You’ve probably seen the headline by now. Ontario jails are using an AI system called SAFER to decide how harshly prisoners should be treated — and it’s handing Black and Indigenous inmates worse conditions. Shocking, right? Everyone’s pointing fingers at the algorithm. “AI is racist,” they say. “Technology has gone too far.”
But here’s the thing that should make your blood run cold: the Ontario Ministry of Community Safety and Correctional Services already knew this would happen. Before SAFER ever processed a single prisoner, internal training documents — viewed exclusively by The Breach — explicitly acknowledged that “Indigenous and racialized individuals face systemic discrimination in our justice system.” The ministry’s own words. Then, in the next breath, the document admits that “assessments like SAFER would likely contribute to the over-representation of Indigenous and racialized individuals in higher security classifications.”
They didn’t discover the bias after deployment. They predicted it, documented it, and then deployed the tool anyway.
That’s not a bug. That’s not an accident. That’s a choice.
We keep having this exhausted national conversation about whether AI is “safe” or “ethical” — as if the technology woke up one morning and decided to be racist. But SAFER didn’t invent anti-Black bias in Canadian prisons. It inherited it. The system trains on historical data — arrest records, sentencing patterns, institutional infractions — all of which are soaked in decades of racial discrimination. When you feed biased history into a machine, you don’t get objectivity. You get bias with a mathematical stamp of approval.
An algorithm doesn’t eliminate human prejudice. It launders it.
Here’s how the laundering works. Before AI, when a prison guard or case worker decided to throw someone into maximum security, that person could be questioned. They could be challenged. They had a face, a name, and a supervisor. There was at least the theoretical possibility of accountability. But when SAFER spits out a risk score, suddenly the decision feels inevitable, scientific, neutral. Who do you argue with? The math? The model? The invisible chain of training data that nobody can fully explain?
That’s the genius of the scheme — and yes, “scheme” is the right word. The government has effectively outsourced racial discrimination to a black box. When a Black prisoner gets assigned to harsher living conditions, the ministry can shrug and say, “The algorithm made the call.” When Indigenous inmates are over-represented in maximum security, officials can point to the data and insist it’s not personal. It’s just what the numbers say.
But numbers have never been neutral. Every dataset is a mirror, and this one reflects centuries of colonial violence and anti-Black policing dressed up as objective risk assessment.
Think about what this means in practice. We’re not talking about an AI that recommends movies or sorts emails. SAFER determines whether a human being spends their days in a cell with natural light or in segregation. It decides whether someone gets access to programs, family visits, or the conditions that might actually prepare them for release. For the people inside Ontario’s jails — many of whom are pretrial detainees who haven’t even been convicted of a crime — this algorithm is the difference between hope and despair.
And the government knew. Let that sink in. Not “found out later.” Not “was surprised by independent research.” They wrote it down in their own training materials, in their own words, before the system went live. They identified the harm, described the mechanism, predicted the outcome — and then pressed go.
If a person did that, we’d call it reckless endangerment. When a government does it, we call it modernization.
This is the pattern we need to stop accepting. Across North America, public institutions are rushing to adopt AI tools for decisions that shape human lives — child welfare, predictive policing, bail determinations, now prison classifications. And every time these tools reproduce the exact same racial disparities that already existed, officials act shocked. They promise to “review” the algorithm. They convene ethics boards. They issue statements about their “commitment to fairness.”
But the Ontario case reveals the playbook for what it is. The review happened before deployment. The ethics assessment was done. The findings were clear. The outcome was predictable. And the decision was made anyway — because the people making it weren’t the ones who’d be sleeping in a maximum-security cell.
AI doesn’t have a racism problem. Institutions have a racism problem, and AI gives them the perfect cover to keep practicing it without consequence.
So no, the story here isn’t that “AI assigned Black prisoners harsher conditions.” The story is that a government knowingly operationalized a tool of racial discrimination, documented its own complicity, and bet that nobody would care enough to stop them.
The question is whether they’re right.
FAQ
Q: Isn't this just a case of AI being imperfect? Surely the government will fix it.
A: No. The government didn't stumble into this — they documented the racial bias in their own training materials before SAFER was deployed and proceeded anyway. This isn't a flaw to be patched. It's a feature they chose to ship.
Q: What does this mean for people in Ontario's prisons right now?
A: Real human beings — many pretrial detainees who haven't been convicted of anything — are being assigned to harsher security conditions based on an algorithm that the government admits reproduces racial discrimination. This affects their living conditions, access to programs, family visits, and ultimately their chances of release.
Q: If the data is biased, can't we just use better data?
A: That's the trap. There is no 'clean' dataset of Canadian criminal justice history. Every arrest record, sentencing pattern, and infraction log is shaped by decades of colonial and anti-Black bias. You can't train an unbiased model on biased history. The real fix isn't better data — it's removing AI from decisions that control human lives until institutions have earned the right to use it.