Health Canada’s Disclosure and Use of AI Raises Questions Around Effectiveness and Accountability

Canada is going all in on AI. Besides the “AI for All” strategy to implement AI throughout public life, the federal government is also piloting and implementing various tools internally. On November 28, 2025, the federal government released a minimum viable product (MVP) version of an AI register documenting how these tools are implemented across different departments.

While there are, at present, 6 tools listed that are being piloted or are already in use at Health Canada, the entries are sparse on details—in one case omitting the name of the vendor—and according to Mark Johnson, a spokesperson at Health Canada and the Public Health Agency of Canada, the descriptions of some of the tools in use are no longer accurate. There are also tools that aren’t listed on the register that are in use at Health Canada, including Copilot, which is subject to US surveillance laws requiring companies to hand over data, even when they’re used by other countries.

Renée Sieber, PhD, an associate professor at McGill University whose research includes government AI adoption, says that the transparency provided by the register isn’t meaningful because there is no accountability or enforcement related to the use of AI tools. Sieber and other experts suggest AI implementation is a labor issue, especially when the government implements tools without evidence of effectiveness as public servants are laid off.

“One should stay away from anyone who says we need AI because of a backlog or because we’re going to save on labor costs,” says Sieber. “Canada needs to hire more civil servants. That’s it.”

What Tools Are Being Used by Health Canada?

Two of the publicly disclosed tools are listed as “in production,” meaning they are currently in use at Health Canada; the remaining 4 are listed as “in development.” Beyond these tools, Johnson writes that GCtranslate, an internally developed AI tool, is in use for translation; Copilot Chat (Basic) is in use among all employees; and M365 Copilot Premium is used by 600 employees at Health Canada and the Public Health Agency of Canada.

One of the tools listed as “in production” by the register, a literature review tool called Covidence, is used to “improve efficiency of conducting scientific literature reviews” by organizing steps in the literature review and getting rid of duplicate citations. Johnson says that, “according to the use case documentation, the current benefit of the tool is a reduction in the time required to review relevant information,” but notes that Health Canada is “not aware of evidence demonstrating that the tool makes for ‘better’ searches and significant time is still needed to screen and review information.” The department is not currently using the tool’s built-in machine learning features.

The other tool that the register lists as “in production” is called the Food Label AI for Monitoring the Food Environment (FLAIME), which “aims to collect publicly available data from the internet relating but not limited to food products, nutrition facts tables and list of ingredients” into databases for use by machine learning algorithms to create data models and generate reports. Though the name of the vendor is not listed in the register, Johnson confirms the tool was developed internally.

Natasha Tusikov, PhD, an associate professor at York University and a visiting fellow with the Justice and Technoscience Lab at the Australian National University, notes the risk of collecting potentially inaccurate information for use in this database, but Johnson says that “The use of AI within FLAIME has evolved since the original GC AI Register entry was published.” Now, it only automates data entry by “extracting data located on pictures of food packaging” using optical character recognition (OCR) algorithms. He emphasizes that “humans remain responsible for the outputs of AI systems,” which may provide a safeguard while also introducing additional labor.

“Are they really saving time by having someone check all the OCR scanning that this tool is now doing?” asks Tusikov.

She also raises concerns with the translation tool being used internally: “This is also why we have language designations, and we have people who have that capability, so we don’t have errors.

Meanwhile, the use of Copilot raises questions about data sovereignty. “Cloud-based technologies that Health Canada uses are subject to departmental security, privacy, and information management requirements,” writes Johnson.

However, Sieber calls these directives “meaningless in an era of AI.” The recent hacking incidents at HuggingFace and the Australian government’s health website, where AI companies tested tools without guardrails, show it’s challenging to insulate these models and keep them away from sensitive information. A 404 Media investigation reported that human contractors regularly read and see images and prompts from users.

Government AI Use Is an Oversight and Labor Issue

Johnson notes that Health Canada “evaluates AI-enabled tools throughout their lifecycle using departmental governance, security, privacy, and business oversight processes.” It does not, however, “routinely publish internal operational assessments of individual tools,” he says, adding that public reporting occurs “where required” through the AI register or other processes.

“I definitely need to know more than is in the register,” says Tusikov. “The register should describe what the tool is and describe how the department is using the tool.

‎

Other enforcement mechanisms, like Algorithmic Impact Assessments—which federal ministries are required to complete at the design stage and before the deployment of AI tools to assess potential risks, biases, and transparency issues—are routinely not released to the public. In a 2025 study, Sieber and her colleagues found that most departments refused to release them; to date, only 40 such assessments have been published. Both the register and the impact assessments may be ill-equipped to even measure the potential risks of generative AI tools.

“The Algorithmic Impact Assessments in the register were not really designed well for generative AI use,” says Sieber.

As the federal government continues to lay off public servants, both Sieber and Tusikov worry that these tools are being used as a replacement for their expertise and labor. “There’s this real assumption here that the automated tools are going to be more accurate, more comprehensive, and more efficient than humans,” says Tusikov. “I’d really like some of these departments to put the brakes on this [...] Could we do it better with people rather than training a tool?”