A grant, not a mandate
On September 16, 2026, at the ALL IN conference in Montreal, Canada and Germany announced up to CAD $300 million in joint funding for LawZero, the Montreal non-profit founded by Turing Award laureate Yoshua Bengio. Canada is contributing up to $150 million through its Strategic Response Fund; Germany is matching with up to €100 million. The money funds LawZero's flagship project, "Scientist AI" — a safety-by-design system meant to monitor and constrain more autonomous, agentic AI models — plus a new sovereign compute build-out in Canada with data-centre partners Hypertec and 5C, and a new LawZero office in Berlin (Canada.ca). Ottawa projects roughly 360 Canadian jobs from the investment; LawZero itself currently employs around 50 people (The Globe and Mail).
What makes this notable isn't the dollar figure — it's the mechanism. Canada tried the legislative route to AI governance once already, and it failed. Bill C-27's Artificial Intelligence and Data Act would have created an AI and Data Commissioner with authority over "high-impact" systems (ISED companion document). It spent over two years in committee, was criticized for vague high-impact definitions and a commissioner who reported to a minister rather than acting independently, and died on the order paper when Parliament was prorogued in January 2025 (McInnes Cooper). Canada has had no comprehensive federal AI statute since. The LawZero grant is the clearest sign yet that Ottawa isn't planning to revive one soon — it's substituting direct public investment in safety research and domestic compute for binding rules on deployment.
The case AIDA's defenders would make
It's worth stating that case fairly before dismissing it. A grant to one non-profit, however well-regarded its founder, is not a law. It creates no enforceable obligations for the commercial labs — OpenAI, Google DeepMind, Meta, xAI — actually shipping the frontier systems Canadians interact with daily. If Scientist AI succeeds as a monitoring layer, nothing compels any company to adopt it. And unlike a statute, a grant is politically contingent: it can be trimmed, redirected, or quietly left to lapse in the next budget cycle in a way an act of Parliament, with its own commissioner and enforcement docket, cannot. Canadians harmed by biased or unsafe deployed AI still have no dedicated regulator or statutory redress mechanism — precisely the gap AIDA, whatever its flaws, was built to close. Funding a safety research lab is a bet that better tools will diffuse into industry practice voluntarily; it is not, by itself, governance.
Why the funding route is still the better bet right now
That critique lands hardest against a counterfactual where AIDA-as-written was actually a good bill. It wasn't. Every specific failure mode critics identified — indeterminate "high-impact system" thresholds, a commissioner without independence comparable to the Privacy Commissioner, exposure to constitutional challenge over federal reach into university research, and no carve-out for pure R&D — was a drafting problem, not an argument against AI governance in the abstract (McInnes Cooper). Rushing a second AIDA through Parliament to fill the vacuum quickly would risk repeating those same mistakes under a new government with different priorities.
Direct funding for measurable capability — a monitoring system, a compute buildout Canada actually owns — sidesteps that failure mode entirely because it doesn't require Ottawa to first solve the hard definitional problem of what counts as "high-impact" AI. It also doesn't force one theory of AI risk onto Canada's AI sector. That plurality matters domestically: Cohere CEO Aidan Gomez has publicly dismissed coordinated-slowdown advocacy — the instinct underlying much of Bengio's public argument for Scientist AI — as amounting to "cartel" behaviour among incumbents (The Globe and Mail). A regulator empowered to pick a single compliance framework would have to adjudicate that disagreement; a grant program can fund one serious approach without foreclosing others.
Germany's participation reinforces the point. Berlin operates under the EU AI Act, arguably the world's most extensive binding AI statute, and chose to co-fund LawZero anyway rather than treat Brussels' rulebook as sufficient. That's a tell: even inside a jurisdiction with hard law on the books, government money for safety research and sovereign compute is being treated as a complement to regulation, not a substitute the way Canada is currently using it by necessity.
The honest read
This is industrial policy dressed as safety policy, and both descriptions are accurate — Ottawa gets a stake in frontier safety research and Hypertec/5C get a subsidized compute customer, while genuinely useful monitoring tooling may emerge either way. It is not a substitute for eventually legislating narrow, enforceable rules for the highest-risk deployed uses of AI in Canada — government services, hiring, credit, policing — the exact places AIDA's critics said Bill C-27 got the scope wrong in the other direction. But as an interim step while Parliament figures out what a better AIDA would even look like, funding a credible safety lab and building sovereign compute capacity is a more defensible use of $150 million than another two years in committee.