Canada and Germany put CAD 150M and EUR 100M behind Bengio's non-goal-seeking AI
LawZero's "Scientist AI" is explicitly built not to act autonomously or pursue its own objectives -- a monitoring layer, not an agent. Two governments just backed that bet over the industry's default architecture.
The Government of Canada and the Federal Government of Germany announced on September 16, at Montreal's All In AI conference, that they're investing CAD 150 million and EUR 100 million respectively in LawZero, the Montreal-based AI-safety non-profit founded and scientifically led by Turing laureate Yoshua Bengio. Per the official announcement, the money is meant to fund LawZero's build-out of "Scientist AI" and comes bundled with a partnership with Canadian compute providers Hypertec and 5C, plus a commitment to create 360 full-time jobs in Canada. Germany's contribution still requires European Commission notification.
A different architecture, not just more safety research
What makes this notable isn't the dollar amount -- it's what the money is funding. Scientist AI is deliberately built to not act autonomously or pursue objectives of its own; instead of an agent that takes actions in the world, it's designed to reason transparently and produce evidence-based outputs, positioned as a monitoring and guardrail layer that can flag misaligned behavior in other, more agentic systems without itself being trained via reinforcement learning toward some goal. Germany's Federal Minister for Digital Transformation, Karsten Wildberger, called it "a novel approach to develop inherently safe AI without pursuing its own goals." Canada's AI minister, Evan Solomon, tied the funding explicitly to sovereignty: "Canada's future in AI depends on developing advanced systems that reflect our values and protect our interests" -- language that places this alongside the broader Sovereign Technology Alliance the two countries launched in February.
Why this matters beyond the safety framing
Every dollar a government puts into a non-goal-seeking monitoring model is a dollar not going toward a bigger, more autonomous frontier system -- and it's a live test of whether "safety as a core capability," as Bengio put it, can be commercially and technically viable rather than just a research agenda. For buyers weighing whether to deploy increasingly autonomous agents against their own systems, LawZero's bet is a useful counter-model to watch: a purpose-built overseer that never becomes an agent itself is a different risk profile than adding another layer of prompting on top of a general-purpose one.