The Canadian submission · Short summary
From Transparency
to Human Control
A Systems-Engineering Response to Canada’s Consultation on Artificial Intelligence Transparency
Read the complete PDFThe argument
Transparency should be connected to operational control.
Canada’s consultation identifies important questions about generated content, AI interactions, system limitations, serious incidents and agent activity. The submission carries those questions one step further: what can a person actually do with transparency when something goes wrong?
Its answer is to connect disclosure with human recourse, independent testing, reliable records, identifiable agents, incident reconstruction and the practical ability to intervene.
Seven arguments
A practical Canadian contribution.
Transparency should lead to human recourse
A notice that AI was used is not meaningful if the affected person has no effective way to challenge what it did. For consequential applications, people should be able to identify AI’s role, the responsible institution, the route of appeal and the accountable human able to review the result.
Canada should independently test important AI systems
High-impact systems should be evaluated by reproducible testing, not only by provider claims. Defined tests can examine reliability, safeguards, permissions, resistance to attack, uncertainty disclosure and human-override mechanisms without turning government into an arbiter of approved opinion.
Provenance must be visible at the point of use
Metadata is useful only when ordinary people can see and understand it. Browsers, email clients, messaging applications and social platforms should present provenance information consistently while recognizing that provenance can be missing, removed or unreliable.
Require audit traces, not artificial “thought transcripts”
Consequential agents should leave trustworthy records of the system and software involved, responsible operator, permissions, tool use, external actions, approvals and timestamps. These records should focus on observable decisions and actions—not private internal reasoning—and should not be erasable by the system being audited.
AI agents should have verifiable identities
Agents able to spend money, access confidential data or affect real-world systems should carry interoperable credentials appropriate to their capability and risk. Higher-risk services should be able to identify the responsible party, delegated authority and transaction, and to reject unidentified automated actors.
Incident reporting requires evidence
Useful reporting depends on records that let investigators reconstruct which system operated, what authority and permissions it possessed, what it accessed, what it did and when humans intervened. Without evidence, incident reporting becomes competing narratives after failure.
Regulate where Canada has practical leverage
Canada cannot govern every model used within its borders, but it has greater leverage over domestic deployment, procurement, regulated industries, critical infrastructure and services offered to Canadians. Policy should emphasize certification, testing, recourse, auditability, agent identity and incident reporting.
The connection
Human Administrator Rights
The paper’s recommendations resolve into seven core controls: Know, Read, Appeal, Override, Audit, Stop and Offline. Together, they express a simple systems principle: society should preserve the ability to inspect consequential systems, understand what happened, change permissions, stop unsafe processes and recover from failure.
Artificial intelligence can be given enormous capabilities and significant autonomy. Humans should retain administrator access.