The Weekly Brief — Issue 04

Hello,

This week, an Ontario risk-scoring system used in provincial jails shows what can happen when an algorithm influences high-stakes decisions without much public visibility. At the same time, Ottawa is asking what AI systems should have to explain.

Plus: Clio’s legal AI arrives in Canada, researchers test an “AI CEO,” and Indigenous technologists at Mila offer a different model for building with data.

Here’s the week.

What happened

Ontario has used the Security Assessment for Evaluating Risk, known as SAFER, in provincial jails since 2021. The automated scoring system uses information such as arrests, charges and disciplinary records to recommend whether a prisoner should be placed in minimum, medium or maximum security. Staff can override the recommendation, but little has been publicly disclosed about the system’s precise inputs, methodology, testing or accuracy.

Government data analyzed by University of Toronto criminologist Scot Wortley and reported by The Breach found that Black people represented 5.4 per cent of Ontario’s population but nearly 27 per cent of prisoners housed in maximum security between 2022 and 2025. White people represented 63.3 per cent of the population and approximately 41 per cent of prisoners in maximum security. Those figures demonstrate substantial overrepresentation, but do not by themselves establish how much of the disparity was caused by SAFER rather than other parts of the justice system.

A proposed class action alleges that SAFER assigns Black inmates to medium- and maximum-security classifications at disproportionately high rates. The allegations have not been proven in court. Materials filed for the certification motion include an expert analysis finding substantial racial differences in SAFER classifications, while noting that additional data would be needed to determine what causes those differences.

Documents reviewed by The Breach also described additional review measures for some Indigenous prisoners. The reporting found no comparable measure for Black prisoners in the documents examined.

Why it matters

A security classification affects where someone is housed, how freely they can move, access to visits, activities and programming, and other conditions of incarceration.

This is an automated scoring system helping inform one of the more consequential decisions the state can make about a person.

A human override does not by itself make a system transparent or fair. Accountability also requires evidence that the system was properly tested, that its effects are monitored across different groups, and that people can understand and challenge the decisions affecting them.

Ottawa’s new AI-transparency consultation is asking related questions: when people should be told that an AI system is involved, what understandable information should be available about its capabilities and limitations, and how serious incidents should be tracked. SAFER is a risk-scoring algorithm rather than a generative AI tool, but it illustrates why those transparency principles matter whenever automated systems influence high-stakes decisions.

Your move

When a public body uses automated scoring, ask five questions:

What information goes into the score? How was the system tested across different groups? How often do staff override it? Can the affected person challenge the result? Are outcomes and independent evaluations published?

The existence of human review is only one part of the answer.

Clio brings its legal AI workspace to Canada

Burnaby-based Clio has launched Clio Work in Canada. The system can analyze matter files, research Canadian law, test arguments and draft documents, with results linked to supporting legal authorities. The Canadian version draws on more than 470,000 cases across more than 40 courts, acquired through Clio’s purchase of Toronto legal-data company Jurisage. Clio says lawyers remain responsible for reviewing its outputs and exercising professional judgment.

Why it matters: In legal work, a polished answer is not enough. The useful feature is traceability: lawyers need to see the authority supporting a claim and verify it before relying on it. This is a practical example of AI transparency becoming a product requirement—not just a policy aspiration.

Skyfall AI plans to put an AI system in charge of a small company

Skyfall AI was founded by Sam Pasupalak, Sumit Pasupalak and Kaheer Suleman. Sam Pasupalak and Suleman previously co-founded Maluuba, which Microsoft acquired in 2017. Skyfall plans to buy a small software or e-commerce business for up to US$1 million and let its AI coordinate areas such as pricing, marketing, customer support, finance and operations. Its stated goal is to double the company’s revenue within six months while documenting the experiment publicly.

Why it matters: The six-month target is a company claim, not evidence that the system works. The more important shift is in ambition: from AI helping people complete individual tasks to AI coordinating decisions across an organization. Watch the evidence—and what happens to the people involved—not just the headline.

Indigenous technologists at Mila are designing AI around data sovereignty

At Mila’s Indigenous AI Gathering, co-hosted with Abundant Intelligences, participants in the seven-week Indigenous Pathfinders in AI program presented tools designed to operate locally and keep community data within community networks. Projects included Landlens, which examines the cumulative effects of land development; SAI Cheese, a preventive dental-care application; and Portage, an education platform built around cultural knowledge.

Why it matters: This approaches AI governance before a system is deployed. Rather than collecting data first and adding safeguards later, the model begins with community control, consent and a defined purpose. It offers a concrete alternative to the extract-first approach common in technology development.

It helps to think of transparency in three levels.

Disclosure tells you that AI was involved.

Explanation tells you what role the system played, what it considered and what it can and cannot reliably do.

Traceability creates a record of the data, sources, instructions and actions involved, allowing mistakes to be investigated and decisions to be challenged.

An “AI-generated” label may satisfy the first level.

For decisions involving employment, money, healthcare, education, government services or legal rights, Canadians will often need all three.

For anything consequential, see whether you can answer:

What information can it access?

What sources support its answer?

What actions can it take without asking me first?

Who is responsible when it is wrong?

When a product cannot answer these questions clearly, that is not merely a documentation gap. It is part of the risk.

A calm, practical briefing on the AI developments that matter in Canada—delivered every Tuesday.

That's the week. See you next Tuesday.

— Padge T.

AI Brief Canada is an editorially independent publication published by Groundshift Advisory Inc. and written by Padge T.

Groundshift Advisory helps Canadian organizations adopt AI responsibly. Sponsored, affiliated, or commercially connected content will always be clearly disclosed.

Issue disclosure: Nothing in this issue is sponsored, affiliated, paid, or commercially connected.

AI use: AI tools supported research, source discovery, outlining and drafting for this issue. All material was reviewed, edited and fact-checked by Padge T. before publication. Read our AI Use Disclosure.

Not affiliated with the Government of Canada.

Corrections or feedback: We welcome corrections and reader feedback. Email [email protected] or review our Editorial Policy.