Share

Writing in the 1930s, economist Harold Innis described Canada’s economic situation in biblical terms: the country, he said, acted as “hewers of wood, and drawers of water.” This, he wrote, was Canada’s resource economy: a country structurally pegged to larger powers, a place to draw resources from—resources which would go on to have high skill, value-added processes performed outside Canada’s borders.

While the Canadian economy has certainly changed a lot since that time, the pattern remains legible in Canada’s resource economy: over 90 per cent of Alberta’s diluted bitumen, for example, is exported for refining in the United States, with refined products returning at a premium. How, then, is that dynamic playing out in the rapidly expanding AI industry?

Data is not bitumen—it is invisible, intangible, and borderless, and that is precisely what makes its extraction easy to overlook. Yet the extractivist logic is identical: companies gather raw data from populations, refine it via underpaid overseas labour, and sell back globally as proprietary infrastructure.

What makes Canada’s position in the AI economy structurally distinct is that the pressure toward asymmetric value extraction runs in both directions simultaneously. Downstream, Canadian-funded AI development draws on cheap annotation labour from the Global South, building products on a foundation of annotation labour whose conditions are often undercompensated and psychologically harmful. Upstream, the foundational intellectual property (the model architectures, the GPU chip designs, the platform infrastructure) is largely held by American technology corporations. 

Canada’s $2 billion sovereign compute investment risks, in substantial part, building infrastructure that runs foreign-owned AI platforms under Canadian hosting arrangements. Canada occupies the precarious middle of a global value chain: extracting from those below while being extracted from above.

Sovereignty over the server rack is not the same as sovereignty over the system, and sovereignty over the system is not the same as capturing the economic return. A July 2025 Asia Pacific Foundation of Canada analysis found that despite launching the world’s first national AI strategy in 2017, Canada still holds under three per cent of global digital services exports, a figure that has barely moved in two decades. The compute infrastructure companies are building (with significant public subsidies) risks repeating the same pattern at a larger scale.

Invisible supply chains

Canadian AI companies are already active in the Global South. TELUS Digital runs a global annotation business employing workers in the Philippines, India, and Latin America. None of this is subject to supply chain diligence—a Philippines government official has described data annotation as an “informal sector” with no regulatory protective mechanisms in place.  Canada’s Fighting Against Forced Labour and Child Labour in Supply Chains Act, in force since January 2024, was designed for physical goods crossing physical borders. Data annotations travel by file transfer and API call, and do not appear in a customs manifest. 

Back in Canada, Alberta’s electricity regulator had over 30 AI data centre projects in its approval queue as of February 2026, representing approximately 93 per cent of Canada’s planned AI compute capacity. The proposed Wonder Valley project alone would emit enough greenhouse gases to erase Alberta’s entire coal phaseout gains. The project also draws up 24 million cubic metres of water annually from the Smoky River under a licence the Sturgeon Lake Cree Nation challenged and lost. Canada’s National Observer confirmed on the day of the AI for All launch that the strategy contains no new protections for water or climate.

The regulatory void

Canada’s only dedicated AI legislative instrument, the Artificial Intelligence and Data Act, died in January 2025 without producing a single enforceable rule. In its place, a patchwork of existing instruments governs AI development: privacy law through the Personal Information Protection and Electronic Documents Act (PIPEDA), human rights and consumer protection statutes, and a voluntary Code of Conduct for Generative AI. None requires companies to disclose training data provenance or upstream labour conditions. 

The gap is not due to lack of legislative capacity. In June 2026, the federal government introduced Bill C-34, the Safe Social Media Act, to protect children from online harms. Similarly, Bill C-35, the Ban on Importing Goods Made with Forced Labour Act, is now at the second reading. The bill would ban imports of goods produced by forced labour, but as drafted it keeps the same physical-goods framing. The gap is less a total absence of law than a mismatch between existing frameworks and the realities of AI supply chains.

The EU AI Act requires providers of high-risk AI systems to document the origin  and collection processes of their training datasets.Canada lacks a comparable general statutory rule requiring public transparency of training data provenance. 

The Centre for International Governance Innovation argues that Canada should anchor its domestic AI rules to international standards and actively build regulatory interoperability with its trade partners, rather than constructing a comprehensive regime that competes with the rest of the world. The authors explicitly carve out areas where domestic rules are warranted: online harms, Indigenous data rights, and federally regulated sectors.

However, the three jurisdictions that dominate the global AI economy do not merely have different rules; they operate on incompatible premises. The EU anchors AI governance in fundamental rights. China subordinates it to state authority. The U.S. governs primarily through executive directives, a tool one administration can issue, reverse, or weaponize with no legislative process at all.

In June 2026, for example, Washington issued an export-control directive to Anthropic restricting foreign access to its most advanced models, citing national security. Two weeks later, the administration partially reversed itself. This abrupt policy swing shows that access to frontier AI can be revoked by foreign executive decision alone; sovereignty over publicly funded server racks means little if the model running on it can be summarily switched off. 

Interoperability across major AI jurisdictions is a values problem, not a technical coordination challenge. Mutual recognition cannot reconcile fundamentally incompatible definitions of fair labour, consent, or public interest. True sovereignty cannot be outsourced to a lowest-common-denominator international standard. Deferring the hard questions—who bears the cost of annotation labour, whose land hosts the infrastructure, and who captures the return on public investment—is abdication. 

Canada’s position as a late entrant into AI governance is not a handicap if it is used wisely. The existing frameworks represent years of hard-won compromise Canada does not need to repeat from scratch. The question is not which regime to align with, but which instruments are consistent with Canadian values: the labour commitments, the Indigenous obligations, the public accountability traditions. That judgment cannot be outsourced. It is precisely the work that a sovereign AI strategy should be doing.

Supply chain sovereignty

Digital sovereignty, then, requires extending its logic to the full supply chain. The five measures below use existing federal authorities; none requires new legislation.

First, reform procurement. The Treasury Board should, through an amendment to the Directive on the Management of Procurement and the Policy on Social Procurement, require companies receiving federal AI funding to attest that training data was produced under conditions consistent with International Labour Organization standards, placing the compliance burden on large subsidized recipients rather than startups. Canada already imposes comparable conditions in other procurement contexts.

Second, assess dependency before buying. The Privy Council Office should, through its existing machinery-of-government authorities, require a standardized dependency assessment before any federal department procures an AI system above a defined threshold: who controls the data, who owns the IP, and what happens to publicly funded assets if the commercial relationship ends. 

Third, bring back the watchdog. The federal government should re-establish the Canada Ombudsperson for Responsible Enterprise (CORE) by executive order, the same mechanism used to eliminate it, with an expanded mandate covering AI annotation labour. The timing makes the case. CORE’s first and only formal finding of human rights abuse, issued in March 2024 against Dynasty Gold Corporation, showed the mechanism could work when resourced to investigate. In June 2026, the government eliminated the office entirely, citing a thin track record that critics attribute to the office never being given power to compel testimony or documents. Closing Canada’s only corporate human rights watchdog just as a new wave of AI-funded supply chains opens abroad is a choice worth reversing.

Fourth, attach consent conditions. The federal government should, as a condition of signing any future federal memoranda of understanding for large-scale data centre approvals, attach free, prior, and informed consent requirements consistent with the UNDRIP Act. Provincial regulators evaluate projects one at a time; the federal funding lever applies across all of them.

Fifth, fund augmentation, not automation with federal funding. A February 2026 Brookings Institution framework distinguishes pro-worker AI, which expands what workers can do and raises the value of their expertise, from automation that simply replaces labour at lower cost. Replacing workers is often simply the path of least resistance. AI for All’s pro-worker language is currently aspirational; requiring recipients of federal AI funding to demonstrate which category their deployment falls into would turn that language into a procurement condition.

Canada has a pattern of letting accountability instruments arrive late: softwood lumber, mining due diligence. AI need not follow the same arc. The question for Canada is whether it will act while the pattern is still interruptible.

AI for All is an ambitious strategy. Its ambition is incomplete insofar as it does not yet address the costs being externalized onto annotation workers in the Global South, onto Indigenous communities where data centres are being built in Canada, and onto lower-income countries that will purchase Canadian AI products without the institutional capacity to govern them. 

Canada can choose to govern the asymmetric value extraction it is already part of. Or it can entrench the dependency that the strategy was designed to escape.

About the author

Si Thu Naing

Si Thu Naing is an MBA candidate in Sustainable Innovation at the Peter B. Gustavson School of Business, University of Victoria.