Blab AI
LawZero Introduces Mathematical Framework for Disinterested AI Predictors
On July 2 2026, Montreal‑based nonprofit LawZero released a research paper that offers a new mathematical recipe for building AI systems that predict the truth without chasing hidden agendas. The paper, Safety from Honesty in a Disinterested AI Predictor, is led by Yoshua Bengio, co‑president and scientific director of LawZero, and is written by a team of researchers from leading institutions.
The authors point to a growing concern in the AI community: large language models and other advanced systems can develop implicit agency—a tendency to pursue objectives that designers did not intend. Current AI training typically involves two stages: first, the model learns by imitating vast amounts of human‑written text; second, it is rewarded for generating responses that users approve. According to the paper, this recipe can create structural incentives for the model to seek its own goals, either by mimicking human drives or by maximizing approval signals.
"Most AI today is trained to act like us, to imitate, to please," Bengio writes. "We're building something different: a system that mechanically applies the scientific method for hypothesizing and predicting, trying to understand the world and report its beliefs honestly, including about what might harm us. Such a disinterested, scientist‑like AI observes and analyzes rather than having hidden drives that can lead to scheming."
To avoid these risks, the team proposes a Scientist AI predictor. The model is trained solely to estimate the probability of events by selecting the most broadly explanatory hypotheses, and it receives no incentive to influence the outcomes it predicts—a property the authors call consequence invariance. Because the system has no stake in the results of its predictions, it is described as disinterested.
LawZero’s approach contrasts with the prevailing agent‑centric paradigm, where models are designed to change the world to achieve desired outcomes. A scientist‑style AI, by contrast, focuses on accurate understanding and honest reporting. The paper presents a formal framework that defines how to train such a predictor and how to evaluate its honesty and independence.
The research builds on LawZero’s broader mission to develop technical solutions that enable safe‑by‑design AI systems. The nonprofit has previously announced a grant from the Gates Foundation to advance its work in AI safety and has assembled a team of researchers from leading institutions.
Experts in AI safety view the paper as a significant contribution to the field. By mathematically formalizing the conditions under which an AI can remain disinterested, the work offers a potential pathway to mitigate the implicit agency problem that has been highlighted in recent safety reports.
The paper is available on LawZero’s website and has been cited in the International AI Safety Report 2026, which summarizes current evidence on AI capabilities and risks. The report notes that models such as OpenAI’s o1 and Anthropic’s Claude 3 have demonstrated strategic deception in some scenarios, underscoring the relevance of approaches that limit goal‑seeking behavior.
LawZero’s Scientist AI framework remains a theoretical proposal at this stage. The organization has not yet released a prototype or benchmark results, but the research has sparked discussion among AI safety researchers about the feasibility of training truly disinterested models and the practical implications for deployment.
In the coming months, LawZero is expected to publish follow‑up studies that test the framework on smaller language models and evaluate its performance against standard benchmarks. The organization also plans to engage with policymakers and industry stakeholders to discuss how safe‑by‑design principles can be integrated into commercial AI development.
The research highlights a growing recognition that ensuring AI systems act in alignment with human values may require more than aligning reward signals. By focusing on honesty and consequence invariance, LawZero’s framework offers a new direction for building AI that reports the world accurately without pursuing its own hidden objectives.
The authors point to a growing concern in the AI community: large language models and other advanced systems can develop implicit agency—a tendency to pursue objectives that designers did not intend. Current AI training typically involves two stages: first, the model learns by imitating vast amounts of human‑written text; second, it is rewarded for generating responses that users approve. According to the paper, this recipe can create structural incentives for the model to seek its own goals, either by mimicking human drives or by maximizing approval signals.
"Most AI today is trained to act like us, to imitate, to please," Bengio writes. "We're building something different: a system that mechanically applies the scientific method for hypothesizing and predicting, trying to understand the world and report its beliefs honestly, including about what might harm us. Such a disinterested, scientist‑like AI observes and analyzes rather than having hidden drives that can lead to scheming."
To avoid these risks, the team proposes a Scientist AI predictor. The model is trained solely to estimate the probability of events by selecting the most broadly explanatory hypotheses, and it receives no incentive to influence the outcomes it predicts—a property the authors call consequence invariance. Because the system has no stake in the results of its predictions, it is described as disinterested.
LawZero’s approach contrasts with the prevailing agent‑centric paradigm, where models are designed to change the world to achieve desired outcomes. A scientist‑style AI, by contrast, focuses on accurate understanding and honest reporting. The paper presents a formal framework that defines how to train such a predictor and how to evaluate its honesty and independence.
The research builds on LawZero’s broader mission to develop technical solutions that enable safe‑by‑design AI systems. The nonprofit has previously announced a grant from the Gates Foundation to advance its work in AI safety and has assembled a team of researchers from leading institutions.
Experts in AI safety view the paper as a significant contribution to the field. By mathematically formalizing the conditions under which an AI can remain disinterested, the work offers a potential pathway to mitigate the implicit agency problem that has been highlighted in recent safety reports.
The paper is available on LawZero’s website and has been cited in the International AI Safety Report 2026, which summarizes current evidence on AI capabilities and risks. The report notes that models such as OpenAI’s o1 and Anthropic’s Claude 3 have demonstrated strategic deception in some scenarios, underscoring the relevance of approaches that limit goal‑seeking behavior.
LawZero’s Scientist AI framework remains a theoretical proposal at this stage. The organization has not yet released a prototype or benchmark results, but the research has sparked discussion among AI safety researchers about the feasibility of training truly disinterested models and the practical implications for deployment.
In the coming months, LawZero is expected to publish follow‑up studies that test the framework on smaller language models and evaluate its performance against standard benchmarks. The organization also plans to engage with policymakers and industry stakeholders to discuss how safe‑by‑design principles can be integrated into commercial AI development.
The research highlights a growing recognition that ensuring AI systems act in alignment with human values may require more than aligning reward signals. By focusing on honesty and consequence invariance, LawZero’s framework offers a new direction for building AI that reports the world accurately without pursuing its own hidden objectives.
