🐯虎嗅•Freshcollected in 31m
When AI Diagnoses Clash With Doctors
💡Real patients are challenging doctors with AI answers—an urgent case study in medical safety, trust, and product design.
⚡ 30-Second TL;DR
What Changed
Doubao reportedly handles approximately 20–30 million health-related consultations per day, according to a 2026 podcast disclosure.
Why It Matters
Consumer medical AI is already affecting real clinical workflows before governance, evaluation, and liability standards are mature. AI builders need to treat uncertainty communication, escalation, provenance, and clinician handoff as core product requirements rather than optional safety layers.
What To Do Next
Prototype a clinician-in-the-loop triage flow with the OpenAI Responses API that cites source evidence, exposes confidence, and escalates red-flag symptoms instead of issuing definitive diagnoses.
Who should care:Enterprise & Security Teams
Key Points
- •Doubao reportedly handles approximately 20–30 million health-related consultations per day, according to a 2026 podcast disclosure.
- •Patients have challenged doctors over CT interpretations, medication choices, anesthesia plans, and available hospital procedures.
- •AI recommendations may misclassify ambiguous symptoms, such as directing possible stroke patients toward benign ear-stone evaluations.
- •Medical AI usage is expanding rapidly: China's medical large-model count rose from 61 in 2023 to 311 by June 2025.
- •Doctors often need to spend significantly more time explaining disagreements, increasing operational and liability pressures.
🧠 Deep Insight
Background and context from public sources — not the original article. 28 sources cited.
🔑 Enhanced Key Takeaways
- •Doubao, developed by ByteDance, was launched in August 2023 and by May 2026 had grown to 330 million users, establishing itself as China's most popular AI chatbot.
- •A 2025 China Digital Life Physician study, surveying 5,000 doctors, revealed that 91% of physicians had used or were familiar with AI tools, and 93% believed AI offered more benefits than drawbacks, although 83% maintained that final diagnostic and treatment decisions should remain with human doctors.
- •Ant Group's AI-native health app AQ, also known as Ant A-Fu, was the most frequently recommended AI service by doctors in a January 2026 survey, reporting approximately 30 million monthly active users and processing over 10 million health-related questions daily.
- •China currently lacks unified healthcare AI legislation, but existing regulations such as the Medical Devices Supervision Regulation (revised in 2024) and the Interim Measures for the Administration of Generative Artificial Intelligence Services (Gen AI Measures) are applicable to medical AI.
- •Doctors in lower-tier cities in China reported a higher familiarity with AI (97%) compared to their counterparts in top-tier cities (89%), suggesting a broader adoption and accessibility of AI tools in regions with potentially fewer medical resources.
📊 Competitor Analysis▸ Show
| Company/Product | Key Features | Pricing | Benchmarks/Usage |
|---|---|---|---|
| Doubao (ByteDance) | Multimodal AI assistant (text, image, voice generation, AI-powered search), conversational memory, content generation (articles, code), image creation, multimodal understanding (PDFs, Word, Excel, PPT, TXT), voice capabilities. International version (Dola) uses OpenAI's GPT and Google's Gemini. | Subscription tiers introduced May 2026. Doubao 2.0 priced at approximately one-tenth the cost of Western rivals like GPT-5.2. | Doubao 2.0 (Feb 2026) benchmarked against OpenAI's GPT-5.2 and Google's Gemini 3 Pro on long-chain reasoning. 330 million users by May 2026. |
| Baidu Health (Youyi Assistant, Melody) | Youyi Assistant: AI agent for doctors (research, writing, clinical task execution), literature retrieval, academic writing, medical record generation, data organization, online consultations, patient management. Melody: Converses with patients, collects symptom data, helps determine doctor visits, builds patient profiles. Baidu Health platform: Customized health systems, online advice, medical tests, hospital services, 3D images. | Not explicitly found. | Baidu Doctor platform has approximately 360,000 registered physicians. |
| Tencent Health (HunYuan, Tencent Miying, Tencent Healthcare mini program) | HunYuan (launched 2023): Intelligent dialogue, case structuring, detailed imaging reports. Tencent Miying: Medical imaging (early esophageal cancer screening, diabetic retinopathy, colorectal adenomas). Tencent Healthcare mini program: Booking appointments, health checks, vaccinations, online pharmacies, health records, health education. Vector database for domain-specific medical knowledge. | Not explicitly found. | Tencent Miying achieved 90% accuracy in early esophageal cancer screening and 97% accuracy for diabetic retinopathy. Cell morphology analyzer achieves up to 95% accuracy. |
| Alibaba Health (AQ/Ant A-Fu, Hydrogen Ion, Coca AI) | AQ/Ant A-Fu: Addresses common health concerns, promotes healthier habits, connects users to hospital digital services, online consultations, AI Doctor Agents trained on physician experience. Hydrogen Ion: Evidence-based medical AI assistant for clinicians and researchers, focuses on reducing hallucinations with traceable information. Coca AI: Early colorectal cancer detection from CT scans. Also Grape (gastric cancer) and Panda (pancreatic cancer) models. | Not explicitly found. | AQ reported 30 million monthly active users and processed over 10 million health-related questions daily by January 2026. Coca AI achieved 86.6% sensitivity and 99.8% specificity for colorectal cancer detection, outperforming radiologists by 20.4% sensitivity. |
🛠️ Technical Deep Dive
- Doubao is powered by the Doubao Large Model, developed by ByteDance's Volcano Engine, and processes approximately 120 trillion tokens per day.
- The international version of Doubao, known as Dola, utilizes OpenAI's GPT series of large language models and Google's Gemini.
- Key strategies to mitigate AI hallucinations in healthcare include Retrieval-Augmented Generation (RAG), knowledge graphs, self-reflection, specialized training, and human-in-the-loop (HITL) approaches.
- RAG methods have demonstrated a 30-50% reduction in hallucinations, while HITL strategies can achieve up to 95% reduction, though they pose scalability challenges.
- AI hallucinations in medical contexts often stem from probabilistic pattern prediction, gaps or biases in training data, a lack of real-time grounding in verified medical databases, and the overgeneralization of patterns.
- Chinese medical large language models (LLMs) are trained on extensive, domain-specific datasets that include clinical guidelines, authoritative medical textbooks, expert consensus reports, scientific literature, case studies, medical licensing exam question banks, online consultation platforms, and traditional Chinese medicine classics.
- The MedGo model's foundational dataset incorporates 15 key categories of medical data, while ChiMed 2.0, a large-scale Chinese medical dataset, contains 204.4 million Chinese characters, 164.8K documents for pre-training, 351.6K question-answering pairs for supervised fine-tuning, and 41.7K preference data tuples for reinforcement learning from human feedback.
- Multimodal LLMs are being developed to integrate textual data with medical images and genomic data, enhancing comprehensive diagnostic and treatment decision-making.
🔮 Future ImplicationsAI analysis grounded in cited sources
The regulatory landscape for medical AI in China will become more stringent and unified.
The current lack of unified legislation and the rapid proliferation of AI models, coupled with concerns over accuracy and liability, will necessitate clearer, more comprehensive regulatory frameworks to ensure patient safety and trust.
AI will increasingly serve as a primary triage and information-gathering tool, particularly in underserved areas, but human doctors will retain ultimate diagnostic and treatment authority.
While AI can significantly improve efficiency and access to basic health information, doctors widely agree that final clinical judgment must remain with human professionals due to the complexities and uncertainties inherent in medicine and the risks of AI hallucinations.
The development of specialized, high-quality Chinese medical datasets will accelerate, focusing on multimodal data and traditional Chinese medicine, to enhance the accuracy and cultural relevance of domestic AI models.
Existing Chinese medical datasets are often limited in size, domain coverage, and support for advanced training methods, creating a strong incentive for further investment in comprehensive, diverse, and multimodal datasets, including TCM, to improve model performance and address unique local needs.
⏳ Timeline
2017-07
China's State Council released the New Generation Artificial Intelligence Development Plan, prioritizing smart healthcare.
2020-01
China FDA granted the first AI medical device license, marking a milestone for medical AI commercialization.
2023-08
ByteDance launched the Doubao chatbot.
2023-09
China's National Medical Products Administration (NMPA) issued Guiding Principles for the Registration Review of Artificial Intelligence Medical Devices.
2024-11
Doubao became China's most popular AI chatbot with approximately 60 million monthly active users.
2026-01
Ant Group's AI-native health app AQ (Ant A-Fu) reported 30 million monthly active users and processed over 10 million health-related questions daily.
📎 Sources (28)
Factual claims are grounded in the sources below. Forward-looking analysis is AI-generated interpretation.
- wikipedia.org
- citynewsservice.cn
- onestopesg.com
- chambers.com
- aixsociety.com
- medium.com
- pandaily.com
- ifdesign.com
- the-digital-insurer.com
- tencent.com
- tencent.com
- nih.gov
- aithority.com
- asianews.network
- insightscaremagazine.com
- chinadaily.com.cn
- precedenceresearch.com
- scmp.com
- nih.gov
- infosysbpm.com
- arxiv.org
- arxiv.org
- sciltp.com
- researchgate.net
- jmir.org
- triagehealthlawblog.com
- chinadaily.com.cn
- worldhealthexpo.com
📰
Weekly AI Recap
Read this week's curated digest of top AI events →
👉Related Updates
AI-curated news aggregator. All content rights belong to original publishers.
Original source: 虎嗅 ↗
This is a summary, not the original. Read the source, or get the weekly briefing.
Weekly AI briefing
One email a week. Unsubscribe anytime.