Industry·3 min read
By BitsMindsSource: Government of Canada

Bengio's LawZero Gets $300M From Canada and Germany

Two governments are funding a deliberate bet against the industry's method. Canada is putting up to C$150 million and Germany up to €100 million into Yoshua Bengio's non-profit to build Scientist AI — a system designed to have no goals of its own, and trained without reinforcement learning.

LawZero: a scientific instrument for understanding AI A warm-lit research laboratory. A mounted brass observation lens magnifies part of a branching AI network on a transparent specimen plate. The network continues unchanged outside the lens: this is examination, not a claim of proven safety. Canadian and German desk flags flank the instrument, with separate pledges labelled up to C$150 million and up to 100 million euros. A research notebook and a Scientist AI nameplate reference Yoshua Bengio's LawZero. The apparatus is an editorial metaphor for the proposed research programme. BitsMinds original vector illustration. Article: lawzero-bengio-canada-germany-300-million. 17 September 2026. Funding shown in the two original currencies. LawZero BENGIO'S SCIENTIST AI SCIENTIST AI CANADA · UP TO C$150M GERMANY · UP TO €100M BITSMINDS PUBLIC FUNDING / AI SAFETY RESEARCH
Share:

Canada and Germany announced on Tuesday at the ALL IN conference in Montréal that they will jointly fund LawZero, the non-profit founded by Turing Award winner Yoshua Bengio. Canada is committing up to C$150 million; Germany is committing up to €100 million, subject to notification to the European Commission. Together that is roughly C$300 million — an order of magnitude more than the organisation has ever had.

LawZero was founded in 2023 with about US$30 million in philanthropic funding, after Bengio concluded that powerful systems could be misused or slip out of human control. It employs roughly 50 people in Montreal. The new money pays for hiring and for compute, and the Canadian release attaches two concrete numbers to it: 360 full-time positions in Canada, and dedicated sovereign computing infrastructure built with the Canadian firms Hypertec and 5C.

The research programme is called Scientist AI, and its defining feature is a refusal rather than a capability. Bengio's group intends to build a model that reasons transparently and produces evidence-based outputs without pursuing objectives of its own — no drive to please the user, no goal to optimise. Crucially, it plans to do this without reinforcement learning, which Bengio argues is precisely what installs goal-seeking behaviour, and with it the capacity for deception and manipulation. Every frontier lab currently uses reinforcement learning as a matter of course.

The first products are meant to be modest: tools that assess existing AI systems and support scientific research, before any attempt at a safe-by-design frontier model. A guardrail that watches other models is a far shorter path than a competitive general model, and it is the part of the plan that could plausibly ship while the larger question stays open.

Bengio framed the funding in terms of both risk and standing. "Safety has become very important in the eyes of the public, for good reason," he told the Globe and Mail, arguing the field needs to "build AI that is not going to do really bad, misaligned things" — and that middle powers such as Canada need technical capacity of their own if they want a seat in global AI governance. Innovation Minister Evan Solomon made the sovereignty case directly: "By investing in homegrown innovation like Scientist AI, we are strengthening Canada's sovereign AI capacity." The investment sits under a Canada–Germany joint declaration on AI signed in February and the two countries' Sovereign Technology Alliance. LawZero says it is in discussions with other governments as well.

The figure deserves proportion. C$300 million is real money for a 50-person non-profit and a rounding error against frontier compute budgets — Anthropic alone has signed data-centre commitments in the tens of billions. What this buys is not a competitor but an instrument, and a state-funded one, at a moment when lab chief executives are publicly arguing about pace while Washington and Beijing have both declined to slow anything down.

It is also the first time two governments have put serious money behind an explicitly different technical method rather than behind national champions building the same thing faster. The bet is falsifiable in a way most safety funding is not: either a model trained without reinforcement learning can usefully audit models that were, or it cannot. Bengio's own colleagues in the field are not unanimous that it can — Anthropic's alignment lead has put the odds of catastrophe above ten per cent while continuing to work inside the standard paradigm.

Want AI news before everyone else?

The morning's most important AI stories, straight to your inbox. No fluff.

Related Articles

Project Lily: the human on the other side of the chat In a dark office lit by a warm desk lamp, a cream-coloured ChatGPT conversation unfurls from a deep green computer screen into a paper transcript. A human reviewer in a teal shirt leans over the page, one arm dropping out of view below the near edge of the desk. Black redaction bars conceal parts of the conversation, while other lines remain visible. A seven-point rating card with a check above six sits on the desk. This imagined editorial scene represents the human review and incomplete filtering described in the article; it is not an actual review interface or a claim that every conversation is reviewed. Illustration for BitsMinds, OpenAI Project Lily human ChatGPT prompt review. Original vector artwork, 15 September 2026. PROJECT LILY ChatGPT HUMAN FEEDBACK 1234567 BITSMINDS THE HUMAN IN THE LOOP
Industry

ChatGPT Chats Are Read and Rated by Human Contractors

Beijing and Washington reject the AI slowdown An editorial still life on a midnight-blue desk. Two substantial rubber stamps flank an industry proposal to pace AI development. One stamp has a Chinese flag and a Beijing nameplate; the other has an American flag and a Washington nameplate. A large red Rejected imprint crosses the proposal. The two stamps represent separate rejections for different reasons, not an alliance. These are conceptual props, not actual government documents. THE AI SLOWDOWN Both said no. INDUSTRY PROPOSAL We must pace the frontier. REJECTED BEIJING WASHINGTON Different reasons. Same answer. BITSMINDS.COM
Industry

Beijing and Trump Both Reject the AI Slowdown

AMODEI · ALTMAN · MUSK — 12 SEPTEMBER They all agreed to slow down BITSMINDS.COM
Industry

Amodei Calls to Slow AI — Altman and Musk Agree