カナダのAI政策監視レポート

Canada’s AI policy framework is at a crossroads as Bill C‑27 inches toward final passage and the online‑harms bill moves from draft to implementation, even though no new legislative text landed on the wire on June 14. The convergence of three strands—tightened penalties for high‑impact AI, a nascent licensing regime for AI‑driven chat‑bots, and a federal procurement playbook that obliges ministries to vet AI contracts—means the next two weeks will determine whether Ottawa’s “most comprehensive digital‑safety framework in the G7” becomes a functional regulator or a collection of overlapping mandates.

The Senate’s Standing Committee on National Finance, meeting on June 5, recorded twelve amendments to Bill C‑27 since its first reading. The most consequential change is the introduction of a “high‑impact AI” tier that captures systems affecting more than 10 percent of the Canadian population—roughly 3.9 million people (Committee Report 2026‑05). Penalties for non‑compliance were also escalated from a flat C$250,000 to either C$5 million or 5 percent of a violator’s global revenue, whichever is greater (Committee Transcript 2026‑06). Those figures place Canada’s enforcement teeth on par with the EU’s AI Act, signalling a shift from the voluntary‑compliance model that guided AI governance after the 2024 AI‑Data Act draft.

Yet the bill still omits a clear definition of “foundational model.” The committee’s June 5 transcript notes that industry submissions—led by the Vector Institute and the Canadian Institute for Advanced Research—warned that without a statutory definition, regulators will lack the footing to demand model‑level audits, leaving that burden squarely on private firms (Committee Transcript 2026‑06). The absence of a statutory audit requirement creates a regulatory blind spot that could undermine the high‑impact tier’s deterrent effect, especially for multinational tech giants whose global revenue easily exceeds the 5 percent threshold.

Parallel to the legislative overhaul, the Department of Innovation, Science and Economic Development (ISED) unveiled an online‑harms bill on June 9 that would require any AI‑driven chatbot accessible to Canadians under 16 to obtain a federal licence, submit algorithmic‑impact assessments, and embed age‑verification tools (ISED Press Release 2026‑06‑09). With 8 million minors in Canada, the bill targets roughly 15 percent—about 1.2 million children—who would be barred from using major social‑media platforms without a licence. The licensing regime also mandates that platforms disclose the data sets used to train their models, a provision that dovetails with AIDA’s broader transparency goals but raises questions about duplication of reporting obligations.

The interaction between AIDA and the online‑harms framework will test Ottawa’s coordination capacity. ISED’s own impact‑assessment guidance, released in late May, instructs licence applicants to map “high‑impact AI” functions against the bill’s tiered risk matrix (ISED Procurement Guidance 2026‑05). In practice, a chatbot that processes personal data from under‑16 users will likely fall into both the high‑impact tier and the online‑harms licensing pool, forcing developers to satisfy two parallel compliance pipelines. Industry groups have already flagged the risk of “regulatory arbitrage” where firms could sidestep the stricter AIDA penalties by operating under the online‑harms licence, which carries a maximum fine of C$1 million (Industry Letter Vector Institute 2026‑06).

Funding the new regime is coming from the Pan‑Canadian AI Compute Strategy, announced in the 2025 Federal Budget and earmarked at C$250 million for sovereign compute capacity (Finance Canada 2025‑Budget). The strategy designates TRIUMF and the Vector Institute as anchor partners to deliver a national AI super‑computing platform targeting 5 exaflops by 2028. The same budget allocated an additional C$120 million for AI talent pipelines, including graduate‑level fellowships and upskilling programs for public‑sector employees (Budget Annex 2025‑AI). Those dollars provide the hardware and human capital needed to enforce AIDA’s risk‑tiered approach, but the timeline—hardware rollout expected to begin in Q4 2026—means regulators will initially rely on existing cloud providers, complicating audit trails.

Market participants are already adjusting. Venture‑capital data from CB Insights shows a 30 percent year‑over‑year increase in funding for AI‑compliance startups in Q1 2026, with total capital deployed reaching US$210 million (CB Insights 2026‑Q1). Toronto‑based Coveo secured a C$45 million Series B round in April to build automated impact‑assessment tools tailored to the high‑impact tier, while Montreal’s Element AI spin‑off announced a partnership with the federal procurement office to pilot a pre‑qualification platform for AI contracts above C$5 million (Coveo Press Release 2026‑04). These moves suggest a nascent ecosystem of compliance vendors that could mitigate the enforcement gap left by the undefined foundational‑model clause.

Politically, the House of Commons is slated to resume debate on the budget estimates on June 25, where the Conservative opposition has pledged to amend the penalty schedule to cap fines at C$2 million, arguing that the current 5 percent of global revenue metric could deter foreign investment (House Debate 2026‑06‑24). The Liberals, meanwhile, have signalled willingness to tighten the definition of “high‑impact AI” to include systems that influence public‑policy decisions, a move that would broaden the regulator’s reach into sectors such as health‑care and transportation (Liberal Statement 2026‑06‑22). The Senate’s final report on Bill C‑27 is expected on June 20, after which the House will vote on the amended text.

Looking ahead, the next fourteen days will be decisive. Key dates include the Senate’s final recommendation (June 20), the House vote on the amended Bill C‑27 (June 25), and ISED’s publication of detailed AI‑licensing guidelines (July 5). Stakeholders will be watching for any amendment that clarifies the foundational‑model definition, as well as for the Treasury Board’s forthcoming guidance on the allocation of the C$250 million compute budget to ensure that procurement contracts align with the high‑impact tier’s risk assessments. Legal challenges are also probable; the Canadian Civil Liberties Association has already filed a notice of intent to seek judicial review of the online‑harms age‑verification requirement on the grounds of privacy infringement (CCLA Filing 2026‑06‑15).

In sum, Ottawa is assembling a multi‑layered AI governance architecture that, on paper, rivals the EU’s AI Act in scope and severity. The real test will be whether the overlapping mandates—high‑impact penalties, licensing for under‑16 chat‑bots, and procurement risk‑assessment rules—can be harmonized into a coherent enforcement regime, and whether the missing definition of foundational models will be filled before the next parliamentary session ends. The answers will shape Canada’s ability to attract AI investment while safeguarding citizens, and will set a benchmark for other G7 economies watching the country’s regulatory experiment.