Biomechanical models of the spine have traditionally assumed that workplace lifting conditions (weight, posture, motion, etc.) precisely dictate the magnitude of individual muscle forces necessary to maintain a biomechanical balance within the trunk. However, because there are a large number of muscle groups within the trunk there is also an infinite number of possible combinations of muscle forces that can satisfy this biomechanical balance requirement for a given condition. Currently there are no methods available to predict this possible variability in muscle activity. Such variability in a multiple muscle system can result in variations in spinal loading. To quantitatively capture this trunk muscle variability during bending motions, such as those involved in lifting, a stochastic (probabilistic) model of trunk muscle activation was developed. The model was based on a simulation of experimentally derived data and predicted the possible combinations of time-dependent trunk muscle coactivations that could be expected given a set of trunk bending conditions. These simulated muscle activities were then used as input to an electromyographically assisted biomechanical model so that the magnitude and variability of the spine reaction forces could be estimated. This procedure allows one to assess the range of spinal loads that would be expected with a particular task. Significant variability in muscle activities was observed for each specific lifting condition and explained biomechanically. The results indicated that the variability in trunk muscle force had a small effect on spinal compression variability (+/- 7% of the mean compression), but greatly influenced both lateral (+/- 90% of mean) and anteroposterior shear forces (+/- 40% of mean). A validation study confirmed that the model predictions were reasonable estimates of muscle activity variability under previously untested conditions. This work could help explain how some repetitive lifting motions could increase the risk of acquiring a low back disorder and the simulation model could help drive electromyographically assisted models without the need for recording actual electromyographic activity.