Canada's Treasury Board Secretariat published the first public federal AI Register on November 28, 2025. It lists more than 400 AI systems from 42 institutions, ranging from early research projects to tools already supporting operations. The register is the closest official answer to a plain question: which AI uses has the Government of Canada actually reported, and at what stage?
Key takeaways
- The register is an open dataset, labeled a minimum viable product, with a version last modified April 28, 2026.
- It covers 42 institutions out of a federal government with more than 200 departments and agencies, according to an independent analysis.
- That analysis of 409 entries counted 44 percent in development and 39 percent in production.
- About 86 percent of entries were designed for internal government staff, not for the public.
- A listing shows that a system exists. It does not show how well it performs or how often staff overrule it.
What the register is and how it was built
The Treasury Board of Canada Secretariat announced the register in a news release and posted it on the Open Government Portal. Each entry gives a system's purpose, description, intended or current use, and whether it was built in house or by a vendor. The release says the register leaves out AI embedded in low-risk commercial products, such as virtual assistants and spell checkers, and notes that government use of such systems goes back decades, with some dating to 1994.
The dataset page says the data were compiled from Algorithmic Impact Assessments, Access to Information requests, responses to parliamentary questions, Personal Information Banks, and the GC Service Inventory. It warns that entries are incomplete in places and formatting varies. The page lists a CSV in English and French, an open government license for Canadian federal data, and a publication date of November 28, 2025. Its update frequency field reads "Not Planned," while the text says a fuller version will be updated regularly after public consultations in 2026.
What an independent analysis counted
An April 2026 preprint on arXiv analyzed all 409 systems then registered. The arXiv page lists it as accepted at the FAccT 2026 conference, and its counts may differ from the current file, which was modified after the launch. Its reported figures are in the table.
| Measure | Reported share or count |
|---|---|
| Systems analyzed | 409, across 42 institutions |
| In development | 44 percent |
| In production | 39 percent |
| Designed for internal government staff | about 86 percent |
| Built in house | 43.3 percent |
| Built by external vendors | 38.1 percent |
| Open-source based | 6.1 percent |
| Use personal information | 21.4 percent |
| Data practices not specified | 24.4 percent |
The same analysis found Microsoft to be the most frequent named vendor, with 46 systems. It also notes that a handful of bodies, among them the Canada School of Public Service, Innovation, Science and Economic Development Canada, the National Research Council, and Statistics Canada, account for many entries.
Reading the numbers with care
The distinctions below prevent common misreadings of a register like this one.
- A count of systems is not a count of people affected. One tool for staff drafting emails and one used in benefit decisions each count as one row.
- Stage matters. Nearly as many entries are in development as in production, so the total overstates what is running today.
- Coverage is partial. Institutions outside the 42 are missing, and so are uses judged low risk.
- Percentages rest on how entries were filled in. A quarter of entries leaving data practices unspecified limits what can be said about personal information.
What the register does not show
The preprint's authors argue that registers shape what counts as accountable AI instead of merely reflecting it. They report that entries rarely describe risks, limitations, or error rates, how officers review or override outputs, or what training staff receive. The preprint also reports a limited presence of completed Algorithmic Impact Assessments in entries, without giving an exact count. These are the authors' findings on an early version, and the Secretariat has said it will refine the design with public feedback.
A short checklist for reading any government AI register
- Note the version date and the file's modification date.
- Filter by status before counting uses.
- Separate staff-facing tools from public-facing ones.
- Look for who built each system and what data it touches.
- Check whether an impact assessment or human review step is mentioned.
- Compare institution coverage with the total number of institutions.
FAQ
When was Canada's AI Register first published?
November 28, 2025, by the Treasury Board of Canada Secretariat, on the Open Government Portal.
Does the register list every AI tool the federal government uses?
No. It covers 42 institutions and excludes AI inside low-risk commercial products such as spell checkers and virtual assistants.
Can the data be downloaded and reused?
The dataset page lists a CSV under an open government license for Canadian federal data, in English and French.
Does being listed mean a system is approved or safe?
No. The register describes what exists. The sources checked do not present it as an approval or a safety rating.
Related reading
- What narrow AI means, with examples
- AI and robotics across industries
- How data science teams use AI
- A research institute's view of AI in industry
More in AI Basics.
Last verified: October 2026
Sources
- Canada launches first register of AI uses in federal government, Treasury Board of Canada Secretariat, November 28, 2025
- Government of Canada AI Register (Minimum Viable Product), Open Government Portal, modified April 28, 2026
- Analysis of the Government of Canada AI Register, arXiv preprint 2604.15514, April 2026