ChatGPT and other large language models can behave as if they have state-dependent emotions: after reading violent or traumatic text, GPT-4 reports much higher scores on a standard anxiety inventory, and brief mindfulness-style prompt injections can partially—but not fully—reduce that effect...