A public AI register is a published inventory of AI systems or uses within a defined reporting scope. The number of records in the register is a count of records, not automatically a count of all systems deployed by the institutions it covers.
The scope determines which uses belong in the inventory. The record structure determines how those uses are counted. Publication practice determines how closely the visible entries match that intended population.
The publication rule selects the population
Canada’s federal automated decision-making regime concerns systems used to make or support administrative decisions within its defined scope. It does not encompass every use of AI in the federal public service. Its Algorithmic Impact Assessment results are published on the Open Government Portal.
An inventory assembled from those assessments therefore starts with assessed uses under that regime. Extending the label to all government AI changes the population without adding the missing evidence. A tool used only to draft internal text and a tool informing a decision about a client cannot be assigned the same reporting status merely because both use AI.
Records and systems are different counting units
One system can have an original assessment, a revised assessment and a peer review report. Counting those documents as three systems overstates deployment. Conversely, a platform used in several services can represent several deployments even when the underlying model is shared.
A reproducible count declares its unit: model, technical system, deployment, assessed use or document. It also records how revisions, translations and retired uses are treated. A keyword search cannot supply those decisions; its results still need classification and deduplication.
Take an independently established inventory of 50 deployments within a reporting rule. Suppose published assessments match 40 of them. Coverage is 80% for that population and date. If the same 40 assessments are visible but the full population is unknown, 40 remains a count. It cannot yield a coverage percentage.
Coverage, detail and currency
Coverage measures whether an in-scope use has an entry. Detail measures what a reader can learn from the entry. Currency measures whether the entry describes the relevant version and operating state. Improving one does not establish the other two.
Canada’s peer review guidance requires review for projects at impact levels 2–4 and permits a justified plain-language summary where security, intellectual property or privacy considerations limit full disclosure. A published summary and a full report therefore differ in detail even when both satisfy the described publication route. Guide to Peer Review of Automated Decision Systems.
An assessment also remains a record of a particular configuration and purpose. If either changes, the earlier document does not become evidence about the change merely by remaining online.
Interpreting an omission
A missing entry creates a question about scope, identity, timing or compliance. It does not select the answer. The distinction requires another record: an internal inventory, procurement description, service notice or system documentation that establishes the use being investigated.
A register becomes evidence of coverage through reconciliation with that wider population. Its public value also depends on whether the entries identify the responsible institution, the decision supported and the version assessed. The presence of a record establishes disclosure of that record; claims about deployment and oversight require its contents.