The Free World Needs Open Source AI: A North American AI Compact for the USMCA Era

Earlier yesterday, I read something on Politico that I decided was worth writing about. Politico published a report on the Trump White House's postponed AI executive order and it was a super fascinating read. There's tons of discussion happening right now in DC in the Trump Administration about Frontier AI capabilities, and those discussions are extremely important. I think the now-cancelled Executive Order should be read as a sign that the United States, Canada, and the broader democratic world are entering a new phase of AI policy, one where older cybersecurity and tech policy categories no longer fit neatly. Frontier AI is no longer only a consumer technology issue, or an innovation issue, or a cybersecurity issue. Frontier AI touches cybersecurity, trade, national security, industrial policy, compute policy, export control, and democratic governance all at once.

I am watching AI policy developments with interest not only as someone who follows the U.S. AI policy landscape closely, but also as a dual U.S./Canadian citizen who believes both countries have an opportunity to get this right. In six weeks, the USMCA review process reaches its July 1, 2026 decision point, and the trilateral trade agreement is imperative for both the US and Canada's commercial success. Additionally, Canada is also expected to release its long-awaited refreshed national AI strategy imminently, according to Prime Minister Mark Carney, who told reporters that the strategy is "coming out next week." Canadian Press reporting also notes that Minister Evan Solomon has said the strategy will consider labor-market impacts, public trust, and the balance between AI optimism and AI skepticism. There may be a tendency to treat these two developments as wholly separate, but I don't think that's the best idea. The US and Canada are trying to answer an extremely difficult question on two different axes: can North America build an AI ecosystem that is open enough to innovate, secure enough to defend, and allied enough to compete with authoritarian powers? My answer is an optimistic yes, but only if we are clear about what Americans and Canadians are trying to protect.

It goes without saying that I want the Free World to win the AI race. The US and Canada have a storied history of robust allyship, ranging from both being founding members of NATO, to free trade partners, to being culturally similar in many regards. I do not want the global AI stack defined by the Chinese Communist Party, authoritarian surveillance systems like those run by the IRGC, or despotic regimes that view human freedom as an inefficiency to be optimized into obsolescence. My ideals are a basic commitment to liberal democracy, civil society, open inquiry, and the imperfect but real freedoms that democratic societies still protect. For those reasons it's imperative to make sure that if the Free World is going to win, it should win as the Free World. That means allies, open research, independent developers, universities, startups, public-interest labs, and open-source communities. It does not mean quietly redefining "AI safety" or "national security" as a permission structure where only the largest closed labs can afford to participate, and it doesn't mean regulatory capture. A democratic AI strategy that treats open-source developers as a problem to be managed has already misunderstood a fundamental value of what a sovereign democracy is trying to defend and protect.

The importance of handling frontier AI in a way that's aligned with the values of the Free World is where today's Politico article becomes important. Dianna Rozi and Sophia Cai describe three ideological factions inside the White House. One camp, associated with former AI Czar David Sacks, favors a lighter-touch approach and worries that AI regulation could slow American companies in the race with China. A second camp, associated with Secretary of War Pete Hegseth and Undersecretary Emil Michael, is more hawkish about advanced "Mythos-type" models and their potential exploitation by rivals. A third middle-ground camp, involving White House Chief of Staff Susie Wiles, Treasury Secretary Scott Bessent, and National Cyber Director Sean Cairncross, reportedly favours a relaxed, voluntary framework in which frontier AI companies give the U.S. government a first look at new models before public release. All three of these factions have value and deserve attention on their own merits. David Sacks is correct that overly burdensome rules can stifle development, Secretary of War Hegseth is correct that Mythos-type frontier models pose risks to cybersecurity and military preparedness, and Chief of Staff Wiles is correct that letting red-teamers from NIST and CAISI look at models before they are released is valuable for research and for promoting social trust.

What I personally find most interesting is that the nearly signed executive order was apparently not a broad licensing regime. The cancelled Executive Order would have created a voluntary oversight system where companies could consult with the U.S. government on their latest models, giving officials a preview before public release without a formal mandate. One senior official described it as "not mandatory," while also saying that the major companies had agreed to abide by it. It appears that Susie Wiles' opinion was the one that made it into the near-final draft, and the voluntary distinction matters. If a voluntary first-look framework was enough to trigger last-minute alarm, then the industry red line may not be mandatory licensing; it might be a formalized pre-release checkpoint to begin with.

I understand and empathize with the concern of Big Tech here. Vague pre-release review can easily become de facto licensing or a shadow veto system. It can even entrench incumbents and foment regulatory capture, creating an authorized channel to government while leaving open-source developers, smaller labs, and academic researchers outside the room. It can turn "safety" into a moat. That risk is real, and anyone who cares about open-source AI should take it seriously. Nevertheless, the opposite extreme is not a serious outcome either. Politico's reporting says the administration's timeline accelerated after Anthropic's release of Mythos, described as a highly advanced model capable of finding and exploiting unknown flaws in IT systems. That detail shifts the debate out of abstract "AI safety" and into concrete cybersecurity territory, akin to the Clinton administration's handling of Y2K. A model that can meaningfully discover and exploit unknown vulnerabilities is not just a conversational chatbot, and policymakers shouldn't legislate it as such, as it falls into a much more serious domain of infrastructure. A pro-AI position cannot simply mean "ship everything and hope incentives work out." Sure, it might make the race against the Chinese Communist Party go faster, but it also pours proverbial gasoline on genuine zero-day vulnerabilities and prevents governments and civilians alike from fortifying critical infrastructure. The hard question the Trump administration needs to ask itself is not whether governments should understand dangerous frontier capabilities before they reshape the threat environment; they absolutely should. The question is how to ensure that understanding does not become a closed syndicate between the government and a handful of frontier labs.

In my opinion, the strategic deficit posed by the advent of Mythos is where the open-source AI community has to be more precise than both AI doomers and e/acc-style accelerationists. Open-source AI researchers should reject broad AI licensing regimes. We should not build bureaucratic systems that treat every independent model developer as a threat vector, nor should we accept public-private "safety" processes that quietly privilege incumbents at the expense of smaller researchers. At the same time, we should also not pretend that national-security concerns disappear just because some of the people raising them are politically inconvenient. I think the right answer is a narrow, fast, technically competent oversight regime for genuinely high-risk capabilities, paired with explicit civil and criminal safe-harbor protection for open research and open-source development. This means that if a small research team or even a solo open-source developer wants to throw their white hat at Claude Mythos or GPT 5.5 Cyber, they should be allowed to, with the full protection of the law. Mandatory disclosures are commonplace for white-hat cybersecurity, and many participants in this field have cordial relations with law enforcement. It should go without saying that my preferred version here is not a general AI permission slip or a blank check for bureaucrats.

A narrow oversight process should focus on frontier systems with demonstrated cyber, biosecurity, autonomous agentic, military, or critical infrastructure capabilities. Another key thing that should be included is persuasion and sycophancy risk, because the literature is vast and deserves more analysis. Pre-release testing should have fixed timelines, clear thresholds, secure model-handling rules, and public, plain-language transparency about the vetting process without disclosing exploit-level details. Most importantly, this framework should not treat openness as the enemy of security. Open-source AI is one of the Free World's core strategic advantages, as it makes models more auditable. Open-source distributes technical capacity, lowers entry barriers, and helps people learn how bigger frontier models work at a smaller scale. Open-source helps researchers, civil society, journalists, startups, educators, small businesses, and even ordinary civilians in their homes understand and adapt to a technology that would otherwise be mediated entirely through closed corporate APIs. Open models are not automatically safe, but closed models are not automatically trustworthy. A truly democratic AI policy should not confuse opacity with responsibility.

That is why the USMCA review matters. The USMCA's July 1, 2026 decision point is not a deadline for completing every negotiation, but it is the point at which the US, Canada, and Mexico must decide whether to extend the agreement for another sixteen years. If any party declines, the agreement enters annual reviews and, absent resolution, expires in 2036. CSIS has explained this mechanism clearly: July 1 is the point at which the clock either resets or starts running down, thereby creating a strategic opening. North America can either treat AI as an afterthought beneath cars, steel, tariffs, and agriculture, or it can recognize that compute, data, cybersecurity, and model governance are now central to North American continental competitiveness. The existing digital trade chapter should be preserved, not casually reopened. CSIS has argued that Chapter 19's commitments on cross-border data flows, data localization, and source code protection remain the legal backbone of North America's digital economy, and that reopening them could create damaging uncertainty. Brookings has similarly framed the 2026 review as an opportunity to modernize digital trade and strengthen regional cooperation on cybersecurity, AI, data governance, and SME inclusion. I truly believe that augmenting Chapter 19 is the right foundation for the US, Canada, and Mexico to build upon. It would be worthwhile for trade negotiators to add a Frontier AI and cybersecurity side letter, annex, or parallel compact that recognizes the world has changed since USMCA was negotiated. The USMCA review should be used to start building a North American AI and cybersecurity compact. Not a grand new bureaucracy, not an AI trade war, and not a closed club for the largest labs, but a practical framework for keeping the democratic AI ecosystem open, secure, and competitive.

I am not an expert, far from it, but I have been involved in AI policy and governance conversations since the spring of 2022. That's roughly four years of learning on the job about how these technologies develop, adapt, and get implemented worldwide. I think that for USMCA discussions focusing on frontier AI, each nation should consider four core facts during their discussions. I believe each of my suggestions is important and deserves attention and scrupulous care.

1. Data: Current AI models run on cross-border infrastructure, ranging from cloud services to enterprise data, research collaboration, digital trade, and the ordinary movement of information between firms, universities, governments, and users. Canada is right to care about sovereignty, privacy, public-sector dependence, and the legal status of Canadian data once it enters foreign systems. That concern shouldn't turn into data nationalism, and Minister Evan Solomon is correct that sovereignty is not solitude. US and Canadian markets need trusted interoperability: rules that let data move when it should, protect it when it must, and give both countries confidence and reassurance that sensitive public and commercial information is not being casually absorbed into someone else's jurisdictional black box.
2. Compute: Canada's push to build large-scale domestic AI supercomputing capacity should not be read as anti-American, even though some in the US might consider it to be. I'm personally quite excited by it, as it means Canada can use some of its own talent here and develop powerful tools we can rely on, paired with Canadian values. I think a better framing here is allied resilience. Canada does not need to become a branch-plant digital economy, and the United States should not treat Canadian capacity as a threat, but as a boon to its own interests. A stronger Canadian compute base makes North America stronger, especially if Canadian compute is tied into trusted supply chains, shared research infrastructure, clean energy, Five Eyes intelligence, and secure cloud standards. Sovereignty through alliance is not a slogan; it is the only workable middle ground between going it alone as a superpower like the US or China, and complete isolation from common markets.
3. Cybersecurity preparedness: If frontier models like Claude Mythos or GPT 5.5 Cyber can exponentially accelerate vulnerability discovery, exploit generation, or automated offensive workflows, then model policy and cybersecurity policy cannot be separated between the two countries. A North American framework should make it completely seamless for the United States and Canada to coordinate vulnerability disclosure, critical infrastructure exercises, incident reporting, patching support, and secure information-sharing around AI-enabled threats. A US/Canada frontier AI communication network is practical and would let both countries defend each other's hospitals, banks, telecom networks, energy infrastructure, transportation systems, and public institutions in a world where cyber capability is becoming easier to scale. Even with heightened US/Canada tensions over trade (which I am not going to go into in this piece), the US and Canada are already coupled deeply in core domains; for example, my home province of Quebec exports hydroelectricity to my birth state of New York. Quebec and New York State have shared their electrical grid for over a century, and power plants are critical infrastructure that need robust protections. My proposed US/Canada frontier AI comms line would help ensure that Hydro-Québec can safely and cheaply export clean power to millions of New Yorkers, and that's only one peaceful use case.
4. Evaluation: In my opinion, the United States and Canada do not need broad licensing regimes for AI, like the EU has with its AI Act. What both nations do need is a shared glossary for genuinely high-risk frontier capabilities. If a model crosses meaningful thresholds in cyber exploitation, biosecurity assistance, autonomous replication, critical infrastructure attack planning, or other national-security-relevant domains, both governments should be able to evaluate that fact using technical evidence rather than policy vibes or political gamesmanship. Those standards should be narrow, public enough to be accountable, and precise enough that "safety" does not become whatever a minister, lobbyist, or CEO says it is on a bad day. I think this is incredibly important in the wake of Canadian AI Minister Evan Solomon's forthcoming AI report next week, and it should be seriously considered once he and Heritage Minister Marc Miller ultimately table their forthcoming Online Harms Act. One major elephant in the room for the USMCA is almost certainly going to be online safety for youth, and we're already seeing the tension at G7 levels. Canada's forthcoming legislation, no matter how well intended, could have the effect of throwing a proverbial wrench into serious trade negotiations in six weeks, and that would be an own-goal.

The Free World argument for open-source AI research involves security without cartelization, openness without naivety, and sovereignty through alliance rather than isolation. I use the phrase "Free World" intentionally, even though I know it sounds old-fashioned. I grew up politically shaped by a pre-2016 Republican vocabulary of allies, national security, markets, democratic values, and institutional seriousness. Think less online populism and more Secretaries Condoleezza Rice or Colin Powell. The institutionalist Republican tradition doesn't really exist in the same way anymore, but it shaped a core part of my worldview as I learned about politics in my adolescence; the closest analogy I can think of is Ambassador Nikki Haley, who served in President Trump's first term as UN Ambassador. In my pre-2016 worldview, allies are not accessories but force multipliers. Markets are not magic, but they are powerful engines of innovation when paired with credible institutions. National security concerns are not an excuse for every power grab, but they are also not fictitious. Technology policy is not only about who ships fastest. It is about what kind of world the technology scales into. Democracy is imperfect, but it is always worth defending. That is why I do not want China to win the AI race. Authoritarian governance models cannot be allowed to define the infrastructure through which knowledge, speech, work, education, surveillance, and state power are increasingly mediated. The most powerful AI systems in the world should not and cannot be optimized around censorship, social control, and regime security. Third World countries should not be forced to choose between closed authoritarian stacks and closed corporate stacks. For these reasons, I do not want democratic AI leadership reduced to a few Big Tech companies behind opaque APIs.

The open-source community has to make a stronger democratic argument. The Free World wins not only by having the most capable models, but by having the broadest base of capable people. It wins when students can learn, when researchers can inspect, and when small companies can build. Open-source wins when civil society can audit, and when public institutions have enough technical capacity not to be bullied by either vendors or panic. Open-source wins when allied countries share enough infrastructure to be resilient but retain enough sovereignty to be democratic. Canada's upcoming AI strategy next week will prove to be a significant test. Ottawa's consultation process has emphasized themes such as democratic values, risk-based governance, sovereign infrastructure, intellectual property protection, AI literacy, security frameworks, and public trust, according to a summary of Canada's national AI strategy consultation. Those are the right ingredients, but it's too early to see what the execution of those ingredients will look like.

Canada has world-class AI research institutions, like Mila, Vector, and Amii. Canada has talent, immense energy potential, and geographic and cultural proximity to the United States. Canada has unique immigration advantages, and Canada has an opportunity to become more than a place where ideas are born and then commercialized somewhere else. Sadly, even with all of the major benefits Canadians share, Canada also has a familiar problem: Canadians often produce research excellence and then lose ownership, compute scale, commercialization, or procurement momentum. Minister Solomon's refreshed AI strategy should be judged by whether it answers a simple question: can Canadians turn AI talent into durable Canadian capacity while remaining integrated with allied markets? For the current Carney government, the plan likely means sovereign compute, but not complete solitude from allies. Canadian data protections, but not isolation from global markets. Domestic commercialization, but not protectionist stagnation. AI safety, but not innovation paralysis. Partnership with the United States, but not dependency on American corporate infrastructure for every critical workload. Canada shouldn't repeat the EU's mistakes with its AI Act implementation. Brussels rolled out full protectionism before it had domestic competition, and then the codified protectionism measures scared away domestic innovation.

The United States faces the mirror-image problem. The US has the largest frontier labs, the deepest capital markets, and the strongest compute ecosystem. It also has a political system that can swing between techno-libertarian denial and panic-driven overreach. The Politico article captures that instability quite well within the Trump administration. The White House appears to be trying to respond to real national-security concerns while also managing industry pressure and strategic competition with China. The result is a postponed executive order, internal factionalism, and an unresolved question about whether voluntary model review is prudent governance or the first step toward bureaucratic choke points. Michael Kratsios, as Director of the White House Office of Science and Technology Policy, sits directly in the middle of the U.S. science and technology policy agenda, including AI and strategic competition with China. Evan Solomon, as Canada's Minister of Artificial Intelligence and Digital Innovation, is trying to define Canada's AI strategy at a moment when sovereignty, safety, jobs, compute, and commercialization are all live questions. In my opinion, Kratsios and Solomon should be reading the same strategic map, sharing the same coffee, and collaborating on a US/Canada bilateral AI compact as part of USMCA negotiations.

A North American AI compact needs to make three things absolutely clear.

1. The United States and Canada should treat allied AI leadership as a shared strategic priority. The Free World is stronger when Canada is technologically capable, not when it is merely dependent, just as the United States is stronger when her allies have resilient infrastructure, skilled workforces, and trusted institutions. Canada is stronger when it has access to U.S. scale, markets, investment, and frontier expertise without surrendering its own sovereignty. All of these things can be true at once, and have been true for a large part of both nations' histories.
2. Open-source AI should be recognized as part of democratic resilience and an asset for both nations. Not every model is going to be open; there are intellectual property concerns that need scrutiny. Not every capability will get widescale release like a mainstream LLM product such as ChatGPT or Google Gemini. Nevertheless, open research and open-weight development are essential to auditability, education, competition, and broad-based innovation. Any bilateral framework that treats openness as inherently suspicious will tilt power toward closed incumbents and away from the public.
3. High-risk frontier capabilities (like Claude Mythos or GPT 5.5 Cyber) require narrow, serious security processes. Cyber-capable models and biosecurity-relevant technology are not ordinary software, and neither are autonomous agent systems that can operate across tools, networks, and institutions. A North American compact should define the truly high-risk categories and create fast, technical, accountable mechanisms to evaluate them. That process should be strong enough to matter and narrow enough not to swallow the ecosystem.

My three points for a North American AI compact are harder to explain than the two binaries of "regulate AI" or "never regulate AI." Still, a nuanced and well-thought-out position is the one democratic societies need. AI accelerationists are right that bad regulation can kill innovation, and that beating the Chinese Communist Party to AGI matters. The safety hawks are also right that some frontier capabilities create real national-security risk, and that cyber automation changes conventional threat models. The open-source community is also right that the future cannot belong only to closed labs, with regulatory capture and the risk of safety becoming a moat. I believe the most statesmanlike answer is to hold all three truths at once.

I want the Free World to win the AI race. I want America to lead, as much as I want Canada to be more than a branch office. I want open-source AI to thrive just as much as I want authoritarian surveillance states to lose the normative battle over what AI is for. I want frontier labs to innovate in conjunction with governments that have built enough technical competence to evaluate real risks. I want small developers to be able to build without asking permission from Washington, Ottawa, or Silicon Valley. Finding the balance is why AI policy and governance research is so important. The next few weeks are therefore unusually critical. The Canadian AI strategy will tell us how Ottawa understands sovereignty, compute, adoption, safety, and commercialization. The USMCA review will tell us whether North America can modernize its digital trade architecture without blowing up the foundations that made cross-border digital commerce work. President Trump's postponed AI order will tell us whether the United States can build serious AI security processes without turning them into either nothing or licensing-by-vibes (and whether an Executive Order will even be signed in the first place). It would be a critical mistake to treat every competing variable as a separate story. They're the same story, because AI policy is becoming alliance policy. Compute policy is becoming trade and national security policy. Cybersecurity policy is becoming frontier model policy and third-party audit governance. Conventional labels don't make sense anymore, and that's the point with novel technologies.

I think the United States and Canada should use the USMCA renegotiation moment to begin building a North American AI and Cybersecurity Compact. The compact should aim to preserve digital trade, strengthen trusted data flows, coordinate high-risk model evaluation, expand allied compute capacity, support SMEs, protect open-source research, and treat authoritarian AI dominance as a strategic challenge without imitating authoritarian centralization. The Free World should win the AI race as the Free World: open, allied, technically serious, and secure without becoming closed.