If a doctor has an AI clone, will his treatment be different? What other obstacles are there?

With the gradual application of AI virtual clone consultation tools, AI multi-agent technology in the medical field has attracted attention. In reality, patients have obtained health consultation services through doctor AI avatars. Can doctor AI clones be integrated into real diagnosis and treatment? What are the remaining obstacles to large-scale application?

Improve service efficiency

In July this year, the international authoritative academic journal "Nature" published an article, citing the Bauhinia AI Hospital built by the Tsinghua University team as a core case in China's AI intelligent medical field. The multi-agent collaborative ecosystem of Bauhinia AI Hospital has also been officially released recently. Human doctors from different departments work collaboratively with intelligent agents to explore medical service models that cover more scenarios.

The system builds doctor clones with different specialty settings. This kind of avatar must first carry out extensive training in a virtual environment that simulates AI patients. After reaching the set competency standards, it must be deployed to physical medical institutions, collaborate with real doctors, and continuously iterate and optimize in real diagnosis. Qiao Yuchen, chief product officer of Bauhinia Zhikang, said that AI clones can learn to replicate doctors' diagnosis and treatment ideas and working habits, and undertake some administrative work, which is expected to alleviate doctors' administrative burden and improve service efficiency. "More importantly, it allows the clinical experience of senior experts to empower young doctors and grassroots doctors, helping high-quality medical resources sink." Qiao Yuchen said that the system currently has built a matrix of 54 AI doctor agents covering 27 clinical departments, which can cover most common diseases, major diseases and some rare disease scenarios.

It is reported that after the AI ​​avatar enters the medical institution, it continuously aligns the doctor's consultation logic and medical record writing habits during the actual diagnosis and treatment process. At this stage, relevant functional tests of the Bauhinia AI hospital system cover 9 hospitals in Beijing, Jiangsu, Guangxi and other places, and nearly 200 doctors have built exclusive avatars. The pilot institutions include tertiary hospitals and community health service centers, and have been tested and verified in various scenarios such as medical teaching and research, and grassroots diagnosis and treatment.

Different from Zijing Zhikang's layout for hospital diagnosis and treatment scenarios, the AI ​​health application Ant Afu launched by Ant Group focuses on the digital reuse of famous medical resources. Led by eight academicians of the two academies, including Liao Wanqing, Dong Jiahong, and Wang Jun, more than 2,000 doctors across the country have opened AI clones on Ant Afu, covering about 28 specialties. AI clones allow famous doctors to answer users' health questions through application software regardless of geographical or time restrictions. The famous doctor AI avatar does not carry out diagnosis and treatment activities, but only provides health science consultation and solves daily and basic health questions, allowing real doctors to focus on clinical treatment.

Currently, the service radius of AI clones is rapidly expanding. Mao Hongjing, deputy director of Hangzhou Seventh People's Hospital and expert in sleep disorder diagnosis and treatment, is one of the first batch of doctors to participate in the development of AI clones of famous doctors. Mao Hongjing's AI avatar can interact with users in a manner that is close to real doctors and users, and synchronizes consultation content to the doctor's workstation to help doctors quickly complete the collection of patients' basic health information. In the past, it was difficult to find Mao Hongjing's expert account, but now the AI ​​clone is online around the clock to answer questions for people who are concerned about sleep health.

Mao Hongjing, who has been in the industry for nearly 30 years, lamented: “I used to serve up to 10,000 patients a year, but now my AI clones serve 3.68 million patients a year.”

Complement rather than replace

In interviews, many doctors positioned the value of AI clones as "filling in" rather than "replacing".

"Existing medical models mostly focus on disease treatment, but are relatively weak in continuous health management and lifestyle intervention. The diagnosis and treatment threshold needs to move from 'treating the disease' to 'treating the disease before it's too late." Ma Liangkun, chief physician of the Department of Obstetrics and Gynecology at Peking Union Medical College Hospital, said that if AI clones can participate in patients' dietary guidance and postoperative rehabilitation follow-up, there is an opportunity to make up for some of the shortcomings of current medical services.

This "fill-in" value is more obvious at the grassroots level and in remote areas. Fan Haining, director of the Affiliated Hospital of Qinghai University, is an expert in hepatobiliary surgery. He said: "Distance itself is a threshold for medical treatment in plateau areas. Some herdsmen brave the cold wind in winter and ride motorcycles from the countryside to the county and then to the state for more than ten hours to see a doctor in the clinic. After the AI avatar is activated, the herdsmen can consult on health issues online and no longer have to travel for a long time for a small problem."

AI clones are also changing the form of doctor-patient interaction. Duan Tao, director of the Department of Obstetrics and Gynecology at Shanghai Dongfang Hospital, opened his own AI clone and has received 160,000 consultations. Duan Tao introduced that many of the offline clinics deal with basic and repetitive questions related to pregnancy lifestyle habits, and AI clones can help doctors share some of the pressure. In addition, in offline diagnosis and treatment, men cannot accompany their wives for prenatal check-ups, and AI avatars allow men to participate. Data shows that 24% of users who consulted about Duan Tao’s AI clone are male.

According to the design positioning, the AI ​​avatar is a tool for doctors, not a replacement. AI is mainly responsible for information collection, sorting and analysis, and outputting reference suggestions; the final decision-making power of diagnosis, treatment, and medication is always in the hands of doctors. Many industry insiders believe that in the diagnosis and treatment process, AI avatars can currently undertake pre-diagnosis consultation, triage, health file collection, as well as speech transcription during diagnosis, and output reference opinions with medical guideline evidence links. For example, in the emergency scene of a car accident, multiple AI clones in the fields of emergency, orthopedics, neurosurgery, and anesthesia can collect and summarize physical signs information separately, but the overall treatment plan is still determined by the judgment of real doctors. "The work of the intelligent agent is limited to the scope of learned knowledge, and requires doctors to complete comprehensive judgments." An outpatient doctor participating in the test gave an analogy. Facing the AI ​​clone is like bringing a real intern. The doctor is using and teaching, and the AI ​​clone is learning.

AI health tools are also extending to primary health service systems. In August 2025, six departments including the National Health Commission jointly issued "Several Measures to Improve the Experience of Family Doctor Contracting Services", clearly encouraging the use of the "Internet +" contracting process to improve the accessibility and convenience of family doctor services. According to reports, health departments in Shijiazhuang, Hebei, Huangpu District, Shanghai and other places have cooperated with Ant Afu and used it as one of the local "family doctor" contract signing platforms. Residents can complete the contract online through their mobile phones. The platform helps family doctors carry out health management and science popularization services for key groups such as patients with chronic diseases and the elderly.

Facing multiple tests

At present, attempts to integrate doctor avatars into real diagnosis and treatment are gradually underway, but there are still multiple structural obstacles to seeing from simulated environments to actual clinical applications.

The first is safety. He Zhongjun, assistant to the president of Baidu Research Institute, gave an example. In the field of software engineering, if an intelligent agent runs incorrectly, it can be rolled back and repaired. However, the fault tolerance space in medical scenarios is extremely small. How to prevent algorithm errors and establish a reliable safety control mechanism is a problem that the industry must solve.

The second is the evaluation level. How to determine whether the diagnosis and treatment performance of the doctor's avatar is close to that of a real doctor? Qiao Yuchen introduced that "associated evolution" aims to promote the alignment of AI and corresponding doctors in dimensions such as value orientation, diagnosis and treatment logic, and risk judgment preferences. The team internally relies on indicators such as adoption rate, coverage rate, modification rate, and editing rate to measure the alignment effect. After training on hundreds of cases, the alignment performed well in internal testing, but it still needs further testing with large-scale real clinical data.

The key is to determine an application "critical line" and clarify what conditions are met before AI can be authorized to participate in the corresponding diagnosis and treatment links. "Critical standards cannot be generalized." Qiao Yuchen said that the judgment thresholds in different scenarios still need to be explored in practice.

A deeper level is the system barrier. According to current regulations, there are strict entry thresholds for physician qualifications. AI avatars do not have independent practice qualifications, and their output can only be used as auxiliary reference. The relevant regulatory rules for Internet diagnosis and treatment are clear, and prescriptions must not be automatically generated by artificial intelligence. Once the reference recommendations provided by AI deviate, there are currently no legal details on how to divide the responsibilities between doctors, medical institutions and technology providers. In addition, the training of AI clones relies on a large amount of medical records and health data. The collection boundaries, authorization methods and desensitization standards of patients’ personal information also require clear answers at the institutional level. If these problems are not solved, it will be difficult for AI clones to truly move from testing scenarios to the clinical frontline.

He Zhongjun believes that AI tools objectively provide new opportunities for the sinking of medical resources. In the long run, ordinary people may have personal medical intelligence. At the same time, he also reminded that large models iterate quickly and the industry should always maintain a reverence for life while pursuing technical efficiency.