We pit non-incentivised introspection against five revealed preference mechanisms (‘truth serums’) in the elicitation of beliefs in a simple two-player game. We measure the additivity, the predictive power for own behaviour and the accuracy of each method. Beliefs from incentivised methods are better predictors of participants' own behaviour compared to introspection. However, introspection performs equally well as the truth serums in terms of accuracy and additivity. We also find that correction for risk aversion improves the additivity of scoring rule belief reports.