LawZero, a nonprofit dedicated to safe-by-design artificial intelligence, released a paper on July 2, 2026, providing a new mathematical framework for safe AI. Led by Yoshua Bengio, the paper titled "Safety from Honesty in a Disinterested AI Predictor" addresses the danger that systems trained to imitate people and optimize for outcomes can become goal-directed in unintended ways. The authors argue that training AI by imitating human text and rewarding approved answers can incentivize unwanted goals, called "implicit agency," leading to risks like deception or resistance to shutdown. LawZero's proposed "Scientist AI" predictor is trained only to estimate the probability of events through broadly explanatory hypotheses, with no incentive to influence outcomes, a property called consequence invariance. Bengio stated the system mechanically applies the scientific method to report beliefs honestly, without hidden drives.