AI in homeopathy

"AI in homeopathy" resolves to four distinct technologies that practitioners now meet inside their software: semantic search over a repertory corpus, large-language-model summarisation of case notes, live audio transcription with rubric suggestion, and image-and-photo extraction from intake forms. None of them prescribes a remedy. Each changes one step in the workflow — the rubric lookup, the case write-up, the consultation note, the intake — and each sits inside a tradition with its own case logic and its own clinical literature.

What the four patterns actually are

Across the vendors observable in May 2026, the term resolves to a small set of recognisable patterns.

Semantic search embeds rubrics and a query into a shared vector space and returns nearest-neighbour rubrics. It is the most widely deployed pattern and ships free on most current platforms over the classic repertory corpus.

LLM summarisation converts unstructured case notes into a structured rubric draft. It is typically gated behind a Pro subscription and an AI-credits budget because the underlying provider calls have variable token cost.

Live transcription records a consultation, transcribes it through a speech-to-text provider, and surfaces rubric candidates as the conversation unfolds. It is currently labelled Beta on most platforms and consumes both AI credits and transcription minutes; the raw audio is generally not retained server-side, only the transcript.

Image extraction reads handwritten or printed intake forms and proposes structured case fields for the practitioner to confirm.

These four patterns do not overlap with the older "expert system" lineage that some classical repertorisation programs still describe themselves with. An expert system encodes a hand-curated rule base — a set of if-then rules written by a domain expert and fixed at compile time. The 2026 systems are learned statistical models: trained on large corpora of text and rubric data, updated by the vendor as the underlying model provider releases new versions, and capable of handling phrasing no rule author anticipated. The user-facing surface — a rubric list, a remedy chart — looks familiar to anyone who learned software in the 1990s or 2000s, but the underlying retrieval has changed in a way the marketing vocabulary does not always make clear. The model and provider are usually disclosed in the privacy or processing documentation rather than the marketing copy, which is where a practitioner who wants to understand the actual system should look first.

How the four patterns map to the case

None of these patterns selects the simillimum. That remains a practitioner judgement grounded in the full anamnesis and case analysis.

What this changes in practice

Five workflow effects show up consistently in practitioner reports. Time-per-rubric drops for ambiguous symptoms. Post-consultation write-up time drops sharply — twenty minutes to two minutes per case is a typical informal figure for LLM-drafted notes. The number of rubrics surfaced per query rises, with the corresponding requirement that the practitioner discard most of them. Note-keeping shifts toward dictation rather than typing. And — the effect worth watching — newer practitioners tend to lean on the LLM's case structure rather than constructing their own, which is the point at which anamnesis discipline starts to fray.

The practical question for an established practitioner is whether the time compression on lookup and write-up justifies the subscription tier and the AI-credit consumption. Semantic search cuts rubric-lookup keystrokes for the ambiguous clusters — anxiety before sleep, burning pain ameliorated by cold application. Live transcription removes the post-consultation transcription pass entirely. LLM summarisation collapses the write-up pass.

Credits are worth sizing before adopting live transcription as a routine practice. A single LLM summarisation of a standard case note costs a small number of credits. A full hour-long consultation transcribed live costs considerably more, because the token count scales with audio length and the downstream rubric-suggestion pass runs over the full transcript. A high-volume practice — ten or more consultations per week — should model the monthly credit draw against the renewal allowance and decide early whether top-up credits are part of the budget. A 14-day trial with all AI features unlocked is the cleanest way to measure actual consumption against your own caseload; you can run the trial on Similia Pro without committing to a paid tier.

Where the literature is most cautious

The clinical-decision-support literature flags two failure modes that map directly onto homeopathic AI features. The first is the confident-but-incomplete summary: LLM systems produce fluent case write-ups that omit the very mental-emotional or peculiar symptom on which a classical prescription would have turned. The summary reads well; the simillimum is no longer visible in it. The discipline is to read the raw notes, not the summary, when the case is unclear.

The second is automated rubric mapping without an uncertainty estimate. Semantic retrieval surfaces more candidate rubrics per query than classical hierarchical lookup — useful for ambiguous symptoms, but it inflates the practitioner's curation burden and, when the ranking is treated as authoritative, can drift toward over-prescription. The classical hierarchy is still doing real work; the semantic ranking sits on top of it, not in place of it.

Live transcription carries its own data-handling weight. The consultation recording contains the patient's voice and medical history in raw form, which sits at the upper end of sensitivity under HIPAA and GDPR. The relevant question for any vendor is whether audio is retained at all, where the speech-to-text provider sits, and whether the transcript itself is encrypted at rest — answers that belong in the processor documentation, not the marketing page.

What stays a practitioner judgement

None of the four patterns selects the simillimum. None of them performs the case analysis. None of them differentiates between two remedies with overlapping rubrics on the basis of a peculiar mental, a modality, or a constitutional reading. Those are the steps where the practitioner's training is doing the work, and where the AI surface is at its most misleading — because a fluent rubric draft and a fluent case summary look like the work is done when the central interpretive act has not yet begun.

A semantic-search hit is a candidate, not a rubric selection. An LLM summary is a draft, not a case. A transcribed consultation is a record, not an anamnesis. The discipline is to treat the AI outputs as the input to the practitioner's analysis, not its output.

References

Boericke, W. (1927) Pocket Manual of Homoeopathic Materia Medica with Repertory, 9th edition, Boericke & Tafel, Philadelphia.

Hahnemann, S. (1842) Organon der Heilkunst, 6th edition.

Hering, C. (1879–1891) The Guiding Symptoms of our Materia Medica, 10 vols., Estate of Constantine Hering, Philadelphia.

Kent, J. T. (1897) Repertory of the Homoeopathic Materia Medica, Ehrhart & Karl, Chicago.

Clarke, J. H. (1900) A Dictionary of Practical Materia Medica, Homoeopathic Publishing Company, London.

Sutton, R. T., Pincock, D., Baumgart, D. C., Sadowski, D. C., Fedorak, R. N. and Kroeker, K. I. (2020) An overview of clinical decision support systems: benefits, risks, and strategies for success, npj Digital Medicine, 3, 17.

Topol, E. J. (2019) High-performance medicine: the convergence of human and artificial intelligence, Nature Medicine, 25, 44–56.