The Role of Confirmation Bias in Reinforcement Learning: Effects of Confidence and Truthfulness

Introduction: Over the last few years, the rise of social media has been accompanied by an increase in the spread of fake news. It has been suggested that the human predisposition to seek information aligning with pre-existing beliefs, i.e., confirmation bias (CBias), is one of the main mechanisms underlying misinformation[1]. CBias contributes to maintaining maladaptive beliefs, even in the face of contrary evidence. This behavior appears in clear opposition to traditional reinforcement learning theories, which indicate that external feedback is a key element in guiding our decisions[2]. Here, we sought to understand the role of CBias in reinforcement learning within contexts related to news consumption, by disambiguating the weight of truthfulness and confidence in the generation of confirmation mechanisms. Methods: 34 healthy volunteers (23.9(3.7); M: 13, F: 21) divided into two experimental groups underwent a 4-phase study. In a first task we measured the participants' ability to learn differential probabilistic associations between rewards and abstract symbols. Second, individual beliefs related to a series of 324 headlines were mapped; d' and c indices were measured to verify possible group differences in the ability to discriminate fake news or in the response criterion. Thirdly, subjects performed the previous learning task using headlines judged as being true (group 1) or false (group 2) rather than symbols. Finally, in the fourth phase, participants have the chance to confirm/disconfirm previous opinions, allowing us to evaluating the effects of rewards association to belief updates. Results: We found that CBias acted as a learning heuristic influencing reinforcement learning. Compared to the learning task performed with symbols, which did not convey any subjective value of credibility, participants showed higher accuracy and were more likely to choose the reinforced element when rewards were linked to the supposedly true news. This effect was especially pronounced when these elements were associated with higher individual confidence. In contrast, we did not observe this outcome when reinforcement was linked to the supposedly fake news. However, different reward association with news truthfulness did not produce differential changes in belief updates, resulting in shifts of criterion toward liberal attitude to both groups. Discussion: Our study confirmed the influence of CBias on reinforcement learning[3]. Notably, simply possessing a previous belief does not appear to increase the salience of a stimulus. However, firmly believing something to be true does seem to affect individual choices in a learning task. However, such an effect strongly depend on the individual confidence. Taken together, our results underscore the interaction between confidence and truthfulness in the facilitation of learning effects attributed to confirmation bias: believing something to be true enhances learning about potential environmental reinforcement, and this effect is more pronounced as an individual's confidence in their beliefs increases.