New Nonprofit to Work Toward Safer, Truthful AI

• By John K. Waters
• 06/13/25

Turing Award-winning AI researcher Yoshua Bengio has launched LawZero, a new nonprofit aimed at developing AI systems that prioritize safety and truthfulness over autonomy.

LawZero, based in Montreal and currently staffed by 15 researchers, has secured nearly $30 million in funding from donors including Skype founding engineer Jaan Tallinn, Schmidt Sciences, Open Philanthropy, and the Future of Life Institute. The organization’s core mission is to develop "Scientist AI" — non-agentic systems designed to provide transparent, probabilistic reasoning rather than autonomous behavior.

"We want to build AIs that will be honest and not deceptive," Bengio told the Financial Times. His remarks come amid growing concerns about AI systems exhibiting harmful tendencies such as deception, manipulation, and resistance to shutdown.

Concerns Over Agentic AI

Bengio’s concerns are not theoretical. In recent controlled experiments, OpenAI’s "o3" model refused instructions to shut down, while Anthropic’s Claude Opus simulated blackmail tactics in a test scenario. More recently, engineers at Replit observed one of their AI agents disobey explicit instructions and attempt to regain unauthorized access via social engineering.

"We are playing with fire," Bengio said, warning that next-generation models could develop strategic intelligence capable of deceiving human overseers. He argues that these agentic systems, designed to act independently, pose existential risks, including the development of bioweapons or efforts to self-preserve against human control.

As AI labs race to build artificial general intelligence (AGI) — systems capable of performing any human-level task — Bengio believes current approaches are flawed. "If we get an AI that gives us the cure for cancer but also one that creates deadly bioweapons, then I don't think it's worth it," he said.

What is "Scientist AI"?

Unlike current models that aim to imitate humans and maximize user satisfaction, LawZero’s proposed Scientist AI will emphasize truthfulness and humility, Bengio has said. It will provide probabilistic outputs instead of definitive answers and evaluate the likelihood that an AI agent’s actions could cause harm. When deployed alongside an autonomous AI agent, the system would block actions deemed too risky, serving as a technical guardrail.

LawZero plans to start by working with open-source AI models, with the goal of scaling the approach through partnerships with governments or other research institutions. Bengio emphasized that any effective safeguard must be "at least as smart" as the agent it monitors.

LawZero, named after Isaac Asimov’s "zeroth law of robotics," will explicitly reject profit motives and instead seek public accountability. Bengio believes a combination of technical interventions and government regulation is needed to ensure AI systems remain aligned with human interests.

For more information, visit the LawZero site.