AI case taking in homeopathy: how it works and what to check
AI case taking records the consultation, transcribes it, picks out the symptoms and suggests rubrics. It saves writing, not judgement: transcripts drop words, models misplace symptoms, and the peculiar symptom is the easiest thing to lose.
Evaluating EzHomeo? Start a reviewable consultation in the app. 10 consultation credits free, no card required.
The standard: the Organon
The Organon asks the physician to write down “accurately all that the patient and his friends have told him in the very expressions used by them”, without interrupting (§84). Questions come afterwards, never framed to suggest the answer (§87). The remedy rests chiefly on “the more striking, singular, uncommon and peculiar (characteristic) signs and symptoms” (§153).
Recording lets you listen instead of write, which helps with §84. An AI that paraphrases the patient into textbook language works against §84 and §153 at once.
How it works
- Record, with consent. The doctor starts the recording.
- Speech to text. A speech model writes the transcript and separates the speakers.
- Structure. A language model sorts what was said into mind, generals and particulars, with modalities.
- Suggest rubrics. Each symptom is matched to candidate rubrics in a named repertory.
- Repertorise. The kept rubrics are scored against the repertory to rank remedies.
- Draft the record. The software drafts a case record for the doctor to correct.
Errors carry forward: a wrong word in the transcript becomes a wrong symptom, then a wrong rubric. EzHomeo’s features page shows how it handles each step.
Where it goes wrong
- Draft notes contain errors. In a 2025 study of two commercial AI scribes, 31 of 44 draft notes (70%) had at least one error. Omissions were the most common (Biro et al., JMIR).
- Errors reach daily practice. In a 2025 randomised trial with 238 outpatient physicians, clinicians reported clinically significant inaccuracies “occasionally” (Lukac et al., NEJM AI).
- Indian accents raise the error rate. On Indian-accented English, the best system tested made 7.2% word errors against 2.7% on a standard benchmark. Cloud models tuned for India made about 21% (Svarah, Interspeech 2023).
- Code-mixed speech is harder. Baseline systems in a 2021 challenge made 31 to 32% word errors on test sets that included Hindi-English and Bengali-English speech (MUCS 2021). These were research baselines, not current commercial models.
- Speech-only AI misses what is not said. AI scribes cannot capture nonverbal cues and can attribute words to the wrong speaker (Topaz et al., npj Digital Medicine 2025).
- General chatbots are inconsistent. Given 100 acute cases, four free chatbots included the practitioner’s first remedy in 36.5% of cases on average, and changed answers when the same case was entered again (Doherty et al., Healthcare 2026).
The World Health Organization warns of “automation bias”, where clinicians overlook errors they would otherwise catch. The Society of Homeopaths says clinical responsibility “must therefore remain entirely with the qualified practitioner”.
Check these before you repertorise
| Check | What to look for |
|---|---|
| Negations | Did “no thirst” become “thirst”? |
| Modalities | Is each worse and better attached to the right symptom, the right way round? |
| Peculiar symptom | Is it there in the patient’s own words, not smoothed into a common one? |
| Speakers | Did a parent’s words get attributed to the child, or your question to the patient? |
| Observations | Add what you saw. Speech-only AI cannot. |
| Omissions | Read the draft against your memory of the visit. |
| Rubric fit | Does each rubric match what the patient said, in the repertory you chose? |
| Nothing invented | Anything not said should be blank or marked as not mentioned. |
Replay the audio for anything that drives the prescription.
Consent and privacy in India
- Ask first. Tell the patient you will record and why, and note the consent in the case.
- DPDP consent rules start in May 2027. Consent must then be free, specific, informed, unconditional and unambiguous, and as easy to withdraw as to give. The Rules were notified in November 2025 with an 18-month phase-in.
- Keep identifiable cases out of general chatbots. The Society of Homeopaths asks practitioners not to share patient data with systems that lack safeguards.
Questions to ask any tool
- Can I replay the original audio to check the transcript?
- Does each symptom show the words it came from?
- Can I add, remove and weight rubrics, and which repertories can I use?
- Does it ever suggest a potency or dose? It should not.
- Where are recordings and records stored, and how do I delete them?
EzHomeo answers yes to the first three, never suggests a potency or dose, and keeps its app servers and patient database in Mumbai. Test EzHomeo against this list with a synthetic consultation.
Related reading
Compare tools in repertorisation software compared, see EzHomeo compared with other software, or read what a complete case record holds.
AI Case Taking in Homeopathy: frequently asked questions
Sources and further reading
Last checked .
- Hahnemann, Organon of Medicine, Dudgeon translation (1916 printing, US National Library of Medicine scan)
- Biro et al., Accuracy and safety of AI-enabled scribe technology, JMIR 2025
- Lukac et al., Ambient AI scribes in clinical practice: a randomized trial, NEJM AI 2025
- Topaz et al., Beyond human ears: the risks of AI scribes, npj Digital Medicine 2025
- Javed et al., Svarah: evaluating English ASR systems on Indian accents, Interspeech 2023
- Diwan et al., MUCS 2021: code-switching speech recognition for Indian languages, Interspeech 2021
- Doherty et al., Comparing AI chatbots to live practitioners of homeopathy, Healthcare 2026
- World Health Organization, Guidance on large multi-modal models (2024)
- The Society of Homeopaths, Public statement on AI
- Press Information Bureau, DPDP Rules, 2025 notified