Leading artificial intelligence models tested by a Toronto research team aligned more closely with Canadian public opinion than with American views on a range of questions, a result that challenges assumptions about how U.S.-built systems reflect and transmit national values.
Transformer Lab, a Toronto-based AI company, compared responses from versions of five widely used systems—OpenAI’s GPT, Anthropic’s Claude Opus, SpaceXAI’s Grok, Meta’s Llama and Alibaba’s Qwen—to patterns in the World Values Survey, a long-running international study of public attitudes. The team selected 10 survey questions where Canadians and Americans diverge the most—touching on religion, immigration, confidence in government, gender roles and interpersonal trust—and then instructed each model to answer three ways: as an American, as a Canadian and with no stated identity.
Across those prompts, the models’ outputs more often landed closer to Canadian responses than to American ones, according to the company. The finding complicates the common belief that AI developed by U.S. firms and trained on predominantly American data will naturally encode and export American viewpoints, especially to neighbouring countries like Canada that depend on those technologies.
“Canada depends on all of these U.S. models, and that’s a massive concern around sovereignty,” said Transformer Lab co-founder Ali Asaria. “We thought it was obvious that these models are trained in the U.S. and they would have U.S. bias.”
Transformer Lab’s analysis suggests otherwise, at least for the questions it examined. When no national identity was specified, the systems still tended to hew closer to positions more commonly held by Canadians on themes such as immigration, national pride and trust in public institutions. That pattern held across the five models tested.
“We can say conclusively that all of the frontier models we tested consistently lean toward Canadian bias,” Asaria said. “That is pretty clear t”
What was tested—and why it matters
The team’s approach was straightforward: pick areas where the two countries’ attitudes are most distinct, and probe whether the models’ baseline answers fell nearer to one public’s views or the other’s. For Canadian readers, the result lands in the middle of two debates: how to safeguard technological sovereignty in a market dominated by foreign platforms, and how AI systems might shape civic life by normalizing particular value frames through everyday interactions.
- Models examined: Five prominent large language models spanning U.S. and international developers.
- Benchmark: Questions drawn from the World Values Survey where Canadian and American responses diverge most.
- Domains: Immigration, religion, national pride, confidence in government, gender roles, interpersonal trust.
| Model | Developer |
|---|---|
| GPT | OpenAI |
| Claude Opus | Anthropic |
| Grok | SpaceXAI |
| Llama | Meta |
| Qwen | Alibaba |
Challenging assumptions about cultural export
The findings push back on the idea that Silicon Valley’s generative AI will serve as a direct conduit for exporting American cultural values to countries like Canada. Transformer Lab expected to see a U.S.-centred tilt, given where the technology is built and the data it is believed to draw upon. Instead, the researchers observed a consistent lean toward Canadian-aligned answers, including when the systems were not instructed to take on a national identity.
For Canada, which relies on foreign-made AI tools across business, education and government, the results cut two ways. On one hand, they suggest that heavy dependence on U.S. systems does not automatically equate to adopting American-leaning value frames in public-facing AI. On the other, they underline the sovereignty concern Asaria highlights: if most models used domestically are controlled abroad, the underlying value orientations—and how they shift with updates—remain outside Canadian oversight.
The sovereignty question
As the federal government and provinces weigh regulatory approaches to generative AI, understanding how systems position themselves on sensitive social questions is increasingly salient. If widely deployed models are already closer to Canadian attitudes on topics like immigration and trust in government, policymakers face a nuanced task: ensuring transparency and accountability without presuming a foreign bias that this study did not detect in its sample of questions.
The study also raises practical questions for Canadian companies and institutions integrating AI into services: how to assess value alignment, how often to retest as models evolve, and how to compare across providers. While the analysis covered five well-known systems and a focused set of questions, its central observation—that responses skew toward Canadian public opinion—adds a new dimension to national discussions about AI adoption and governance.
Ultimately, the research highlights a paradox at the core of Canada’s AI posture: high reliance on external platforms paired with a desire to preserve domestic priorities and norms. The indication that prominent models may already reflect Canadian-leaning answers on some divisive issues does not settle the sovereignty debate. It does, however, sharpen the questions that matter most for Canadians who will live with these systems: whose values guide them today, and who decides tomorrow.
