When Quebec Premier Christine Fréchette met U.S. Trade Representative Jamieson Greer in Washington in late April, she named two priorities as non-negotiable during the upcoming review of the Canada-United States-Mexico Agreement (CUSMA): supply management and the French language. Digital sovereignty, including artificial-intelligence governance, was not on the list.
The omission is the story.
Quebec is, by infrastructure and electricity policy, the most-exposed Canadian jurisdiction to whatever the CUSMA review settles for the data centre, cloud and AI economy. To leave digital sovereignty off the priority list is not a choice of focus. It is a choice of authority surrendered.
The same week, the United States and Mexico opened formal CUSMA review talks. Canada-U.S. talks have not yet begun but Ottawa is also lagging in its approach because it will enter the negotiations with no AI legislation in force and no publicly stated position on the digital-trade chapter that will most directly shape the next decade of how AI is governed in this country.
Here is what that position should include: confirmation of non-discriminatory measures necessary to govern high-risk AI systems; preventative auditing authority over algorithmic systems for labour and consumer regulators; and explicit language preserving provincial authority to attach data-governance conditions to energy-access agreements with data-centre operators.
The issues are serious
AI governance is not a separate question from the broader CUSMA digital-trade debate. However, it is the most consequential instance of it.
The current provisions of the agreement that constrain Canadian sovereignty over data residency, online-platform regulation and government procurement of cloud services also constrain how Canada can supervise the AI systems that increasingly determine wages, hiring, health care, credit and access to government services. Whatever Ottawa concedes or fails to defend on the broader digital question, it concedes or fails to defend on AI as well.
Who governs the digital sphere? How U.S. proxy lobbying erodes Canada’s digital sovereignty
Canada’s digital sovereignty debate overlooks software risks
Canada must ensure its digital sovereignty in the face of U.S. threats
The American agenda has two voices and the relationship between them is the key to reading what is coming.
The first voice is industry. The Computer and Communications Industry Association (CCIA), the principal Washington trade association for the largest American technology firms, including Amazon, Google, Meta and Microsoft, filed CUSMA review comments in November, calling for a new AI annex to CUSMA that would prohibit any requirement on its members to disclose the inner workings of AI systems as a condition of putting an AI product on the Canadian market.
The technical term is model weights: the millions of numerical parameters, learned during training on large datasets, that determine how a trained AI system actually behaves. Disclosing them is the difference between a regulator being able to audit how an algorithmic decision was made and being told to take the developer’s word for it.
CCIA submissions are not U.S. government policy. They are industry advocacy. However, they matter analytically because they have functioned recently as a reliable forward indicator of what the United States trade representative subsequently takes to the negotiating table.
The second voice confirms the pattern. In March, the U.S. trade representative released its 2026 National Trade Estimate Report, an annual catalogue of foreign measures the United States considers barriers to American commerce.
For the first time, the report named Canada’s sovereign cloud and digital sovereignty initiatives as trade barriers. The framing in that report tracks closely the CCIA submission filed five months earlier. Six days later, at the Hudson Institute in Washington, Greer signaled that Canadian digital measures would be a priority in the joint review, with the sovereign-cloud framework named among them.
The CCIA filed in November. The USTR adopted the framing in March. Industry advocacy has become state policy in the space of four months.
Canada, meanwhile, has nothing on the table. The federal Artificial Intelligence and Data Act died in January 2025, when Parliament was prorogued when the April 2025 election was called. AI Minister Evan Solomon has confirmed the act will not return in its previous form.
In January 2026. Ontario’s Information and Privacy Commissioner and its Human Rights Commission jointly released six principles for the responsible use of artificial intelligence. However, these are principles, not law.
Canada is still walking into negotiations over the legal architecture of AI governance with nothing binding to defend, while the other side has a draft.
The structural problem sits in Article 19 of CUSMA, the digital-trade chapter.
Article 19.16 prohibits the parties from requiring access to source code or algorithms (the technical instructions that make a software product work) as a condition for selling that product in the country. The article contains a narrow exception for specific enforcement proceedings. There is no general exception for routine regulatory oversight.
Consider what this means in plain terms.
Imagine if a trade agreement prevented Health Canada from inspecting drug-manufacturing processes. Yet that is exactly what Article 19.16 does for algorithmic-management systems – the software that determines wages, scheduling and performance evaluation for millions of Canadian workers.
The AFL-CIO Technology Institute documented in March that employers are increasingly relying on automated systems to hire, monitor, evaluate, discipline and compensate workers across sectors. Canada’s labour regulators cannot meaningfully audit those systems. Its consumer-protection bodies cannot require systematic transparency into recommendation algorithms. The article was drafted in 2018 before commercial generative AI existed and it has not been updated since.
Article 19.12 compounds the constraint. It prohibits Ottawa from requiring a foreign firm to use or locate computing facilities (data centres in plain terms) on Canadian soil as a condition of doing business here.
Unlike the comparable provision in the Comprehensive and Progressive Agreement for Trans-Pacific Partnership, which preserves an explicit exception for measures pursuing a legitimate public-policy objective, CUSMA Article 19.12 has no equivalent carve-out. The only escape valve is the narrow general-exceptions clause carried over from Article XIV of the General Agreement on Trade in Services, the World Trade Organization (WTO) treaty signed in 1995 that outlines how member countries regulate trade in services.
That clause was not designed for AI systems and does not on its face accommodate them. The WTO’s e-commerce moratorium, the multilateral instrument that had kept countries from imposing customs duties on cross-border digital transmissions since 1998, lapsed at the end of March when Brazil and Turkey blocked its renewal. That lapse opens a narrow window for Canada and other countries to revisit the basic rules. Canada should not close it before CUSMA talks begin.
Three specific demands before talks open
Canada enters this negotiation with a new majority government. The mandate is there. What is missing is a position on digital trade.
First, Canada should demand a digital public-interest exception modelled on Article 32 of CUSMA, the agreement’s general-exceptions clause. The demand is not for a blank carve-out, but for explicit treaty language confirming that the non-discriminatory measures necessary to govern high-risk AI systems fall within the agreement’s existing public-interest exceptions. This is a clarification, not a rewrite. Canada should insist on it in writing before substantive talks proceed.
Second, Canada should demand an algorithmic audit carve-out from Article 19.16 for labour- and consumer-protection purposes. The existing provision permits regulatory access to source code only in the context of a specific investigation or judicial proceeding – that is, only after harm has occurred and a complaint has been filed.
What Canadian labour and consumer regulators need is preventative auditing authority – the capacity to examine algorithmic systems that make consequential decisions before those decisions become entrenched practices and on a defined schedule rather than only on suspicion. The carve-out should be scoped to defined categories of algorithmic systems, not to all software, and tied to existing domestic legal frameworks.
Third, Canada should secure conditional energy-access authority for provinces. Quebec’s hydroelectric advantage is a genuine negotiating asset. Under Article 19.12 as currently drafted, however, a provincial requirement that ties the supply of electricity for a data centre to data-governance compliance would be exposed to challenge as a de facto prohibited data-localization measure.
Canada should negotiate explicit language preserving provincial authority to attach data-governance conditions to energy-access agreements with data-centre operators. This is infrastructure sovereignty made concrete. It is also the one mechanism that gives Canada meaningful leverage with the large-scale data-centre operators that have built their AI compute strategy on Canadian, and especially Quebec, electricity.
None of these demands is unusual.
The European Union’s AI Act of 2024 requires that providers of high-risk AI systems supply detailed technical documentation to regulators, including information about the data used to train the model and the design choices that shape its behaviour. The United Kingdom’s AI Safety Institute conducts pre-deployment evaluations of AI models, testing those systems for capabilities and safety risks before they are released to the public. Brazil’s AI bill creates an algorithmic auditing authority.
These regulatory architectures are not theoretical. They are operating now in other jurisdictions. However, they would be foreclosed for Canada by the AI annex that the U.S. is proposing in the CUSMA talks.
The demands above are not maximalist. They are the minimum preconditions for a digital-trade chapter that does not preclude AI governance before the next piece of Canadian AI legislation has been drafted.
The cost of failing to set that floor is not theoretical.
The views expressed in this commentary are the author’s own and do not reflect those of the Law Commission of Canada or any affiliated institution.
