Use of large language models in German general practice: A cross-sectional survey

In the quiet hum of a doctor's office, a new kind of assistant has begun to appear, not in a white coat, but as a line of code. This assistant is a large language model, a type of artificial intelligence trained on vast libraries of text to understand and generate human language. Unlike older computer programs that follow rigid, pre-written rules, these models can answer questions, summarize complex information, and draft documents by predicting what words should come next, much like a very advanced autocomplete that has read almost everything ever written. For decades, the relationship between a doctor and a patient has been a private conversation, a two-person exchange of symptoms, history, and care. Now, a third element is entering that space, raising a simple but profound question: how do family doctors in Germany feel about bringing this digital tool into their daily work, and are they ready to let it help them decide how to treat their neighbors?

Researchers at the Institute of Family Medicine at Philipps-Universität Marburg set out to answer this by asking German family doctors directly. They conducted a survey between February and June 2026, inviting physicians from solo practices, group clinics, and medical centers across the country to share their experiences. The goal was not to guess what might happen in the future, but to capture a clear snapshot of what was actually happening right then. They asked 206 doctors who completed the questionnaire about which tools they used, how often, and for what tasks. They also inquired about the hurdles that stopped them from using these tools more, whether they trusted the information the models provided, and how they thought this technology would change the bond between doctor and patient.

The results revealed that these digital assistants are already part of the landscape, though their role is still being defined. About 80 percent of the doctors surveyed said they used large language models in their professional work. The most popular tool was ChatGPT, used by roughly two-thirds of those who used any model at all. However, the way doctors used these tools was quite specific. They did not typically use them to write patient reports or handle administrative paperwork, tasks that might seem obvious for a computer. Instead, they turned to these models primarily for knowledge and decision support. When a doctor encountered a rare condition they were less familiar with, or needed to double-check a treatment plan, they asked the model for information. The doctors found these tools most helpful for digging into unfamiliar medical topics, supporting their diagnostic thinking, and generating treatment suggestions.

Despite this growing usage, the doctors approached the technology with a healthy dose of caution. The most significant barrier preventing them from using these models more freely was the fear that the information provided might be wrong. These models are known to sometimes produce answers that sound confident but are factually incorrect, a phenomenon the researchers noted as a major concern. Alongside this fear of inaccuracy, doctors were worried about data protection and the legal gray areas surrounding liability. If a model gave bad advice, who was responsible? These uncertainties kept many from integrating the tools more deeply into their workflow. Interestingly, the doctors who used these models most frequently were generally younger and had fewer years of professional experience, while the frequency of use did not differ between men and women.

The study also looked closely at the moment of care itself, when the doctor sits with the patient. While 41 percent of the doctors admitted to using these models while a patient was in the room, a notable portion of them did so without telling the patient. About 26 percent of all respondents said they used the tool during a consultation without disclosing it, while only 16 percent said they were open about it. This gap between practice and transparency is significant, especially since half of the doctors believed that patients would actually respond positively if they knew the doctor was using an AI tool to help. The doctors generally did not believe that patients would stop visiting them just because they had looked up their symptoms online first. Instead, they expected that patients arriving with information from these models would lead to more discussion, turning the visit into a three-way conversation between the patient, the doctor, and the information the patient brought with them.

Perhaps the most reassuring finding for the future of the profession was the doctors' view on their own role. The vast majority firmly rejected the idea that these models could replace them. They believed that while the tools could help with facts and data, they could not replicate the human capacity for empathy or the deep expertise required to navigate complex, real-life situations. Most doctors felt that their own knowledge of these tools was still somewhat low, and a large majority had never received any formal training on how to use them safely or effectively. Yet, there was a strong desire to learn; nearly two-thirds expressed a willingness to attend training sessions in the future.

Looking ahead, the doctors surveyed expect these tools to become a standard part of primary care, much like a stethoscope or a computer is today. They anticipate using them more often as the technology improves and the rules around them become clearer. The study suggests that the integration of artificial intelligence into family medicine is not a matter of if, but how. It will likely happen slowly, driven by doctors who want to make their work easier and their decisions more informed, but only if the tools are reliable, the legal frameworks are clear, and the training is available to ensure that the human element of care remains at the center of the practice.