What You Should Not Ask a Language Model About Your Body

A language model has read more papers than I have. It has not taken your history, looked at your joint, or ordered a blood test. It will still write as if it has. Wearable biosensing plus machine learning is being sold as coaching support (Biosensors, 2026). Support is the right word. Diagnosis is a different job.

WHOOP is not a medical device. Overnight heart rate can agree with a lab in a protocol like Miller's (Miller, 2022). That does not turn a recovery score into a disease label. The 2024 WHOOP review is a methods paper, not a license to treat a yellow morning as pathology (Panchawagh, 2024).

Do not ask it if you have overtraining syndrome. Meeusen's joint consensus is a clinical picture: performance, mood, sleep, and rest that doesn't restore you, assessed by people who can examine you (Meeusen, 2013). A CSV cannot finish that sentence.

Do not ask it to name a supplement stack, change a prescription, or interpret a full lab panel as a treatment plan. Do not ask it to clear you for running on a tendon that still hurts. Do not upload other people's health records. Do not send photos of a child's rash, a stranger's chart, or a report that still has your address on the letterhead.

Load questions can still go wrong. Acute:chronic workload ratios show up in injury conversations (Blanch and Gabbett, 2016; Wang, 2025). They are a research tool with limits, not a red light on your phone. If the model says you will get injured because a ratio crossed a number, that is not a diagnosis. It is a spreadsheet with confidence.

Where did load accumulate this week. Which days should stay easy. Did overnight HRV and resting heart rate drift together. Did heavy lower body cost more than strain admitted. What should next week protect. Those are calendar questions. The model is a fast clerk for calendar questions.

If the answer starts to sound like a clinic, stop. Push it back to the week. If you cannot complete easy days without compensating, if pain is new or changing side, if mood and sleep stay wrong after you've already cut quality, close the chat and see a person.

Write the refusal into the Project or the custom GPT so you don't have to remember it when you're tired. No diagnoses. No drugs. No invented lab accuracy. No claim that a wearable measured muscle damage or VO2 unless that is actually in the file. Propose training changes. Leave the body to the people who can see it.

This isn't medical advice. It is a boundary. The model will cross it if you ask nicely. Don't ask.