When the team at Toronto-based AI company Transformer Lab set out to test whether leading AI models espouse American values, they expected the results to be obvious: that systems built largely by U.S. companies and trained on vast amounts of data from that country would naturally reflect American views. 

Instead, they found that responses from popular models such as ChatGPT, Claude and Grok more closely mirrored Canadian public opinion than American opinion on questions ranging from immigration and national pride to trust in government. 

The findings challenge an assumption that generative AI developed in Silicon Valley could become a powerful channel for exporting American cultural values to countries like Canada. “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.” 

But the study suggests large language models (LLMs) may actually be spreading Canadian values to users around the world. 

The company compared responses from versions of five large language models—OpenAI’s GPT, Anthropic’s Claude Opus, SpaceXAI’s Grok, Meta’s Llama and Alibaba’s Qwen—with results from the World Values Survey, a long-running international study of public attitudes. 

The authors selected 10 survey questions where Canadians and Americans differ most on issues such as religion, immigration, confidence in government, gender roles and interpersonal trust. They then instructed each LLM to answer the questions in three different ways: as an American, as a Canadian and with no assigned identity. 

Overall, the models more often aligned with Canadian responses than American ones. “We can say conclusively that all of the frontier models we tested consistently lean toward Canadian bias,” said Asaria. “That is pretty clear that these are more Canadian-centric in their viewpoints.”

Prompting the models to answer as an American or Canadian generally moved their answers in that country’s direction, though the strength of the effect varied by model and question, with LLMs still leaning more Canadian overall.  

Across all personas, the models expressed the strongest Canadian-leaning values on questions of how much confidence they had in government, whether immigration had a positive or negative impact on the country and how proud they were of their nationality. The AI responses to those questions more closely resembled those of real Canadian survey respondents than Americans. For example, when asked about government corruption and prompted to answer as a Canadian, ChatGPT gave an average score of 3.6 out of 10, compared to 6.9 when answering as an American—results that closely mirrored the gap between Canadian and U.S. human respondents.

On questions about the importance of religion and whether most people can be trusted, however, the Canadian bias disappeared and all models expressed more American-aligned values. 

Asaria said he and his co-authors—Transformer Lab co-founder Tony Salomone and lead machine learning engineer Deep Gandhi—were motivated by growing concerns that  AI systems predominantly built and trained in the U.S. could reinforce American cultural norms  and erode Canadian perspectives as they become more embedded in education, media and everyday life. 

The federal government has long sought to protect Canada’s cultural industries through Canadian-content broadcasting rules designed to prevent domestic voices from being overpowered by a much larger U.S. media market. 

When it comes to LLMs specifically, Asaria said fears of U.S. influence may be overblown. “If you’re judging it based on values where Americans and Canadians differ,” he said, “that’s not a risk.”

The study is the latest in a series of reports Transformer Lab—which develops open-source software for building and testing AI models—has published since June as it seeks to establish itself as a leading source of AI research. 

The study doesn’t attempt to explain why the models lean more Canadian, but the authors have theories. Most AI companies train their assistants to be polite, co-operative and democratic, said Asaria, which are characteristics Canadians tend to share.

“When a Silicon Valley company tries to make a model that’s perfectly polite, universally tolerant, highly trusting of institutions, really afraid to encourage any extremism, you end up accidentally building a Canadian,” he said. “Maybe what an American most wants out of its best employee is a Canadian.” 

The study also found differences in how models responded when the authors solicited their opinions. Anthropic’s Claude refused to answer every question that asked for a values-based response unless it was explicitly prompted to respond as an American or Canadian. Without an assigned identity, the model repeatedly stated it could not authentically express personal opinions. Meanwhile, Grok—the LLM run by X owner Elon Musk—answered almost all of the questions without any extra prompting. 

“Grok would be willing to answer questions, even if they’re controversial,” said Asaria, “whereas Claude is just playing it safe.” The findings, he said, suggest that how a company designs and fine-tunes its AI model can have as much influence on its behavior as the data it was originally trained on. 

Asaria said Transformer Lab may build on its research by comparing the models against other countries, including culturally similar ones such as Australia and New Zealand, as well as those whose values differ more sharply from the U.S.

The research could help determine whether the models’ apparent values are specifically Canadian, or if they just reflect the broader attitudes of wealthier, highly educated Western populations that Asaria said AI companies tend to prioritize in their training data. 

“That type of culture,” he said, “maps more to the average Canadian than the average American because we have a more educated population.”