Affective polarization – dislike between political outgroups – has multiple negative consequences. Political discussions between members of opposing groups have the potential to reduce affective polarization, but studies to date take place in heavily moderated settings, avoid politics and quell potential division. This leaves open the question of whether their findings apply in a real-world, affectively polarized environment in which people are divided on contentious issues. To answer this question, this study first uses data from political chat discussions organised in the run-up to the Dutch 2021 general election. Random matching of subjects to an unresponsive partner is exploited for causal inference. Using expectations derived from contact theory and the deliberation literature, results show that a political discussion of only 20 minutes increases a subject’s expressed sympathy for people from the other side of the political spectrum. This result also apply to the most polarized supporters of far right parties. The results are then replicated and extended with an online experiment on the same chat platform, which tests the effect of a conversation on polarisation, social distance, and support for political compromise. The second study also varies the topic of conversation, the ideological distance between discussing partisans, and conducts quantitative and qualitative text analyses of the chat discussions, to shed light on possible mechanisms.