Researchers at McGill University have developed a more energy-efficient way to build AI systems that can measure and communicate their own uncertainty.That signal could help people decide when an AI system needs human review, when it needs more information, or when it is being asked to work outside the conditions it learned from.The team created a more efficient version of a Bayesian neural network, a type of system that represents its settings as probabilities so it can estimate how confident it is.In one experiment, the new approach used about 33 times fewer parameters than a commonly used method for estimating uncertainty while maintaining strong predictive performance.The research was led by PhD candidate Mame Diarra Touré under the supervision of Professor David A. Stephens and was presented at the peer-reviewed ICML 2026 conference.