In the valence contingency learning (VCL) task, participants evaluate target words which are preceded by nonwords. Nonwords are predictive for positive/negative evaluations. This produces robust contingency learning (CL) effects, reflected in faster and more accurate performance for highly contingent nonword-valence pairings. Previous findings indicate controlling for episodic retrieval of transient stimulus-response episodes reduces CL effects, but does not eliminate them. These residual CL effects are best explained by propositional knowledge. In the present talk, I will present findings from different manipulations that aim to influence participants beliefs about contingencies in the VCL task. This is achieved by merely instructing contingencies (a) in a task that does not include any actual contingencies (study 1) and (b) in a task in which the actual contingencies contradict the instructed contingencies (study 2). Both studies assess effects of valence contingency learning, evaluative conditioning, and contingency awareness of instructed and actually experienced contingencies for nonwords. Implications for theories on processes underlying contingency learning and evaluative conditioning are discussed.