During treatment for fear and anxiety disorders, patients are often encouraged to face their fears and experience that the feared outcome does, in contrast to their expectations, not occur. Yet treatments do not always result in the desired behavioural changes, and even if effective, some individuals experience a return of symptoms afterwards. There is thus great value in better understanding the conditions that can facilitate learning from unexpected outcomes, and how we can manipulate those conditions to our benefit with the goal of improving treatment effects. The current dissertation comprises experimental and theoretical work detailing how prediction errors (i.e., the difference between an expected outcome and an actually occurring outcome) affect learning processes such as extinction learning and memory reconsolidation. We further aimed to develop a physiological marker of prediction error occurrence in a fear-conditioning paradigm. Our findings present novel insights into the measurement and manipulation of prediction errors in humans. In sum, while the importance of prediction errors for learning is unquestioned, we find little support for the notion that maximizing prediction errors during learning may be the key to more effective treatments. However, our findings also highlight some important conceptual and translational challenges in investigating the effect of prediction errors on learning outcomes. Further possibilities to investigate the optimal conditions for learning and long-term behavioural changes are discussed.