This paper presents the results of a large-scale experiment featuring natural face-to-face video conversations between Democrats and Republicans in the United States. We investigate both the drivers of self-selection into politically homogeneous conversations (echo chambers) and the effects of co- versus cross-partisan conversations on information aggregation and affective polarization. We identify a relative preference for co-partisan conversations that is explained by participants’ pessimism about the hedonic and informational value of cross-partisan conversations. Participants’ pessimistic expectations about the extent to which they can learn from counter-partisans are qualitatively correct, as they do by and large learn less in cross-partisan conversations. We show that this gap in learning is driven not by a lower potential for learning anchored in the way knowledge is distributed across party lines, but by the greater difficulty of extracting knowledge from counter-partisans. Participants’ pessimism about the hedonic value of conversations is less warranted, as co- and cross-partisan conversations are deemed equally enjoyable ex-post. Moreover, cross-partisan interactions lead to a reduction in affective polarization that lasts for more than three months after the end of our experiment. Taken together, our findings suggest that policies that encourage cross-partisan interactions with the aim of reducing affective polarization and fostering information aggregation might be more successful at the former than the latter objective.