Is AIDA in force in Canada?

AIDA never came into force in Canada. It was Part 3 of Bill C-27, and the bill died when Parliament was prorogued on 6 January 2025. It has not been reintroduced as of August 2026, and the federal government has signalled that AI will be governed through privacy legislation and targeted measures while any successor statute is designed from scratch.

What actually governs AI in Canada right now?

PIPEDA governs private-sector personal information federally and in provinces without substantially similar legislation. Quebec's Law 25 adds section 12.1 automated-decision rights. OSFI guidance covers model risk at federally regulated financial institutions, and the Treasury Board Directive on Automated Decision-Making covers federal government systems. Alberta and British Columbia have their own private-sector acts.

What does Law 25 section 12.1 require in evidence terms?

For a decision based exclusively on automated processing, you must inform the individual, and on request supply the personal information used, the principal factors and parameters that led to the decision, and the ability to have the decision reviewed by a person. In practice that is a per-decision record tying the input to the output and the applied policy, written when the decision request is made.

Do vendor logs satisfy a Canadian privacy review?

Vendor logs reflect the model provider's view of the API call. They lack your identity context and your policy decisions, and they sit outside your custody. PIPEDA places accountability on the organization that controls the personal information, so the record has to exist on your side, showing who sent what, where it went, and which control fired.

Should we build for AIDA anyway?

The artifacts are the same ones PIPEDA accountability, Law 25 section 12.1, and OSFI model risk expectations already ask for: identity binding, prompt classification, destination logging, and per-decision records. A future Canadian AI statute would layer risk tiering on top of that base rather than replace it, so the work holds its value regardless of what Parliament does next.

Canada AIDA AI Audit Evidence: What an Audit Asks For Now That AIDA Never Passed

Canada has no federal AI statute. The Artificial Intelligence and Data Act (AIDA) sat in Part 3 of Bill C-27, and the bill died on the Order Paper when Parliament was prorogued on 6 January 2025. As of August 2026 it has not been reintroduced, and the federal AI minister has signalled that any future statute will be a fresh design rather than an AIDA revival.

The audits happen anyway. A Canadian organization running AI over personal information gets reviewed under the Personal Information Protection and Electronic Documents Act (PIPEDA) federally, under Quebec's Law 25 if it operates in Quebec, under OSFI expectations if it is a federally regulated financial institution, and under the Treasury Board Directive on Automated Decision-Making if it is a federal department. Each of those reviews reaches for the same class of artifact, and it is the artifact AIDA's record-keeping clause would have demanded. I want to walk what a reviewer actually asks for, because Canadian teams have spent two years waiting for a law and building nothing.

The accountability record (PIPEDA Principle 4.1)

PIPEDA Schedule 1 opens with accountability: the organization is responsible for personal information under its control and must be able to demonstrate that responsibility. For AI traffic, demonstration means a per-request record binding a named identity to the specific prompt, the destination model, and the response. Most Canadian deployments cannot produce this, because the application authenticates the user and then calls the model with a shared service credential. The identity the accountability principle depends on gets dropped one layer before the data leaves.

The artifact that satisfies a PIPEDA review names the caller, the classification of what was in the context window, and the policy that permitted the call. That is a different thing from a privacy policy asserting the organization takes accountability seriously.

The automated-decision record (Quebec Law 25, section 12.1)

Quebec's Law 25 carries the most demanding automated-decision requirement in any Canadian jurisdiction. Section 12.1 came into force on 22 September 2023, and it gives an individual the right to be informed when a decision is based exclusively on automated processing of their personal information, plus the right to the personal information used, the reasons and principal factors behind the decision, and a right to have the decision reviewed by a person.

Answering that request needs decision lineage: the input prompt, the output, and the policy state at the moment of the call. This mirrors the evidence burden under GDPR Article 22 automated decisions, and a general description of the model satisfies neither regulator. The record has to be about the one decision the individual is contesting.

The high-impact-system record AIDA would have required

AIDA's March 2023 companion document described record-keeping obligations for high-impact systems: documentation of how the system was built, how it was monitored, and what it did in operation. That obligation never became binding. The interesting part is that OSFI's model risk guidance for financial institutions and the Treasury Board Directive's Algorithmic Impact Assessment both ask for the operational half of it today, and both are live now.

An Algorithmic Impact Assessment scores a system before deployment. What no assessment produces is proof the deployed system behaved the way the assessment described. That proof is a running log of authorized AI calls, which is the same artifact tracked in AI audit trail requirements by regulation.

The cross-border record

Canadian privacy law permits transfers to processors outside Canada under PIPEDA's accountability principle, provided comparable protection is maintained, and Quebec Law 25 requires a privacy impact assessment before communicating personal information outside Quebec. A prompt sent to a model endpoint in Oregon or Dublin is that transfer. The evidence a reviewer wants is a per-request log of the destination endpoint, its region, and the classification of data in the prompt, which is what AI data residency controls produce at the moment of transmission.

Why self-written application logs fail

An application that makes the AI call and also writes the compliance log is attesting to itself. That record breaks in three ways: the application logs its successes and misses its edge-case failures, the log can be modified by the same system that failed, and a crash between the model response and the log commit leaves the action taken with the evidence gone. I set this out in You Own the AI Liability, Not the Vendor.

My candid view: AIDA's death gave Canadian AI programmes a two-year excuse, and the excuse was always thin. PIPEDA accountability and Law 25 section 12.1 were binding the whole time.

DeepInspect

This is the evidence layer Canadian AI programmes usually lack. DeepInspect sits inline between your users or agents and the LLM APIs they call. For every request it evaluates identity, data classification, and destination, then writes a signed, tamper-evident record before the response returns. The record holds the caller, the classification in the prompt, the model endpoint and its region, the policy decision, and the response.

When the Office of the Privacy Commissioner asks what personal information reached which model, or a Quebec resident exercises the section 12.1 review right, the answer is a query against records the application never controlled. The Canada AIDA controls mapping shows how each obligation lines up with an enforcement point. Book a technical deep dive at deepinspect.ai.