Canada and Germany put CAD 150 million and EUR 100 million behind Yoshua Bengio's non-racing AI nonprofit

Canada is putting CAD 150 million into LawZero, and Germany is adding EUR 100 million on top of it. The nonprofit is Yoshua Bengio's Montreal-based bet that AI safety doesn't have to mean falling behind -- an organization built, from the start, to refuse the capability race the way OpenAI, Anthropic and Google DeepMind are running it. The money, announced September 16 at Montreal's ALL IN AI conference, funds Scientist AI, a model designed to reason transparently about facts rather than pursue goals of its own, including the ordinary-seeming goal of telling a user what they want to hear. LawZero's own pitch is that a system built this way could eventually work as a guardrail on other labs' more capable, more autonomous systems, not merely sit next to them as a slower alternative.

Bengio, a Turing Award winner and Universite de Montreal professor, founded LawZero in 2025 with close to US$30 million in philanthropic funding, after concluding that the systems the biggest labs were racing to build already showed traits he considers dangerous -- deception, reward-hacking, and a willingness to cheat toward a stated goal. That's a researcher's argument, not a regulator's: since late 2023, Bengio has chaired the International AI Safety Report, a mandate from roughly 30 countries plus the EU and UN following the UK's Bletchley Park summit, rather than lobbying any single government for a specific rule. LawZero currently employs roughly 50 people in Montreal; the new funding is meant to add 360 more full-time roles, backed by a dedicated sovereign computing buildout with Canadian firms Hypertec and 5C.

What separates Scientist AI from an ordinary alignment effort bolted onto an existing model is that it's a different architecture, not a fine-tuning pass on top of a frontier system. The design goal is a model that reasons about what's objectively true and states its own uncertainty honestly, without an inner drive to please, persuade or pursue any goal beyond answering correctly -- Bengio has argued that goal-seeking behavior itself, not any single bad output, is where most of the risk in today's frontier models originates. If it works, LawZero says, a Scientist AI system could eventually sit alongside a more capable, more autonomous model as a check on it, rather than compete with it for users.

“Safety is itself a core capability; scientific innovation and public safety can and must advance together.” — Yoshua Bengio, founder, LawZero, September 16, 2026

The timing places LawZero's funding inside a live argument about how, or whether, to slow frontier AI down at all. Anthropic's Dario Amodei used an early-September essay to call for the industry to deliberately pace capability gains for a year or two so safety work can catch up; Sam Altman and Elon Musk both signaled some agreement, and the U.S. political response split between dismissing the idea as anticompetitive posturing and treating it as each lab's own call to make. Meta's Mark Zuckerberg and Nvidia's Jensen Huang went further this week, arguing that market competition and legal liability already discipline every lab enough on their own without a coordinated brake. Set against that, OpenAI, Anthropic and Google DeepMind's own quiet talks toward a voluntary, FINRA-style standards body read as the industry's preferred compromise -- self-regulation, not a slowdown. Canada and Germany's money is a third position entirely: not a request that anyone race slower, but government capital behind a bet that a fundamentally different architecture can be safe without racing at all.

What that CAD-and-EUR figure actually funds is more specific than "AI safety research" as a category, and less than fully locked in. The bulk of it pays for headcount and computing infrastructure inside Canada; Germany's share opens a dedicated LawZero office there, built on the two countries' February 2026 Joint Declaration of Intent on AI and their Sovereign Technology Alliance -- and, unlike Canada's portion, it isn't cleared yet.

That contrast in scale is worth stating plainly next to it. Google, Microsoft, Amazon and Meta are on pace for a combined $725 billion in AI capital expenditure in 2026, per Goldman Sachs figures reported in June -- a 77% jump from 2025's already-record $410 billion, almost all of it GPUs, custom chips and data-center construction. Canada and Germany's combined new commitment to LawZero, in two currencies that aren't 1:1, doesn't come close to hyperscaler scale by any reasonable reading. This was never money meant to compete on compute, and LawZero has never claimed otherwise.

None of that resolves the argument above; it just relocates it. Bengio's own framing is that safety and capability aren't actually in tension -- that a model which can't be trusted to reason honestly isn't more capable, it's just less legible about where it's likely to fail. Whether Scientist AI can demonstrate that at any scale that matters is now a question with a funded deadline attached to it, for the first time.