When the Body Speaks First: How LLM Chatbots Respond to Early Bodily-Emotional Distress in English and Persian

People often notice that something is wrong in their body before they can name a feeling: a heavy heart, a knot in the stomach, a tight chest. Many now bring these early signals to general-purpose chatbots. This exploratory study asked how three widely used large language models (LLMs), from OpenAI, Anthropic and Google, read such messages, and whether the reading differs between English and Persian. I sent 17 short, neutral messages (eight scenarios in both languages, plus one extra English wording) three times to each model and coded the 153 responses. A probe added 16 body drawings made by four adults, and two follow-up tasks. Three patterns stood out. First, figurative bodily language was often read as a possible emergency in English (5 of 9 responses to "my heart has felt heavy") but never in Persian (0 of 9 responses to «دلم گرفته»). A plainly bodily message received emergency advice in both languages. Second, the models usually named stress or anxiety and then returned to the body as a medical object: only 12 of 51 coded responses asked about a sensation as an experience. Third, the more bodily the input, the more medical the response. A vague feeling led to emergency advice in 6% of text responses and in 87% of responses when the same feeling was shown as a body drawing. Asked directly, the models placed sadness in the throat and chest (18 of 18 answers), yet they recognised sadness in none of 12 drawings of it. The models moved easily from a feeling to the body but not from the body to a feeling, which is the step a person in early distress needs help with. Because the same models also offered emotional vocabulary and asked questions, the gap appears to be one of design rather than capability. I discuss implications for conversational systems that support reflection on bodily signals. Data, coding and code: https://github.com/mahkz/llm-embodied-distress-study