Processes of attitude learning were investigated through a game requiring discrimination between good and bad objects, where feedback about object valence (involving gain or loss) is contingent on approach. Previous research demonstrates a preponderance of false-negative errors, with some good objects (‘learning asymmetry’) and most novel objects (‘generalization asymmetry’) being judged as bad, but provides no direct evidence concerning how participants appraise alternative strategies and their own performance. To compare alternative strategies, participants received advice, supposedly from a previous participant, that most objects were bad and should be avoided, or good and should be approached. Learning and generalization asymmetries were replicated, especially among participants who followed the former (risk-averse) advice. Additionally, participants' evaluations of their own game strategy were inversely related to amount of negative feedback (the number of bad objects approached), but unrelated to positive feedback (from good objects approached), pointing to the salience of negative information for self-appraisals. Copyright © 2007 John Wiley & Sons, Ltd.