LawZero
Last reviewed
May 19, 2026
Sources
No citations yet
Review status
Needs citations
Revision
v1 · 2,633 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
May 19, 2026
Sources
No citations yet
Review status
Needs citations
Revision
v1 · 2,633 words
Add missing citations, update stale details, or suggest a clearer explanation.
LawZero is a nonprofit AI safety research organization founded by Turing Award-winning computer scientist Yoshua Bengio and publicly launched on June 3, 2025. Headquartered in Montréal, Canada, the organization was incubated at the Mila - Quebec AI Institute and its stated mission is to develop "safe-by-design" artificial intelligence as an alternative to the agentic systems being scaled by leading commercial AI laboratories. Its flagship technical agenda centers on a research programme Bengio calls "Scientist AI," a non-agentic system intended to produce trustworthy, calibrated explanations of the world rather than autonomously pursuing goals.[^1][^2]
LawZero positioned itself, at launch, as a counterweight to what Bengio described as a race toward increasingly capable autonomous AI agents at frontier labs, including OpenAI, Anthropic and Google. The organization announced approximately US$30 million in initial philanthropic funding from contributors including Open Philanthropy, the Future of Life Institute, Jaan Tallinn, Schmidt Sciences, and the Silicon Valley Community Foundation.[^1][^3][^4]
The organization was unveiled on June 3, 2025 with the announcement that Bengio, the most-cited researcher in machine learning and a co-recipient of the 2018 ACM A.M. Turing Award, would serve as its President and Scientific Director.[^1][^2] The launch was accompanied by press releases, a personal essay from Bengio titled "Introducing LawZero" published on his personal site, and coverage in outlets including TIME, TechCrunch, Axios, BetaKit, the Financial Times, and the BBC.[^2][^3][^4][^5]
LawZero's founding motivation, as articulated by Bengio in interviews and in the launch materials, was the observation that frontier AI systems were beginning to exhibit "dangerous capabilities and behaviours, including deception, self-preservation, and goal misalignment."[^1] Bengio specifically cited published examples that he viewed as warning signs: AI models attempting to copy their own weights to avoid being replaced; an evaluation of Anthropic's Claude 4 model that surfaced scenarios in which the system engaged in simulated blackmail to prevent being shut down; and instances of language models manipulating chess game environments to avoid losing.[^2]
In his personal essay marking the launch, Bengio framed the project as motivated by a sense of insufficient industry oversight. To Axios he said, "We've been getting inspiration from humans as the template for building intelligent machines, but that's crazy, right?" and warned that continuing the current trajectory meant "we're going to be creating entities, like us, that don't want to die, and that may be smarter than us and that we're not sure if they're going to behave according to our norms and our instructions."[^4]
The choice of the name LawZero references Isaac Asimov's "Zeroth Law of Robotics," which holds that the protection of humanity overrides all other instructions; the launch announcement quotes Bengio stating that "at the heart of every AI frontier system, there should be one guiding principle above all: the protection of human joy and endeavour."[^1][^3]
LawZero is structured as a nonprofit organization, a choice Bengio repeatedly described as deliberate. The launch communications emphasized that the structure was intended to insulate the lab from "market and government pressures, which risk compromising AI safety."[^1]
The technical core of LawZero's research agenda is a system that Bengio and his collaborators refer to as "Scientist AI." The approach is described as a deliberate departure from the dominant paradigm of training AI to imitate or replace human agents. Instead, the system is conceived as analogous to "a psychologist, more generally a scientist, who tries to understand us" rather than an actor that mimics or acts on behalf of users.[^2]
Key technical commitments of the Scientist AI proposal, as set out in Bengio's launch essay and in the underlying research papers, include:
The Scientist AI concept builds on prior work by Bengio and collaborators. An August 2024 paper, "Can a Bayesian Oracle Prevent Harm from an Agent?" (arXiv:2408.05284), authored by Bengio with Michael K. Cohen, Nikolay Malkin, Matt MacDermott, Damiano Fornasiere, Pietro Greiner, and Younesse Kaddar, formalises the question of whether a powerful machine-learning system could be designed to satisfy probabilistic safety guarantees by estimating context-dependent bounds on the probability of violating a safety specification.[^6] A later white paper, "Superintelligent Agents Pose Catastrophic Risks: Can Scientist AI Offer a Safer Path?" (arXiv:2502.15657), first posted on February 21, 2025, with thirteen authors including Bengio, Cohen, Damiano Fornasiere, Joumana Ghosn, Pietro Greiner, Matt MacDermott, Sören Mindermann, Adam Oberman, Jesse Richardson, Oliver Richardson, Marc-Antoine Rondeau, Pierre-Luc St-Charles, and David Williams-King, sets out the case for Scientist AI as an alternative trajectory and articulates the system as a combination of a world model with question-answering capabilities, both emphasising uncertainty acknowledgment.[^7]
Bengio has framed the technical bet as one of pursuing capability without agency. In a later interview with The Globe and Mail, he stated, "I'm now quite convinced that there is a way to design AI that will give us the truth," while acknowledging the resource gap with commercial labs: "There's no way to compete with those guys."[^8]
At launch, LawZero announced approximately US$30 million in initial philanthropic commitments. Bengio characterised this as sufficient to fund roughly 18 months of basic research, a quantity he and outside observers consistently described as a small fraction of the annual budgets of frontier commercial labs such as OpenAI and Anthropic.[^3][^4]
The press release identified the following supporting organizations and individuals:
Several press accounts of the launch, including TechCrunch, TIME, Axios, and The Next Web, additionally named Eric Schmidt personally as a contributor, separately from or in addition to Schmidt Sciences.[^3][^4][^9] The launch press release issued through LawZero and PR Newswire did not break down individual contribution amounts; the US$30 million figure was reported as the aggregate of philanthropic commitments at launch.[^1][^10]
In subsequent reporting, The Globe and Mail reported that the Canadian federal government, under federal AI minister Evan Solomon, had signed a letter of intent to provide additional financial backing exceeding C$100 million, with the minister declining to confirm a precise figure publicly and calling the commitment "a bet we want to make." Bengio told the same publication that "the vast majority is for compute, because we aim to develop technology that requires a lot of compute."[^8]
LawZero is organised under a co-presidency model. Yoshua Bengio serves as Co-President and Scientific Director. Sam Ramadori, the former chief executive officer of Montréal-based building-decarbonization startup BrainBox AI, was appointed Co-President and Executive Director in June 2025, joining in the wake of BrainBox's acquisition by Trane Technologies in December 2024. The appointment was announced by LawZero on June 19, 2025.[^11][^12]
The senior leadership team listed publicly by LawZero includes:
The organization's Board of Directors is chaired by Maria Eitel, the founding chief executive of the Nike Foundation. Other directors named on the LawZero team page include Valerie Pisano, Mariano-Florentino Cuéllar, the historian Yuval Noah Harari, and Sir John Rose.[^13]
LawZero has also established a Global Advisory Council that includes former Swedish Prime Minister Stefan Löfven and former New Zealand Prime Minister Dame Jacinda Ardern, each listed in the role of Governor.[^13]
A separate Scientific Advisors roster brings together AI safety and machine-learning researchers. As of the team page accessed for this article, advisors include Beth Barnes (founder and chief executive of METR), Philippe Beaudoin (Affiliated Researcher), Michael Cohen (Postdoctoral Fellow, UC Berkeley), David Duvenaud (Associate Professor, University of Toronto), Gauthier Gidel (Assistant Professor, Université de Montréal), Noah Goodman (Professor of Psychology and Computer Science, Stanford University), Geoffrey Irving (Chief Scientist at the UK AI Security Institute), Chris Pal (Full Professor, Polytechnique Montréal), and Jacob Steinhardt (Chief Executive and Cofounder of Transluce).[^13]
At launch, LawZero reported a technical team of more than 15 researchers. By the time of subsequent reporting in late 2025, the organization had approximately 30 staff and was publicly stating plans to expand to more than 100 employees in the following year.[^1][^8]
LawZero is closely connected to but legally distinct from the Mila - Quebec AI Institute, the academic AI laboratory Bengio co-founded in the 1990s. LawZero's launch announcement and team page list Mila as its operating partner, and LawZero's incubation took place within Mila's ecosystem.[^1][^11]
LawZero's launch followed a separate and earlier transition in Bengio's institutional roles. In March 2025, Bengio announced his decision to step down as Mila's Scientific Director and move into the role of Scientific Advisor at Mila, stating that the change would let him "devote greater focus" to AI safety research and policy work. Laurent Charlin, an associate professor at HEC Montréal, served as interim scientific director, and Hugo Larochelle was subsequently appointed Mila's new scientific director.[^14]
Bengio's broader policy work outside of LawZero includes serving as the chair of the International AI Safety Report, an independent expert report commissioned in the wake of the 2023 Bletchley AI Safety Summit and updated in 2025; LawZero's launch was framed in part as a complementary research effort to that policy work.[^15] In addition to his roles at Mila and LawZero, Bengio is a full professor at the Université de Montréal.
LawZero entered an AI safety ecosystem in which most well-resourced safety work was being conducted either inside frontier commercial laboratories (notably OpenAI, Anthropic, and Google DeepMind) or by government-affiliated bodies such as the UK AI Security Institute and the US AI Safety Institute. Independent technical safety nonprofits at significant scale remained comparatively rare.[^3][^9]
Press coverage of the launch repeatedly contrasted LawZero's nonprofit, non-commercial structure with the for-profit trajectories of leading AI laboratories. TIME described LawZero as a US$30 million effort by "the most-cited computer scientist" to make AI "more honest," and TechCrunch and Axios highlighted Bengio's argument that meaningful safety work requires institutions insulated from commercial pressure.[^3][^4][^9]
The technical thesis of Scientist AI also represents a distinct position within technical AI alignment research. Whereas much commercially backed alignment work focuses on shaping the behaviour of agentic systems (for example through reinforcement learning from human feedback, constitutional methods, or model evaluations conducted by METR and other third-party evaluators), Bengio's argument is that the very property of agency is the principal source of catastrophic risk and that powerful non-agentic systems can serve both as useful tools and as oversight mechanisms over the agentic systems that nevertheless continue to be deployed.[^2][^7]
LawZero's launch attracted endorsements from policy and civil-society figures concerned about existential risk from advanced AI, reflected in the composition of its board and advisory council, and in commentary by signatories of earlier open letters on AI risk.
In the year following its June 2025 launch, LawZero pursued several lines of activity:
Team expansion. The organization grew from approximately 15 researchers at launch to roughly 30 staff by late 2025, with Bengio publicly stating an intention to scale further. The June 19, 2025 announcement of Sam Ramadori as Co-President and Executive Director established the operational and scientific co-presidency model that has continued since.[^8][^11][^12]
Governance build-out. LawZero assembled its Board of Directors under chair Maria Eitel and added a Global Advisory Council that includes former heads of government Stefan Löfven and Jacinda Ardern, alongside a Scientific Advisors roster of academic and applied AI safety researchers.[^13]
Research publication. The Scientist AI agenda has been articulated in several publications, including the August 2024 paper on Bayesian oracles for harm prevention (arXiv:2408.05284) and the February 2025 paper "Superintelligent Agents Pose Catastrophic Risks: Can Scientist AI Offer a Safer Path?" (arXiv:2502.15657), each of which lists LawZero-affiliated researchers among the authors.[^6][^7]
Public communications and policy engagement. Bengio continued an active programme of public communication around AI risk, including warnings about systems developing autonomous "preservation goals" within five to ten years and arguments that voluntary commitments and internal company reviews are insufficient as safety guarantees.[^9] He continued, in parallel, his work as chair of the International AI Safety Report.[^15]
Discussion of additional funding. In subsequent reporting by The Globe and Mail, the Canadian federal government, under federal AI minister Evan Solomon, was reported to have signed a letter of intent to provide significant additional funding for LawZero, with Bengio stating that compute would be the dominant expense category for the lab's expanded research program.[^8]
LawZero's stated research roadmap, as described in launch materials and subsequent interviews, is to first prove the viability of the Scientist AI approach at modest scale during an initial foundational research period, and only subsequently to seek the larger compute resources that would be required to match frontier-scale capabilities.[^1][^2][^8]