LAWZERO

Institutional Archetype

THE LAB THAT BUILDS THE GUARDRAIL MODEL — LawZero is a Montréal nonprofit that does not test other people’s models. It is trying to build a different one: a “Scientist AI” that predicts and explains and, by design, wants nothing. Its pitch is that such a model could sit over the agentic systems the labs are shipping and block what they do. Its money comes from philanthropies that also fund much of the evaluator layer, from NVIDIA, from the Gates Foundation, and, since September 2026, from the governments of Canada and Germany. The file is open because the organization that says the safe alternative should be built outside market and government pressure has just taken a state commitment of up to CAD 300 million, and its founder sits on the Prime Minister’s AI council.

Mandate & Origin

“LawZero is a nonprofit startup developing technical solutions for highly-capable, safe-by-design AI systems” (LawZero). It was launched on 3 June 2025 by Yoshua Bengio: “LawZero is structured as a nonprofit organization to ensure it is insulated from market and government pressures, which risk compromising AI safety.” The launch release says it was “founded in response to evidence that today’s frontier AI models are developing dangerous capabilities and behaviours, including deception, self-preservation, and goal misalignment,” that it had “a growing technical team of over 15 researchers,” and that “LawZero was incubated at Mila - Quebec AI Institute, a non-profit founded by Professor Bengio. Mila now serves as LawZero’s operating partner” (LawZero, launch).

Its team page lists the leadership as Yoshua Bengio, Co-President and Scientific Director; Sam Ramadori, Co-President and Executive Director; Maggie Da Prato, Vice-President, Talent; Justine Gauthier, Vice-President, Legal & General Counsel; Joumana Ghosn, Senior Director, Research; and Catherine Saine, Vice-President, Development & Operations. The board is Bengio, Maria Eitel (Chair), Valerie Pisano, Yuval Noah Harari and Sir John Rose; the Global Advisory Council is Stefan Löfven and Jacinda Ardern; the Scientific Advisors are Beth Barnes, Michael Cohen, David Duvenaud, Gauthier Gidel, Noah Goodman, Geoffrey Irving, Chris Pal and Jacob Steinhardt (LawZero, team). The board announced on 15 January 2026 also named Mariano-Florentino Cuéllar, president of the Carnegie Endowment for International Peace; the team page now lists five directors without him (LawZero, board appointments; LawZero, Ardern appointment). Of the advisory council it says: “Global Advisory Council positions are voluntary and unremunerated.” The board’s stated priorities for the year ahead “will include stewarding LawZero’s fundraising efforts and compute strategy, and helping grow its global presence.”

Funding & Backers

LawZero’s home page carries a donors block under the line “Thanks to the support of generous donors, we launched a major nonprofit AI safety startup, insulated from market pressures that often prioritize speed of deployment over safety and trustworthiness.” It shows logos with captions: the Future of Life Institute, captioned “A donor advised fund of Silicon Valley Community Foundation”; the Gates Foundation; Founders Pledge, “Made possible by the generosity of Jaan Tallinn and Founders Pledge Inc. (US)”; Coefficient Giving; Schmidt Sciences; the Silicon Valley Community Foundation, “Made possible by the generosity of Jaan Tallinn and Survival and Flourishing DAF, a donor advised fund of SVCF”; and NVIDIA. No amounts are published per donor (LawZero). The launch release names the first donors: “Future of Life Institute, Jaan Tallinn, Open Philanthropy, Schmidt Sciences, and Silicon Valley Community Foundation have made donations to the project as part of its incubation phase” (LawZero, launch).

• The Gates Foundation grant. LawZero’s release of August 2025: “This latest grant brings the total amount raised by LawZero to over $35M USD” (LawZero, Gates grant). The Foundation’s own record lists LawZero, Montréal, a committed amount of $4,995,896, Global Health division, July 2025, 35 months, with the purpose “To support the development of safe AI to apply to key drug discovery challenges for diseases that impact LMICs” (Gates Foundation, committed grants).
• NVIDIA. In July 2026 LawZero “announced it has received a donation of computing resources from NVIDIA” to support Scientist AI, with Ramadori saying NVIDIA’s contribution “will be instrumental in advancing the fundamental science behind the Scientist AI” (LawZero, NVIDIA).
• Canada and Germany. On 16 September 2026 the two governments “announced funding of up to CAD $300 million for LawZero to develop a fundamentally new form of advanced, safe and capable AI.” Canada’s own release says the two governments are “planning to invest CAD 150 million and EUR 100 million, respectively,” that “Canada’s funding through the Strategic Response Fund will support research and engineering talent,” that “The project will create 360 full-time jobs in Canada,” and that “The funding from Germany is subject to notification by the European Commission” (LawZero, release; Government of Canada, ISED). The same release describes the backing as “catalytic funding from philanthropy, including the Gates Foundation, Coefficient Giving, Schmidt Sciences, the Future of Life Institute, NVIDIA, and Jaan Tallinn via the Silicon Valley Community Foundation, and from the governments of Canada and Germany.”
• Grant listings that do not carry it. The Survival and Flourishing Fund’s 2024, 2025 and 2026 recommendation tables do not list LawZero or Mila, so the Tallinn money LawZero names reaches it through the donor-advised-fund channel its page describes and not through an SFF round (SFF 2024; SFF 2025; SFF 2026). Coefficient Giving’s grant feed lists no LawZero grant; its only LawZero hit is a staff page (Coefficient Giving, grants search). See The EA to AI-Safety Funding Pipeline.

Institutional Voice & Intent

LawZero’s own account of the technical bet: “The Scientist AI is inspired by an ideal scientist: a mind that has internalized the laws of nature and uses them to make predictions, but without predilection about how things unfold” (LawZero, research). Its September release says Scientist AI “is being developed to reason transparently and provide reliable, evidence-based outputs unbiased by goals of its own” and “is expected to provide guardrails for today’s frontier models, accelerate scientific breakthroughs, and advance understanding of the risks posed by AI and how to avoid them” (LawZero, release). Bengio’s founding statement: the lab is “the result of the new scientific direction I undertook in 2023, after recognizing the rapid progress made by private labs toward Artificial General Intelligence and beyond” (LawZero, launch).

Stated intent: Build non-agentic, safe-by-design AI as a global public good, insulated from commercial and state pressure, and use it to oversee agentic systems and to speed science.

Observed intent: Build the model with a mix of philanthropic and state money, and place its compute on Canadian soil. The September release says LawZero “will establish a dedicated, sovereign compute infrastructure in Canada, in partnership with Canadian partners Hypertec and 5C,” and open a Berlin office; the German minister’s quote frames the project as strengthening “our technological sovereignty” (LawZero, release). The July 2026 safety-case paper, “Safety from Honesty in a Disinterested AI Predictor,” is the first formal argument that the model is safe by construction, with authors from LawZero, Université de Montréal, Mila, UC Berkeley, McGill, Arb Research, Charles University, Oxford and the Sapien Institute (LawZero, research).

Gap: The structure is stated and public: nonprofit form, a named donor list, a governance council, two named governments. What is not published is the split. No page gives the amount per philanthropic donor, the repayment terms of the Strategic Response Fund money, or the price of the compute from Hypertec and 5C. The founding claim, insulation “from market and government pressures,” is made by an organization that has since taken both philanthropic money from one network and a government commitment. The record raises the question and cannot settle it. For the apparatus reading, it does not need to.

Interest: LawZero argues that agentic AI is the risk and a non-agentic, goal-free model is the remedy, and that advanced AI should be built on sovereign infrastructure. What adoption would pay it is set out from its own pages; no fee or contract value for any product is published.

• A non-agentic model as the guardrail layer. Position: Scientist AIs “are non-agentic and primarily learn to understand the world rather than act in it,” and could “provide oversight for agentic AI systems” (LawZero, launch). What it pays: if developers or regulators adopt a monitoring model of this kind, the only design on offer is the one LawZero says it is building, and the CAD 300 million already committed pays for building it. Best case, Serban’s: Scientist AI “will more accurately and honestly evaluate whether their actions and responses may cause harm and, if so, block them” (LawZero, safety case). Convergence: the labs gain an external monitor they can point to; the evaluator layer that Bengio advises, including METR and Transluce, gains a model-level counterpart to its tests.
• Sovereign AI capacity. Position: Canada’s minister says the investment strengthens “Canada’s sovereign AI capacity”; Bengio says the backing moves LawZero into the next phase “while improving both nations’ technological sovereignty” (LawZero, release). What it pays: the commitment funds jobs, a Berlin office and a compute facility, and Canada’s AI strategy names “building sovereign AI infrastructure” as an object of its new national council, on which Bengio sits (Prime Minister of Canada). Best case, Germany’s: the safe-by-design model “offers an alternative to potentially uncontrollable autonomous AI agents.” Convergence: two governments get a domestic AI program; the compute suppliers Hypertec and 5C, whose chief executives are quoted in the release, get the facility; NVIDIA’s donated compute sits in the same supply chain.

Interest is a documented fact about each position, not a claim about motive, and not a verdict that the argument is wrong. Each best case is stated at the strength its author states it.

Position in the Apparatus

• Founder with seats on the evaluation layer. Bengio is listed as an adviser to METR and to Transluce (METR; Transluce), and LawZero’s Scientific Advisors include METR’s founder Beth Barnes and Transluce’s CEO Jacob Steinhardt: the team page describes Barnes as “Founder & CEO, METR” and Steinhardt as “CEO & Cofounder, Transluce.” Both evaluators are therefore tied to the lab through the same person, and the lab’s guardrail model is the thing they would be testing. See Evaluator Alliances.
• A former UK AISI chief scientist on the advisory list. The team page lists Geoffrey Irving under “Chief Scientist, UK AI Security Institute”; his page there reads “Geoffrey Irving is Co-founder and Chief Scientist at Resolution. Previously Geoffrey was Chief Scientist at the UK AI Security Institute” (LawZero, Irving). The government institute that tests the labs and the nonprofit that proposes a guardrail model share an adviser.
• Staff moves. David Duvenaud: “He formerly led the Alignment Evaluations team at Anthropic” (LawZero, Duvenaud). Noah Goodman: “Since 2024, Professor Goodman is also a Research Scientist at Google DeepMind” (LawZero, Goodman). Barnes: at OpenAI “she helped OpenAI develop safety targets” (LawZero, Barnes). Steinhardt “did a postdoc at OpenAI” (LawZero, Steinhardt). Matt MacDermott, a research scientist at LawZero, moved to the funder: “Matt joined Coefficient Giving in January 2026. Previously, he was a research scientist at LawZero” (Coefficient Giving). The executive director came from outside the labs: “Prior to joining LawZero, he was the CEO of BrainBox AI” (LawZero, Ramadori), and the research director “was a Senior Director of Applied Machine Learning (ML) Research at Mila” and did a PhD “under the supervision of Yoshua Bengio” (LawZero, Ghosn).
• Size. In June 2026 LawZero “currently has a team of approximately 25+ researchers” and “aims to grow its research team to approximately 70 members by the end of 2026” (LawZero, ICML).
• Publications. The research page lists “Superintelligent Agents Pose Catastrophic Risks: Can Scientist AI Offer a Safer Path?” (24 February 2025), “The Scientist AI: Safe by Design, by Not Desiring” (5 February 2026) and “Safety from Honesty in a Disinterested AI Predictor” (2 July 2026) (LawZero, research).

Actions & Leadership Choices

CONDUCT: WATCH — INDEPENDENT BY STRUCTURE, FUNDED BY ONE NETWORK AND TWO STATES. No wrongdoing is alleged. LawZero published its donor list and its governance, took a mixed funding base and says so. The structural point is the one it shares with the evaluators on whose boards its founder sits: the money comes from a small set of philanthropies that also fund much of the field, and from 2026 from states with an interest in the product being built at home. The finding is the arrangement, not the intent.

• 3 June 2025. Launch, with the Scientist AI program and the first donors (LawZero, launch).
• July and August 2025. Gates Foundation grant committed, announced as bringing the total raised to over $35 million (Gates Foundation; LawZero).
• 15 January 2026. Names its board and first Global Advisory Council member, Stefan Löfven (LawZero).
• 5 March 2026. Adds Jacinda Ardern to the council (LawZero).
• 2 July 2026. Publishes the Scientist AI safety-case paper (LawZero).
• July 2026. Receives a donation of computing resources from NVIDIA (LawZero).
• 16 September 2026. Canada and Germany announce up to CAD 300 million (Government of Canada).
• 2 October 2026. Bengio is named to the Prime Minister’s National Council on Artificial Intelligence (Prime Minister of Canada).

Alliance Map

Where ties on different axes run at the same time, the overlap is the finding. The lab whose founder advises METR and Transluce counts METR’s and Transluce’s chiefs among its scientific advisers; several of its philanthropic backers also fund METR and FAR.AI (see METR and FAR.AI); it takes NVIDIA’s compute and Canadian and German public money for a Canadian-sited compute facility; and its founder sits on the council advising the Canadian government on “building sovereign AI infrastructure.” The documents show disclosed arrangements, not coordination. Quebec’s provincial government is not a documented funder: the Quebec-regions agency named in the release is the federal Canada Economic Development for Quebec Regions, whose minister is quoted.

Reach Assessment

Institutional: A founder with seats on the evaluation layer and the government layer, two governments’ money, and a program whose output, if it works, would be installed as a monitor over other labs’ models. Memetic: “Scientist AI” and “implicit agency” are LawZero’s coinages, argued in a TED talk and a Bloomberg interview listed on its newsroom and a blog series (LawZero, newsroom; LawZero, blog).

Civilizational: If a goal-free guardrail model becomes the accepted answer to agentic AI, the people who specify it sit upstream of what other models are allowed to do, with a governance council of former heads of government and a donor base of a few philanthropies and two states. LawZero’s pages say the right things about that: a published donor list, an unremunerated council, a nonprofit form chosen to avoid commercial pressure. The state commitment narrows the distance between a nonprofit’s model and national industrial policy: Canada’s own release describes jobs and sovereign compute as outcomes of the investment, and the council advising on sovereign AI infrastructure includes the founder.

Score basis. Atk 5 is reach: LawZero describes Scientist AI as “expected to provide guardrails for today’s frontier models,” Canada and Germany have committed funding for it, and Canada’s national AI council names its founder (LawZero, release; Prime Minister of Canada). The points under a higher number are that nothing requires any lab or regulator to run the model, that no model is released and that no lab release document found names LawZero. Peers on atk 5: Epoch AI, a public audience without a deployment hook; Transluce, Redwood Research and Alignment Research Center hold 6 on lab and government release documents that cite their findings or protocols. Def 6 is funding insulation: the nonprofit form is stated as a defence against “market and government pressures,” no lab pays it and the donor list is published (LawZero, launch; LawZero). The points off are that several named philanthropic donors also fund METR and FAR.AI, that up to CAD 300 million now comes from two governments, that NVIDIA donated the compute, and that no per-donor amount is published. Peers on def 6: FAR.AI and Palisade Research, on one philanthropic network supplying most of the money; Transluce and Alignment Research Center hold 7 on a published revenue split and a policy barring funder control, and Redwood Research holds 8 on a stated policy of no payment from the labs it advises. Hp 6 is entrenchment: a commitment of CAD 150 million and EUR 100 million, a stated 360 full-time jobs, a sovereign compute facility and a Berlin office (Government of Canada, ISED). The points off are that it launched in June 2025, that nothing is yet released, that the German funding awaits notification and that no statute or program of record installs it. Peers on hp 6: Transluce, FAR.AI and Palisade Research; Redwood Research and Alignment Research Center hold 7.

Sources: LawZero — home; LawZero — team; LawZero — research; LawZero — newsroom; Yoshua Bengio Launches LawZero; LawZero Receives Grant — Gates Foundation; LawZero advances safe-by-design AI with support from NVIDIA; LawZero receives a commitment of up to $300M from Canada and Germany; LawZero Appoints 7 Global Leaders; Jacinda Ardern joins LawZero’s Global Advisory Council; LawZero Lays out a Formal Safety Case; LawZero to Attend ICML 2026; Gates Foundation — LawZero grant; Coefficient Giving — Matt MacDermott; Coefficient Giving — grants search; Government of Canada — Canada and Germany invest in LawZero; Prime Minister of Canada — National Council on Artificial Intelligence; SFF 2024; SFF 2025; SFF 2026; METR — about; Transluce — about.