AI does not diagnose

The second rule of this track completes the first. In a clinic, general AI services are never used for four things: diagnosing (guessing an illness from symptoms or findings), triaging (deciding who goes first or how serious a case is), recommending treatment (choosing a medicine, procedure or referral), and setting doses. That includes interpreting lab results, reading a photo of a wound or rash, and answering a patient's questions about their symptoms by message.

The reason is not that the model is always wrong. A language model works by predicting the most likely continuation of text, as the lesson on why models make things up explains. It does not examine the patient, does not hold their history, does not know what medicines they take, and carries no professional responsibility. When its footing is weak, it still writes a fluent, confident answer. In other fields a made-up answer is a nuisance; in a clinic, a made-up answer can hurt someone.

Clinical decisions in Indonesia are made by medical and health workers within their competence and authority, under the standards in force at that clinic. The clinic has official references for that, such as clinical practice guidelines, a formulary and referral procedures, written and checked by clinicians. An answer from a general AI service is not one of those references, even for a question that sounds general, such as a medicine's usual dose. If your clinic provides a staff-only tool that answers from the clinic's approved documents, such as its formulary or clinical SOPs, follow that tool's rules: its answers are checked against the source document, it is for staff only, and it never becomes a dose or treatment decision for a particular patient. A clinician decides.

To decide quickly, use one test question: would the AI's answer affect what is done to a patient's body? If yes, stop. That question belongs to a clinician.

Most clinical requests can be turned into a safe task. "Which medicine for this complaint?" becomes "make a reply template that points the patient to the nurse on duty". "Should this patient go first?" becomes "tidy the registration flow that holds the emergency criteria from the doctor in charge", with criteria written by the doctor, not by AI. "Explain this lab result" becomes "make a general leaflet on how the clinic delivers test results and who explains them". The clinical content always comes from clinicians; AI only helps with shape, wording and order. This is what human in the loop means in a clinic: the machine helps write, an authorised person decides.

The biggest temptation comes when the clinic is busy: a long queue, the doctor with another patient, and a colleague thinking AI could answer faster. Have your steps ready before that moment arrives.

First, stop it calmly and name the rule, not the person: "In our clinic, questions about patients do not go to AI." Second, offer the right path: call the nurse or doctor on duty, or follow the clinic's referral and emergency procedure. Third, if patient data has already been pasted, do not hide it; report it the same day to your manager and the person appointed for data protection, as the lesson on data that must not go in explains. Fourth, if a patient arrives with an answer from AI, do not argue; note that the patient wants to discuss it and refer them to a clinician.