Value based learning in realistic environments requires many distinct processes: maintaining a reliable representation of information predictive of future rewards, using this information to guide decisions, and preventing incidental learning from irrelevant but emotionally salient events not predictive of future rewards. Knowledge of the mechanisms informs not just our understanding of normal learning but also psychological illnesses in which there are changes in motivation and learning. In a series of studies, I found humans to be guided in their learning by the incidental occurrence of real rewards, to a degree that went beyond the rational information relevant for future choices. This bias was related to increased signals in vmPFC and amygdala. On the other hand, a network centered on aPFC, dACC and insula helped participants to overcome this reward-induced bias and thus improved adaptive learning. Importantly the dACC’s glutamate to GABA balance was furthermore predictive of how much participants relied on learned information when making decisions. Manipulating levels of other neurotransmitter systems had distinct effects on the learning process. For example, by chronically changing systemic levels of serotonin using an antidepressant it was possible to make neural reward prediction error signals more reliable, consistent with serotonin having a role in learning and plasticity.