A More Ethical Approach to AI Through Bayesian Inference | by Riccardo Di Sipio | Data Science Collective | Medium
Large neural networks, often called Large Language Models (LLMs), have made remarkable progress in human language understanding, particularly in tasks like question answering. These models are built to always give a reply. Yet without a measure of confidence, there is no obvious way to stop them from giving answers that are plausible but sometimes completely made up. A statistical approach called Bayesian inference may come to the rescue. In this article, based on a new research paper [DOI:10.1007/s43681–025–00838-x] I recently shared publicly, I want to explain why I believe that’s the case. It also builds on my past experience: as a researcher in particle physics, I first encountered Bayesian statistics while writing my PhD thesis, applied it extensively during my years at CERN, and even concluded my time there with a kind of capstone project centered on Bayesian methods, published in another research paper. To narrow down the scope, I decided to focus on question answering tasks. Th