Consultations on the Government of Canada AI Register: What We Heard

Concerns

Respondents reported that the Register is difficult for non-expert users to understand. The Register presented detailed technical and administrative information better suited to data professionals or governance researchers than the general public. Respondents suggested that it would be helpful to identify the Register’s intended audiences and provide tailored information for these audiences on what the Register is and how to interpret it.

Respondents also noted that the Register was difficult to navigate and use. The CSV format did not support practical filtering or searching. Searches were made more complex by inconsistencies in how information was entered. Respondents also noted issues with accessibility and bilingual usability.

“The AI Register would be easier to use with clearer orientation and guidance for first-time users. A short plain-language overview explaining what the Register is, who it is for, and how the information should be interpreted would help set expectations.”

“Reduce overly technical jargon or replace it with plain language. Remove repetitive or redundant descriptions. Simplify long sections into concise summaries with optional ‘read more’ details. Revise unclear terminology to make it more accessible to the general public.”

“The Register appears to be only made for users comfortable with data, not for those who just glance at the page.”

Improvements suggested by respondents

• Clarify audiences. For each audience, provide a short explanation of what the Register can help them do, the types of questions it cannot answer, and any other important limitations or caveats.
• Create an “About the Register” landing page with a brief section of frequently asked questions.
• Improve navigation and usability by expanding search and filtering options to enable search and filter by:
  • department or organization
  • use case, purpose or type of use (for example, triage, routing, recommendations) or capability
  • risk or algorithmic impact assessment impact level or decision type
  • affected stakeholder types
  • public‑facing versus internal use
  • data type (for example, personal, health, financial)
  • vendor
  • life-cycle stage
• Introduce standardized brief system summaries (“system cards”) written in plain language to accompany more complex technical descriptions.
• Improve accessibility compliance and bilingual presentation consistency.

Concerns

Feedback highlighted wide variation in the completeness, depth and quality of Register entries. Blank fields, inconsistent terminology, and varying entry quality across departments made the Register more difficult to use for research, comparison or analysis. Respondents noted that key fields were often left blank, populated inconsistently, or described using department-specific terminology that limited comparability. Status labels (for example, “pilot,” “in development,” “retired”) were inconsistently applied, making it difficult to understand system maturity or operational use.

“There is a lack of documentation – there is no codebook, so it is hard to understand what the variables mean.”

“Inconsistency of records: Copilot goes by many names, some names are very generic (for example, ‘machine learning model’). The results variable mixes desired outcomes and results.”

“Blank cells are indistinguishable from fields where information was not collected versus not applicable, creating confusion about completeness.”

Issues with data completeness and data quality in the MVP arose from compiling data from existing sources rather than using a dedicated collection instrument. Nonetheless, the concern highlights the importance of establishing strict data standards in future versions.

Improvements suggested by respondents

• Create a public-facing data dictionary or codebook explaining what each field means and how it should be interpreted.
• Use drop-down lists where possible to standardize responses.
• Establish controlled vocabularies for core fields such as system status, AI capability type and data categories.
• Introduce a visible completeness indicator for each entry.
• Define a minimum disclosure standard that all entries must meet.
• Require last updated dates and a scheduled review cadence for each system.

Concerns

Respondents reported difficulty understanding which AI systems pose higher risks or have more significant impacts. This issue was compounded by gaps in accountability and governance context, including missing human oversight details, life-cycle information and links to algorithmic impact assessments. Respondents said that the Register neither identifies which systems are high-impact nor explains in accessible terms what the systems actually do. This made it difficult to prioritize scrutiny and distinguish internal tools from systems affecting Canadians’ rights and services. Presenting low-risk internal tools and high-impact decision systems in the same way can flatten risk perception, confuse non-expert readers, and potentially raise unnecessary concern about low-risk tools.

“Right now, the Register tells that AI is being used, but it would be far more useful if it included risk or impact level, clearer information on data sources and model provenance (including whether third-party vendors are involved), and how governance actually operates in practice.”

“The Register should clearly display each system’s Impact Level, based on the Directive on Automated Decision Making, with a direct link to its AIA [algorithmic impact assessment] to enable public scrutiny of risk assessments.”

“Risk categories should not be limited to privacy and security. They should include risk to employment, human agency, environment, etc.”

“The Register currently does not explain what the listed systems actually do in terms that affected individuals can use to determine whether AI was involved in their case.”

“Would love to see links to the system’s AIA [algorithmic impact assessment], if one exists, as well as to PIAs [privacy impact assessments] and the [GC] Service Inventory.”

Improvements suggested by respondents

• Introduce risk-tiered classifications (low, medium, high impact), aligned with the Directive on Automated Decision-Making.
• Require explicit linkage to the system’s algorithmic impact assessment level and documentation or a clear justification if exempt.
• Add links to other governance documents such as privacy impact assessments, personal information banks, and the GC Service Inventory. Where possible, avoid repeating information that is already provided in existing governance tools and provide direct links to those tools.
• Add structured fields identifying affected stakeholders and primary risk domains (privacy, bias, security, safety, environment, financial cost).
• Include a brief summary of safeguards and human oversight mechanisms in plain language.
• Visually differentiate high-impact systems within the Register interface.

Concerns

A common concern was that the Register did not clearly explain what role AI plays in decisions. Respondents said that it was not clear whether AI systems are advisory, partially automated or fully automated. Descriptions of AI systems often employed vague terminology such as “AI-assisted,” which obscured how AI outputs influenced decisions that affected individuals, organizations or outcomes. Additionally, human accountability for decisions supported or influenced by AI was not consistently identified, and the location of the AI system within the workflow was rarely explained.

“Terms such as ‘AI-assisted,’ ‘AI-enabled,’ or ‘automation’ are used without consistently indicating how much independent action the system can take. It is often unclear whether AI is merely providing recommendations, actively shaping decisions, or executing decisions autonomously.”

“It would be helpful if the Register included clearer information about how each AI system affects decisions or services. Right now, the Register explains what the systems are designed to do, but it is not always clear whether they are used to support staff decisions or directly influence decisions about people.”

“The Register would be significantly more useful if it focused less on listing AI systems in isolation and more on explaining decision architecture – how AI actually interacts with human judgment, authority, and accountability in practice.”

Improvements suggested by respondents

• Introduce a mandatory field that clearly identifies whether an AI system supports decision-making.
• Introduce a mandatory, categorical field clearly classifying the AI system’s decision role (for example, advisory, human in the loop, fully automated).
• Add a short plain-language description explaining where the AI system sits in the decision workflow and how its outputs are used.

Concerns

Respondents noted that information in the Register about data use gave limited detail. It often did not clearly indicate whether personal, sensitive or protected data were used. Data residency, vendor involvement and third-party system dependencies were not always transparent.

Improvements suggested by respondents

• Standardize disclosures on data categories (for example, personal, health, financial, biometric).
• Clearly indicate data residency (in Canada or outside Canada).
• Identify the role of vendors (tool provider, operator, co-developer).
• Where disclosure is limited by contract or security restrictions, include a transparency limitation statement.

Concerns

Respondents called for mechanisms enabling individuals to:

• understand whether AI was involved in a decision affecting them
• seek recourse where appropriate
• enquire further about specific systems

“Citizens need to know who to contact and what process exists if an AI-assisted decision affects them. The Directive on Automated Decision-Making already requires human review at higher impact levels; the Register should surface this plainly.”

“Redress mechanisms: Explicitly indicate how individuals can review and contest automated decisions and exercise their right to an effective remedy.”

“Establish a direct public reporting mechanism within the Register enabling individuals and civil society to flag materialized AI harms, algorithmic bias, or system failures.”

“Without an accountability mechanism, the Register risks being ‘transparency washing.’”

Improvements suggested by respondents

• Require each entry to include a named point of contact (or functional role) and a description of how individuals affected by AI-assisted decisions can seek clarification or redress.
• Link each high-impact system entry directly to the relevant appeal or redress mechanism.
• Add a feedback and incident reporting feature to allow users to flag AI harms or system failures.
• Include a mechanism for affected individuals to flag that they believe an unlisted AI system was used in a decision affecting them.