Bayesian Inference and Predictive Processing: Why AI Needs Evidence

This article examines the roles of Bayesian inference and predictive processing in AI, emphasizing that they are distinct mechanisms from LLM next-token prediction. It uses an intrusion-detection example to illustrate how evidence updates probabilities and why strong alerts don't imply strong diagnoses. The article argues for enterprise architectures that separate prediction, evidence, permission to act, and feedback loops, ensuring trustworthy AI systems can explain changes in assessment and remaining uncertainties. It clarifies common misconceptions about AI's internal models and probability claims, offering practical guidance for technical teams.