Language models can recognize when they are generating hallucinated references, shedding light on the nature of open-domain hallucination and suggesting potential solutions.
