To better understand how neural networks learn to simulate writing, researchers trained simpler versions on synthetic children’s stories.