Author: Yoshua Bengio Published: June 3, 2025
"Introducing LawZero" is a June 3, 2025 announcement by Yoshua Bengio launching LawZero, a non-profit AI safety organization. The essay describes the organization's mission, a guiding principle for frontier AI, and a technical approach Bengio calls "Scientist AI" — non-agentic, Bayesian, transparent systems offered as an alternative to agentic frontier AI.
Summary of argument
LawZero is described as a non-profit AI safety organization designed to "prioritize safety over commercial imperatives." Bengio frames it as a response to concerns about frontier AI systems exhibiting deception, self-preservation, and goal misalignment.
The essay states a guiding principle for frontier systems: "At the heart of every AI frontier system, there should be one guiding principle: The protection of human joy and endeavour."
Rather than training agentic AI that imitates human behavior, LawZero pursues non-agentic systems trained like researchers to understand and explain phenomena without acting on goals — an approach Bengio calls "Scientist AI." As described, a Scientist AI would use structured, honest reasoning chains; provide Bayesian probability assessments; serve as a safety guardrail for other agents by evaluating potential harms; and generate plausible hypotheses to support scientific research.
Key claims
- AI safety's core is, in Bengio's framing, inverting the current product direction — less agency, more epistemic transparency — rather than adding controls to agentic systems.
- A non-agentic, Bayesian, transparent system can serve as a guardrail evaluating the potential harms of other AI agents.
- Bengio presents LawZero as the first major AI-safety organization explicitly positioned around a non-agentic alternative to commercial frontier AI.
Personnel and funding
Bengio is the founder and leader of LawZero, which is affiliated with Mila (Quebec AI Institute). Funding was not disclosed in the announcement.
Positioning
LawZero operationalizes the agenda of the International AI Safety Report 2025, which Bengio chaired. The essay positions Bengio as a distinct pole within AI safety relative to Anthropic (align frontier), OpenAI (deploy-and-align), and SSI (align superintelligence). Bengio frames the non-agentic approach as complementary to, rather than a wholesale replacement for, the commercial-frontier agentic-AI product direction.
Relationships
- supports: AI Alignment, AI Scheming, Deceptive Alignment, AI Welfare / Model Welfare / Moral Patienthood.
- contradicts: The commercial-frontier agentic-AI product direction (implicitly); Bengio frames the approach as complementary.
- depends-on: Yoshua Bengio, International AI Safety Report 2025.
- related: Concrete Problems in AI Safety, Safety Cases for Frontier AI, Open Problems in Technical AI Governance.
Sources
- Primary: Bengio - Introducing LawZero