On May 13, 2025, Alibaba Health (阿里健康) officially launched Hydronium (氢离子) — a medical AI assistant targeting clinical physicians and researchers. The product debuts with an exclusive partnership with BMJ Publishing (70 journals) and a declared hallucination rate 2–3 times lower than domestic competitors.
Key takeaways
- Hydronium launch date: May 13, 2025, by Alibaba Health
- Hallucination rates in general LLMs: Grok 3 — 33.6%, DeepSeek DeepThink — 25% (Royal College of Surgeons of England study)
- RAG in clinical contexts increases unsupported claim rate from 5% to 43.6% (8.7x) — medRxiv study
- Hydronium implements a four-layer evidence architecture: PICO+GRADE, structured RAG, fine-tuning with Rubrics, Experts-in-the-Loop
- Knowledge base: 60,000+ drug monographs, 30,000+ clinical guidelines, 4 million+ case studies
The problem: hallucination in medical AI
Clinical medicine requires source certainty, not probabilistic answers. A study published in the official journal of the Royal College of Surgeons of England found that over one-third of citations generated by popular AI models in a surgical context were fabricated or incorrect. Grok 3 hallucinated in 33.6% of cases; DeepSeek DeepThink in 25%. Nearly half of leading models do not disclose sources in medical answers by default.
The industry's standard workaround has been RAG (retrieval-augmented generation) — feeding models fragments of patient histories, guidelines, and scientific publications before generating responses. However, a study published on medRxiv in February 2026 challenges this assumption: after implementing RAG in a clinical context, the unsupported claim rate rose from a baseline of 5% to 43.6%, representing an 8.7-fold increase in factual error risk. The cause lies in the structure of clinical literature — semantically similar passages may refer to different patient populations, different time points, or mutually contradictory trial results.
Architecture: four layers of evidence certification
Hydronium addresses this with a four-layer architecture that Alibaba Health calls "evidence-based medicine certification" (循证医学).
Layer one — understanding medical evidence. The system converts clinical texts into structured evidence units using the PICO framework (Population, Intervention, Comparison, Outcome) and assesses their reliability using the GRADE scale — one of the two principal evidence grading systems in global medicine, adopted by the WHO and over 100 medical organizations.
Layer two — structured RAG. PICO is injected as a structured query rather than a keyword search. Instead of searching for "ibuprofen in children with fever," the system automatically formulates a clinical question: in children with fever (P), does ibuprofen (I) compared to paracetamol (C) show different speed of action and adverse effects (O)? This prevents semantic match without clinical match.
Layer three — fine-tuning with Reward and Rubrics. The model is fine-tuned not on linguistic style, but on evidence compliance rules: a Reward model defines "what constitutes a good answer," while Rubrics translate evidence-based medicine requirements into measurable scoring criteria.
Layer four — Experts-in-the-Loop. A committee of over 300 specialist physicians — functioning as "attending physicians" and "chief examiners" — reviews AI outputs and identifies weaknesses in the preceding three layers. Validation is not an endpoint: every identified weakness is a signal to correct layer one, two, or three.
Data and partnerships
Validation architecture has value only with appropriate source data quality. Hydronium declares access to:
- 60,000+ drug and active ingredient monographs
- 30,000+ domestic and international clinical guidelines
- 4 million+ case studies from scientific literature
- PubMed, Google Scholar, and domestic scientific journals
On May 13, 2025, Alibaba Health announced an exclusive partnership with BMJ Publishing — making Hydronium the first AI assistant in China enabling access to 70 BMJ journals from a single platform. For 76% of Chinese physicians who previously lacked access to leading journal literature, this represents direct access to top-tier evidence.
Comparison to UpToDate
The product is positioned as a rival to UpToDate (UTD) — the widely used evidence-based clinical decision support system. The declared difference from Alibaba Health: Hydronium accepts queries in natural language, by voice, and with images — shifting the interaction from a knowledge base search to a conversation with a fellow clinician.
One beta-testing cardiologist described a clinical case on a medical community platform: a STEMI patient with acute heart failure requiring confirmation of a ticagrelor dose based on eGFR 65. A query to Hydronium returned a result in 3 seconds, with a citable reference to the 2025 Chinese Cardiovascular Society guidelines and the ticagrelor prescribing information. The same process using PubMed, a PDF guideline, and a drug leaflet had previously taken 15–20 minutes.
The physician logged into Hydronium 193 times over 88 days during the closed beta period before the official launch.
Why this matters
The hallucination problem in medical AI is structurally deeper than in most other domains. Errors in marketing content or code are detectable and reversible. In a clinical context, a wrong dosage recommendation or a missed contraindication carries direct patient safety risk. Until now, the industry response was RAG — but as the medRxiv study shows, RAG without deep understanding of clinical document structure can worsen factual accuracy by an order of magnitude.
Hydronium proposes a layered approach: rather than fixing a single problem, it attempts to close the loop from evidence structuring through retrieval, fine-tuning, and expert validation. This differentiates it from general-purpose models supplemented with medical knowledge — and positions it closer to a clinical decision support system with a native conversational interface.
The key question remains open: can validation by 300 physicians adequately cover the full spectrum of clinical specialties and the continuous evolution of guidelines? The scalability of the expert layer will be the test Hydronium must pass as it moves beyond beta.
What's next?
- Alibaba Health announced continued expansion of the BMJ partnership and domestic medical association agreements — additional exclusive content deals are planned for the second half of 2025.
- The system has been available for download since May 13, 2025; beta testers evaluated it in a closed group before official launch.
- The Experts-in-the-Loop architecture requires continuous specialist updates — the ability to scale this layer as the user base grows will determine the product's long-term competitive position.
Sources
Royal College of Surgeons of England — Trust, truth and transparency: analysing the references underpinning AI-generated surgical information
Alibaba Health — official Hydronium product launch presentation, May 2025