To learn or not learn from AI? Unpacking the effects of feedback valence on novel insights recall

Artificial intelligence (AI) tools are unique pedagogical agents, in that (1) they can generate novel insights, and (2) they tend to be treated differently by humans, because they are opaquer and more intelligent than traditional technologies. Thus, it is not clear if and how people learn novel insights through AI feedback, and if the learning process and outcomes differ compared to when the source of insights and feedback is humans. Here, we seek to address this gap with four preregistered experiments (total n=1,766). They use a learning paradigm in which we manipulate the feedback source, feedback valence, and test instructions. The key finding is that people learn novel insights better when they receive negative feedback from AI than from humans. There is no difference in recall for positive feedback. This implies that AI as pedagogical agents are likely to outperform human feedback sources (peers or experts), or at least, be no worse. Our results further reveal that (1) people present typical human-human learning biases when receiving negative feedback from AI, namely mnemic neglect, and (2) ego bruising (as manifested in reduction in state self-esteem) and anger mediate such effects when receiving negative feedback from AI.