Recherche

Damiano Fornasiere*, Mirko Bronzi*, Spencer Kitts*, Alessandro Palmas, Yoshua Bengio†, Oliver Richardson†

We provide evidence that language models can detect, localize and, to a certain degree, verbalize the difference between perturbations applied to their activations. More precisely, we either (a) mask activations, simulating dropout, or (b) add Gaussian noise to them, at a target sentence. We then ask a multiple-choice question such as “Which of the previous sentences was perturbed?” or “Which of the two perturbations was applied?”. We test models from the Llama, Olmo, and Qwen families, with sizes between 8B and 32B, all of which can easily detect and localize the perturbations, often with perfect accuracy. These models can also learn, when taught in context, to distinguish between dropout and Gaussian noise. Notably, Qwen3-32B’s zero-shot accuracy in identifying which perturbation was applied improves as a function of the perturbation strength and, moreover, decreases if the in-context labels are flipped, suggesting a prior for the correct ones—even modulo controls.
Because dropout has been used as a training-regularization technique, while Gaussian noise is sometimes added during inference, we discuss the possibility of a data-agnostic “training awareness” signal and the implications for AI safety.

Damiano Fornasiere*, Oliver Richardson*, Gaël Gendron, Iulian Serban, Yoshua Bengio. 

Scientific theories aspire to describe what is, as opposed to prescribe what ought to be. At LawZero, we take this idea as a design principle for safe artificial intelligence: that understanding—even of arbitrary depth and scope—can be disentangled from preference over how the world unfolds.

We distill into a non-technical overview the motivations and core components of the Scientist AI, a system that aspires to this ideal. Agency, we argue, rests on three pillars—affordances, goal-directedness, and intelligence—each a matter of degree. By limiting the first two while pursuing the third, we aim to build a system that is highly intelligent yet incapable of holding or pursuing goals of its own. The Scientist AI comprises a generator held accountable by a neutral estimator, allowing for creative thought without compromising safety. Two key ingredients are (i) contextualization, a transformation of the training data that disentangles facts from statements about such facts (e.g., opinions), and (ii) consequence invariance, a property of the training process that prevents feedback about downstream outcomes.

We believe this approach offers a promising path toward systems that are at once powerful, transparent, and safe, and that may serve as trustworthy anchors in a world of increasingly autonomous AI.

Yoshua Bengio, Michael Cohen, Damiano Fornasiere, Joumana Ghosn, Pietro Greiner, Matt MacDermott, Sören Mindermann, Adam Oberman, Jesse Richardson, Oliver Richardson, Marc-Antoine Rondeau, Pierre-Luc St-Charles, David Williams-King

The leading AI companies are increasingly focused on building generalist AI agents — systems that can autonomously plan, act, and pursue goals across almost all tasks that humans can perform. Despite how useful these systems might be, unchecked AI agency poses significant risks to public safety and security, ranging from misuse by malicious actors to a potentially irreversible loss of human control. We discuss how these risks arise from current AI training methods.