Two national governments have decided that AI safety research is too important to leave to the companies building the AI. On September 16, at the ALL IN conference in Montréal, Canada and Germany announced a joint commitment of up to CAD 300 million to LawZero, the 15-month-old non-profit founded by Turing Award winner Yoshua Bengio — the largest public funding package ever directed at an independent AI safety lab, and the first time Ottawa has put federal money into one.
The split is roughly even. Canada is putting in CAD 150 million through Innovation, Science and Economic Development Canada's Strategic Response Fund. Germany is contributing EUR 100 million — an amount the two governments are treating as the other half of the CAD 300 million headline. Both figures are ceilings, not cheques already cashed: Ottawa's release says the two governments are "planning to invest," LawZero describes it as a "commitment of up to $300M," and the German money is explicitly subject to notification by the European Commission, the state-aid clearance step that has to happen before Berlin can disburse.
What the money buys
LawZero's pitch is a bet against the prevailing design of frontier AI. Its flagship research direction, Scientist AI, is a deliberately non-agentic system: trained on verifiable facts, built to make calibrated probabilistic predictions, and — critically — given no goals of its own. The intended use is as a guardrail, a model that sits alongside an agentic system and predicts whether a proposed action could cause harm. Bengio has argued that deception, sycophancy, reward hacking and self-preservation are emergent properties of goal-directed training, and that a predictor with nothing to want cannot acquire them.
The grants fund two things: people and compute. LawZero will open a Berlin office, where the German money will be spent, and stand up dedicated sovereign compute in Canada with Montréal-based Hypertec and its data-centre arm 5C. Ottawa says the project will create 360 full-time jobs in Canada. LawZero, launched in June 2025 and incubated at Mila, had grown to almost 50 staff by this month, according to The Logic; before these grants it had raised roughly US$35 million in philanthropy from the Gates Foundation, Coefficient Giving, Schmidt Sciences, the Future of Life Institute, NVIDIA and Jaan Tallinn via the Silicon Valley Community Foundation.
"We are getting a commitment of governments that they care about our mission at a time when the world is realizing that this is a really important mission," Bengio told The Logic. "We need to figure out how to build AI that will follow our instructions and not do crazy stuff." In LawZero's statement he framed the announcement as a threshold: "Recent events demonstrate that safety is itself a core capability: scientific innovation and public safety can and must advance together."
Why two governments
The two-flag structure is not incidental. Canada and Germany signed a Joint Declaration of Intent on AI and launched a Sovereign Technology Alliance in February 2026, and Germany was ALL IN's "country of honour" this year; the same week, Toronto's Cohere finalized its merger with Germany's Aleph Alpha. Both capitals used one word repeatedly: sovereignty. Ottawa's release says anchoring the research in Canada will "reduce reliance on foreign-controlled AI systems." Germany's Federal Minister for Digital Transformation, Karsten Wildberger, called Scientist AI "an alternative to potentially uncontrollable autonomous AI agents" and said the project strengthens "our technological sovereignty."
Berlin gets a seat at the design table early enough that, in Ottawa's words, European norms and regulations can be "incorporated from the outset." Canada gets a research asset it named in its own National AI Strategy, AI for All, launched June 4, 2026, which promised to back foundation-model research built with safety in mind and help export it to countries seeking trusted alternatives.
Why it matters
Almost all serious AI safety research today is paid for by the labs racing to build the systems, or by a small cluster of philanthropies. That is a structural problem no amount of good intentions fixes: the people auditing the technology answer, ultimately, to the people shipping it. Two G7 governments writing nine-figure cheques to an organization explicitly structured to be "insulated from market and government pressures" is the first real attempt at a counterweight. Whether it stays a counterweight is the open question — LawZero is now substantially dependent on two treasuries with their own industrial-policy incentives, and Ottawa's framing leans heavily on jobs, compute and sovereignty rather than on safety outcomes.
The technical bet is riskier still. Scientist AI has to catch what a frontier agent is about to do, which means keeping pace with models trained on budgets an order of magnitude larger. LawZero's own formal safety case, published this year, concedes it rests on assumptions rather than guarantees, and does not cover deliberate human misuse or agentic systems built on top of the predictor. Outside observers are supportive but not convinced it is sufficient: Jaxson Khan of the University of Toronto's Munk School told CBC there is "no silver bullet to this problem right now," while calling the grant worthwhile. And two existing funders, the Future of Life Institute and Coefficient Giving, are closely tied to the existential-risk movement — a lineage critics will read as public money underwriting one particular theory of AI harm.
What to watch
Three things. First, whether the European Commission clears Germany's EUR 100 million, and when; until then half the headline number is provisional. Second, milestones — neither government has published what LawZero must deliver to unlock successive tranches, and "up to" is doing a lot of work in both releases. Third, adoption: LawZero plans to sell Scientist AI as a guardrail layer, through model makers or directly to their customers. A safety system nobody deploys is a research paper. The test of this CAD 300 million is not whether Scientist AI works in Montréal, but whether anyone running frontier agents in production agrees to put it in the loop.
“We need to figure out how to build AI that will follow our instructions and not do crazy stuff.”— Yoshua Bengio, Founder and Scientific Director, LawZero