THE SIGNAL IN ONE SENTENCE

Canada has published the public diagnosis of its first federal AI register. The diagnosis is unusually specific. The register is hard to search. Its CSV format is awkward for ordinary people. Entries use inconsistent terms and leave fields blank without explaining whether information is missing, not collected or not applicable. System status labels do not mean the same thing across departments. Risk is difficult to compare. Human oversight is vague. The role of AI in decisions is often hidden behind phrases such as AI-assisted. Data residency and vendor involvement are unclear. People affected by a system may not know who to contact, how to seek review or where to report a failure. That is not a complaint from a cranky spreadsheet enthusiast. It is the summary the Treasury Board of Canada Secretariat published on November 18 after consulting the public from January through March 2026. Canada released the first register in November 2025 as a minimum viable product. It assembled information from existing sources, including algorithmic impact assessments, access-to-information requests, parliamentary responses, personal information banks and the federal service inventory. The government described it as an early version meant to collect feedback. Now the feedback is in. Canada says an improved register will replace the MVP and be updated regularly. The useful next step is not merely to ship a cleaner interface and declare the consultation complete. The plain signal is this: publish a field-by-field change log with the replacement. For every major consultation request, the change log should state whether the government accepted it, changed it, deferred it or declined it. It should name the affected field or feature, the reason, the owner, the implementation status, the target version and the evidence that the change works. If a recommendation cannot be implemented because of security, privacy, procurement, technical or legal limits, say that too. This would turn the consultation from a collection of thoughtful paragraphs into a public product decision. The distinction matters because the What We Heard report is a summary, not a commitment. It describes themes that respondents raised. It does not publish a response count, claim that respondents represent everyone in Canada or promise to implement every suggestion. The government has not announced a launch date for the replacement register on the report page. A change log would preserve that honesty. It would show where public feedback changed the register without pretending that every comment became policy. Start with usability. Respondents said the MVP was difficult for non-experts to understand and difficult to navigate. They asked Canada to clarify the audiences, create an About page and frequently asked questions, write plain-language system cards and improve accessibility and bilingual consistency. They also asked for search and filtering by department, purpose, capability, risk, impact level, decision type, affected stakeholder, public-facing versus internal use, data type, vendor and life-cycle stage. That is a product backlog hiding in a consultation report. The public change log should break it into inspectable decisions: Audience guidance, accepted or not. Plain-language system cards, accepted or not. Accessibility target, owner and test status. Bilingual field parity, owner and unresolved gaps. Search filters, delivered field by field. Download format, machine-readable schema and update cadence. Do not mark usability complete because a search box exists. Test whether a person can answer a real question. Which federal systems influence decisions about benefits? Which use health or financial data? Which are fully automated? Which rely on an outside vendor? Which are retired? Which have a public appeal route? If the interface cannot answer those questions, the filter may be decorative. Next comes data quality. The MVP inherited information from several existing sources. That was a practical way to publish early, but it produced inconsistent entries. Respondents reported blank fields, department-specific terminology, generic system names, mixed meanings and status labels that made comparison difficult. The report recommends a public data dictionary, controlled vocabularies, drop-down choices, a visible completeness indicator, a minimum disclosure standard, last-updated dates and a scheduled review cadence. Those are not glamorous features. They are what make a register a register. A list of systems becomes useful when the fields mean the same thing from one row to the next. The change log should identify every normalized field and publish its definition. What does deployed mean? Is a pilot processing live personal information or only synthetic data? Does retired mean the model stopped running, the service stopped using it or the contract ended? Is vendor the company that built the tool, operates it, hosts it or helped configure it? Blank must have a meaning too. Use separate states for not applicable, not collected, withheld, unknown and pending update. A silent empty cell can conceal five different governance conditions. Then show completeness. Each register entry should display which required fields are present, which are overdue and which are withheld with a stated reason. The change log should state when the minimum disclosure standard took effect and which legacy entries have not yet met it. That will be uncomfortable. Good. A transparency tool should make missing transparency visible. Risk and impact need their own layer. Respondents said the register made it difficult to distinguish a low-risk internal helper from a system that may affect a person’s rights, services or access. They asked for risk tiers aligned with the Directive on Automated Decision-Making, direct links to algorithmic impact assessments, privacy impact assessments and other governance documents, affected stakeholder fields, risk domains and plain-language summaries of safeguards and human oversight. The replacement should not turn the risk tier into a shiny badge. Show the assessment date, version, responsible institution, result, link, exemptions and changes since the last assessment. If the system changed materially after the assessment, flag the mismatch. If an assessment is not required, explain why. Risk domains should go beyond privacy and security. The consultation report names bias, safety, environmental impact, financial cost, employment and human agency among the concerns people may need to see. The change log should show which domains became structured fields, which remain narrative and how Canada tested whether the categories help people prioritize scrutiny. Decision architecture may be the most important missing field. Respondents said descriptions such as AI-assisted did not reveal whether a system merely offered a suggestion, actively shaped a decision or executed one autonomously. They asked for a mandatory categorical field identifying the decision role, with a plain-language explanation of where the system sits in the workflow and how people use its output. Ship that field with a controlled vocabulary. Advisory. Recommendation with mandatory human review. Recommendation with optional review. Automated triage. Automated decision with human appeal. Fully automated operational action. Other, with explanation. The exact labels can improve. The point is to stop using one soft phrase for several very different kinds of authority. Then name the human role. Who is accountable for the final decision? Who reviews the output? What evidence can the reviewer see? Can the reviewer reject it? Does an override require reasons? Who can stop the system? Human in the loop is not a sufficient entry. Sometimes the loop is a rubber stamp with a login. The change log should show that the old vague field was replaced, why the new categories were chosen and how many existing entries still need classification. Data and vendors require the same discipline. Respondents asked for standard disclosure of personal, health, financial, biometric and other data categories. They wanted to know whether data stays in Canada or is processed elsewhere, whether a third party is involved and whether that party is a provider, operator or co-developer. Some information will properly remain limited for security, privacy or contractual reasons. The report suggests using a transparency limitation statement when full disclosure is not possible. That is better than silence. The register can say that a field is withheld, identify the category of reason, name the approving authority and give a review date without publishing protected details. The change log should make those limitation rules public before departments start applying them. Otherwise every difficult field may acquire a bespoke fog bank. Recourse is where the register meets a real person. Respondents asked each entry to include a named point of contact or functional role, a description of how affected people can seek clarification or redress, direct appeal links for high-impact systems and a way to report harms, bias or failures. They also suggested a route for someone to report that an unlisted AI system may have affected a decision. Build the register around that journey. A person should be able to find the system, understand its role, identify the responsible institution, reach a human, request review, correct data, follow an appeal and report an incident. The register should not become the appeal system for every federal service. It should connect people to the right one. Measure whether those connections work. Test the links. Publish response-time expectations. Record broken routes. Distinguish a general information address from an office authorized to change a decision. Provide an incident reference that a person can keep. The change log should state which recourse fields became mandatory and how the government will handle reports about unlisted systems. Layered disclosure can make all of this usable without turning each entry into a filing cabinet. The first layer should answer basic questions in plain language: what the system does, whom it affects, whether it shapes a decision, its current status, risk level, responsible institution and review route. The second layer can show data, vendors, oversight, assessments, safeguards and update history. The downloadable layer can provide structured fields for researchers, journalists, auditors and civil society. Different audiences need different depth, not different facts. Version the schema and the entries. Every register page should show when it changed and what changed. A system that moves from pilot to deployment should not silently replace its earlier state. A vendor change, new data source, revised risk assessment, expanded user population or removed human review should create a visible history. The public change log for the register itself should use the same habit. Request. Decision. Rationale. Owner. Status. Target version. Verification. Link. That is enough to let a person follow the work without reading meeting minutes. Canada’s report includes one small disclosure that models the approach. It says generative AI helped draft parts of the document, that no personal consultation information was processed by the tool and that government employees reviewed and validated all content. That note does not prove the drafting process was perfect. It does show the value of naming the AI role, data boundary and human accountability in plain language. Put the same clarity into the register. The initial MVP deserves credit. Publishing imperfect data invited scrutiny that a private database never would have received. Respondents broadly valued the existence of the register, the iterative approach and the amount of information disclosed. Iteration earns trust only when people can see the iteration. The next register should not merely look more finished. It should show what the government learned, which public requests changed the product, what remains unresolved and how a person affected by federal AI can act on the information. Ship the replacement. Ship the schema. Ship the change log beside it.

01

WHAT ACTUALLY CHANGED

Canada published its What We Heard report on the federal AI Register consultation on November 18, 2026.

The Treasury Board of Canada Secretariat conducted the public consultation from January through March 2026.

The first register was released in November 2025 as a minimum viable product assembled from existing government information sources.

Canada says an improved register will replace the MVP and be updated regularly.

Respondents valued the existence of the register, its iterative approach and the breadth of disclosed information.

Reported concerns included weak search, difficult CSV navigation, accessibility and bilingual inconsistency.

Entries were described as uneven, incomplete and difficult to compare because of blank fields, inconsistent terminology and unclear status labels.

Respondents requested clearer risk, impact, human oversight and decision-role information.

They also requested structured data categories, residency and vendor roles, named contacts, redress links and incident reporting.

The report discloses that generative AI helped draft parts of it, with no personal consultation data processed and employee review and validation.

02

WHY THIS MATTERS

A register cannot support accountability if ordinary people cannot search or interpret it.

Inconsistent fields make comparisons among departments and systems unreliable.

A blank cell can conceal missing, withheld, unknown or inapplicable information.

Low-risk internal tools and high-impact decision systems should not appear equivalent.

The phrase AI-assisted can hide whether software advises, shapes or executes a decision.

People need to know which data and vendors participate in systems affecting public services.

Transparency becomes actionable when each entry connects to a responsible institution and review route.

A consultation report records feedback but does not show which recommendations became product decisions.

A public change log can distinguish accepted, deferred and declined requests without pretending every comment became policy.

Visible version history can reveal when a system’s authority, data, vendor or safeguards change.

FIG. 681THE REGISTER CHANGE LOG
1NAME THE PUBLIC REQUEST→
2IDENTIFY THE FIELD OR FEATURE→
3RECORD THE DECISION→
4EXPLAIN THE RATIONALE→
5ASSIGN OWNER AND VERSION→
6VERIFY AND PUBLISH THE RESULT
Public feedback becomes accountable when every product decision leaves a visible trail.

03

WHERE IT COULD HELP

  • Publish a field-by-field change log with the replacement AI register.
  • For each consultation request, state the decision, rationale, owner, status, target version and verification evidence.
  • Create an About page and audience-specific guidance in plain language.
  • Add search and filters for department, use, capability, risk, stakeholder, data, vendor, status and life-cycle stage.
  • Publish standardized system cards for non-expert readers.
  • Test accessibility and bilingual consistency as product requirements.
  • Create a public data dictionary and controlled vocabularies for core fields.
  • Distinguish not applicable, not collected, withheld, unknown and pending update.
  • Show a completeness indicator and scheduled review date for every entry.
  • Publish minimum disclosure requirements and identify legacy entries that do not meet them.
  • Display risk tiers with assessment dates, versions, exemptions and links to supporting documents.
  • Add structured fields for affected stakeholders, risk domains, safeguards and human oversight.
  • Classify the AI decision role with a mandatory controlled vocabulary.
  • Describe where AI sits in the workflow and who owns the final decision.
  • Disclose data categories, residency and each vendor’s operating role.
  • Use a standard transparency limitation statement when information must be withheld.
  • Include a responsible contact, clarification route, correction path and appeal link.
  • Provide incident reporting and a route to flag potentially unlisted systems.
  • Layer plain-language summaries, governance details and downloadable structured data.
  • Version the schema and preserve entry histories when authority, data, vendors or status change.

KEEP A HAND ON THE WHEEL

The report is an official summary of consultation themes. It does not publish the number of responses, claim that respondents are nationally representative or commit the government to every suggestion. Canada says feedback will inform an improved register that replaces the MVP and is updated regularly. The report page does not announce a launch date for that replacement. The original MVP was assembled from several existing information sources rather than a dedicated collection instrument. The report identifies this as a cause of data-quality and completeness problems. A public register does not prove that every federal AI system is listed, accurately described or safely governed. Risk tiers and system cards can simplify information while concealing important uncertainty if definitions, evidence and version history are absent. Security, privacy, contractual and legal restrictions may properly limit some disclosure. Limitation statements still need consistent categories, accountable approval and review dates. Named contacts and appeal links must connect to offices with actual authority. A general inbox is not necessarily a recourse mechanism. Incident reporting requires privacy, security, triage, response-time and records-management safeguards. The report states that generative AI helped draft parts of the document, that no personal consultation information was processed by the tool and that government employees reviewed and validated all content. Watch for the replacement register, its schema, public implementation decisions, update cadence, completeness rules, risk fields, decision-role categories, data and vendor disclosures, recourse routes and entry version history.

04

TERMS WORTH KEEPING

SOURCES AND VERIFICATION STATUS

This article was written from the materials below. Product claims and dates were checked against those sources on November 18, 2026.

THE PUBLICATION ENGINE

WANT A SIGNAL OF YOUR OWN?

We build source-grounded publications, private briefings, and editorial systems for organizations with something useful to say.

WORK WITH US