Context: Forensic geotechnical engineering aims to determine the most likely causes leading to geotechnical failures. Standard practice tests a set of credible hypotheses against the collected evidence using backward analysis and complex but deterministic geotechnical models. Geotechnical models involving uncertainty are not usually employed to analyze the causes of failure, even though soil parameters are uncertain, and evidence is often incomplete. Method: This paper introduces a probabilistic model approach based on Bayesian Networks to test hypotheses in light of collected evidence. Bayesian networks simulate patterns of human reasoning under uncertainty through a bidirectional inference process known as “explaining away.” In this study, Bayesian Networks are used to test several credible hypotheses about the causes of levee failures. Probability queries and the K-Most Probable Explanation algorithm (K-MPE) are used to assess the hypotheses. Results: This approach was applied to the analysis of a well-known levee failure in Breitenhagen, Germany, where previous forensic studies found a multiplicity of competing explanations for the causes of failure. The approach allows concluding that the failure was most likely caused by a combination of high phreatic levels, a conductive layer, and weak soils, thus allowing to discard a significant number of competing explanations. Conclusions: The proposed approach is expected to improve the accuracy and transparency of conclusions about the causes of failure in levee structures.