Estimates of spinal forces are quite sensitive to model assumptions, especially regarding antagonistic co-contraction. Optimization based models predict co-contraction to be absent, while electromyography (EMG) based models take co-contraction into account, but usually assume equal activation of deep and superficial parts of a muscle. The aim of the present study was to compare EMG based and optimization based estimates of spinal forces in a wide range of work tasks. Data obtained from ten subjects performing a total of 28 tasks were analysed with an EMG driven model and three optimization models, which were specifically designed to test the effects of the above assumptions. Estimates of peak spinal forces obtained using the different modelling approaches were similar for total muscle force and its compression component (on average EMG based predictions were 5% higher) and were closely related (R > 0.92), while differences in predictions of the peak shear component of muscle force were more substantial (with up to 39% lower estimates in optimization based models, R > 0.79). The results show that neither neglecting antagonistic co-contraction, nor assuming equal activation of deep and superficial muscles, has a major effect on estimates of spinal forces. The disparity between shear force predictions was due to an overestimation of activity of the lateral part of the internal oblique muscle by the optimization models, which is explained by the cost function preferentially recruiting larger muscles. This suggests that a penalty for active muscle mass should be included in the cost function used for predicting trunk muscle recruitment. © 2005 Taylor & Francis Group Ltd.