Adhesively bonded joint has gained increasing popularity due to weight reduction and relief of stress concentration. However, damage in the mode of adhesive failure and cohesive failure, as well as the adherend failure, could still occur, and it has been realized that the reliability of the adhesively bonded joint depends on numerous complex and even nonlinearly interacting factors. Consequently, the prediction of load-bearing capacity and damage localization for an adhesively bonded joint can be difficult due to the not well-known effects of the variations or uncertainties in the imperfected adhesive-adherent bonding surfaces or the environmental conditions as well as the material and geometrical nonlinearities. Although abundant uncertainties are present, the standard analysis tool in industries is still deterministic finite element analysis (FEA). The routine practice of such analysis is applying a cohesive zone model (CZM) implemented with a proper traction-separation law to predict the onset and gradual degradation of adhesion which may underpredict or overpredict the load-bearing capacity of an adhesively bonded joint. In this study, deterministic FEA with a CZM is first applied to predict the load-displacement curve of an adhesively bonded polyurethane-to-steel double butt joint. A comparison of the prediction with the experiment reveals the inability of a deterministic approach for accurately predicting the load-bearing capability of the joint and the failure propagation route. Then, uncertainty analysis using polynomial chaos expansion (PCE) is applied to examine if it can enhance the prediction of the joint failure. The results show the predictions generated by the uncertainty analysis correlate better than the deterministic analysis with the test data, hence demonstrating the potential of uncertainty analysis in improving the prediction of the failure mode and load-bearing capability of an adhesively bonded polyurethane-to-steel double butt joint.