A Bayesian network is a way to represent uncertain knowledge as a graph: each node is a variable (say, “it is raining,” “the sprinkler is on,” “the grass is wet”), and arrows between nodes capture how the variables depend on one another probabilistically. Given some observations, the network can compute updated probabilities for the things you cannot see directly. It is a disciplined answer to a hard problem: how a machine should reason when it does not know things for certain but does know how likely they are.
The approach was developed by computer scientist Judea Pearl, whose 1986 paper “Fusion, propagation, and structuring in belief networks” (Artificial Intelligence journal, DOI 10.1016/0004-3702(86)90072-X) laid out the propagation algorithms that make inference in these networks practical. Pearl’s broader program, captured in his 1988 book “Probabilistic Reasoning in Intelligent Systems,” gave AI a rigorous mathematical language for handling uncertainty, and it later grew into his formal theory of causality. He received the 2011 ACM Turing Award for this body of work.
Bayesian networks were a dominant paradigm in AI through the 1990s and 2000s, especially for diagnosis, risk assessment, and decision support, areas where you need not just a prediction but a transparent chain of probabilistic reasoning. They sit on the more interpretable, structured side of AI, in contrast to today’s neural networks: a Bayesian network’s logic can be read off its graph, whereas a deep model’s reasoning is opaque. The trade-off is that they require someone to specify the structure and probabilities, which is hard to do for messy, high-dimensional data like raw images or text.
Why business readers should care: Bayesian networks remain valuable wherever you need explainable reasoning under uncertainty and an audit trail for how a conclusion was reached, common in medicine, insurance, fraud, and regulated decisions. They are also a reminder that AI is not only deep learning. Pearl’s insistence that genuine understanding requires causal models, not just pattern matching, is an active and influential critique of the current AI mainstream.