level was overridden, they show the override rather than the original recommendation. The letter also says the data do not show actual housing or management. Human-rights needs and operational feasibility can supersede placement by score.

The plaintiff’s own evidence illustrates the separation. Newell swore that he was placed in maximum security throughout a 2023 detention, while SAFER reports attached to his affidavit show an initial medium result followed by later maximum results. That does not resolve which conditions he experienced; it shows why assessment records cannot be used as a direct proxy for day-to-day placement.

The operating system around SAFER further limits what classification can accomplish. A chief coroner’s expert panel described SAFER as a potentially useful alternative to Ontario’s historical default to maximum security. It also found that the province’s facilities were overwhelmingly built as maximum security and that staffing shortages, lockdowns and poor record quality could frustrate more tailored custody.

At Toronto East Detention Centre in 2022, prisoners and advocates said reviews were delayed and programs needed to lower scores were scarce. A correctional union president defended the pilot as fair and effective, saying it let staff separate different risk levels and move lower-risk people out of maximum conditions. That contemporaneous account captured the unresolved operational tradeoff: classification may help allocate scarce staff and space, but only if reassessments, programs and lower-security placements exist in practice.

Oversight identified data and transparency failures after rollout began

Development of SAFER began in 2019, and the first implementation occurred at two Thunder Bay facilities in March 2021. A formal province-wide policy was not released until August 2024.

The lag is visible in an access-to-information dispute over records from January 2021 through July 2022. In a 2025 interim order, Ontario’s information and privacy adjudicator accepted evidence that final versions of the requested policy manual and data-definition guide did not exist during that period. The order did not require release of the draft policy materials: it found them exempt as policy advice but required the ministry to reconsider whether it should nevertheless disclose them. It also rejected the ministry’s attempt to treat a training manual as an excluded employment record and ordered a new access decision on that manual.

Data quality became a separate concern in a 2024 coroner’s inquest. Witnesses said misconduct outcomes and withdrawn or reduced charges were not always updated in OTIS, the system SAFER reads. The jury recommended scheduled audits, sufficient staffing to maintain records, written classification results for prisoners and an independent third-party evaluation after implementation. The inquest did not find SAFER contributed to Marco Michaud’s 2018 death; the tool had not yet been implemented at his Ottawa jail when the inquest was held. Its recommendations addressed the reliability and governance of the system being rolled out.

The Ontario Ombudsman’s latest annual report adds an operational warning. Prisoners complained that scores did not change after programming and that it was unclear which programs counted. The office also received concerns about disproportionate effects on Black and Indigenous prisoners and said the ministry was reviewing implementation. Complaints are not a prevalence measure or proof that the model failed, but they identify the same gap between formal reassessment and an explainable route to a lower level.

The retained record does not supply a peer-reviewed Ontario validation broken down by race. The archived page for a 2019 article in The Prison Journal identifies a study on developing and validating a system that predicts severe and frequent prison misconduct, but its available body does not provide methods or results that can be verified here. It therefore cannot establish SAFER’s performance in Ontario.

The next test is an auditable link from score to conditions

The certification motion is scheduled for April 28, 2027. That hearing will decide whether the proposed class action may proceed collectively, not whether the plaintiffs have proved discrimination.

Ontario’s policy review can address the central evidence gap sooner. A credible evaluation would need to publish, at minimum:

• performance, calibration and error rates on the Ontario population, reported by race and gender with denominators and time windows;
• the completeness and correction history of the charges, misconducts, program records and other inputs used in each assessment;
• the generated recommendation, any override and its direction, the final recorded security level, and the person’s actual housing and conditions; and
• whether access to programs and the effect of completing them differ by institution or security level.

Without those links, the record supports two conclusions and no more: Black assessment records receive maximum-security results far more often than white records, and Ontario knew unequal justice-system data could reproduce disadvantage. It does not yet show which inputs, model rules, managerial decisions or infrastructure constraints created the gap—or how often the gap translated into harsher confinement. Those are the facts needed to judge SAFER as a safety system rather than merely a scoring system.