DeepSeek Use Reaches a Meaningful Share of Medical Interns in a Chinese Teaching Hospital
Artificial intelligence is no longer an abstract topic in medical education. In a cross-sectional survey at the Second Affiliated Hospital of Anhui Medical University, DeepSeek was already part of the learning routine for a substantial minority of medical interns, with about one-third reporting use. Among those who used it, the pattern was not casual or experimental. Most users returned to the tool regularly, and more than half said they used it several times per week.
The finding matters because intern-level training is where habits become durable. When a digital tool becomes part of the daily workflow at this stage, it can shape how learners search for information, draft assignments, translate material, and judge the reliability of what they read. The study adds a concrete view of how a domestic large language model is entering routine educational practice in China, not as a theoretical innovation but as an ordinary study aid.
Assignment Support and Translation Drove Most Use
DeepSeek was used most often for completing assignments, searching for information, translation, and polishing written text. Those uses fit the strengths and risks of large language models. They are fast, fluent, and convenient, which makes them attractive when interns are managing dense coursework and clinical duties. They also create a dependence on generated prose that can be polished enough to look authoritative even when the underlying content needs verification.
A smaller subset of interns reported using the tool for data analysis. That detail is easy to overlook, yet it points to a broader shift in how learners may come to view artificial intelligence: not only as a writing assistant, but as a general-purpose academic utility. In medical training, that raises a practical question about boundaries. A tool used to rephrase text may be harmless if supervised, but a tool used to interpret data or synthesize clinical reasoning demands much stricter checking.
The advantages and disadvantages were broadly balanced in the eyes of most users. Only a minority thought the disadvantages outweighed the benefits. That balance is telling. It suggests interns are not rejecting the technology, but neither are they blindly embracing it. They appear to be working out, in real time, where convenience ends and judgment begins.
Residence and Major Shaped Who Used the Platform
After adjustment, residence and major remained associated with DeepSeek use. Age and sex did not. Urban interns were more likely to report using the platform than rural interns, a difference that likely reflects access, exposure, and familiarity with digital learning tools rather than clinical aptitude. In practical terms, this is a reminder that adoption of AI in education does not happen on an even playing field. Connectivity, prior experience with online tools, and the culture of a training environment can influence whether a platform becomes part of daily study habits.
Major also mattered. Compared with clinical medicine interns, preventive medicine interns were less likely to report use, while nursing interns were more likely to do so. The result hints at differing educational needs and digital workflows across training tracks. Clinical medicine interns may turn to AI for broad information retrieval and text support, whereas nursing interns may find translation, explanation, and writing assistance especially useful in their coursework. Preventive medicine interns may rely less on this style of support, or may have different learning norms that reduce uptake.
Among users, overall attitudes differed across majors, while usage frequency and the perceived balance of benefits and drawbacks did not. In other words, trainees from different disciplines may agree on how often the tool is useful and whether it feels balanced overall, yet still judge it through different academic lenses. Those differences are likely to shape future expectations for AI literacy in medical school and hospital-based training.
Verification and Academic Integrity Sit at the Center of the Debate
The educational appeal of large language models is easy to explain. They compress time, lower the effort needed to draft text, and provide instant access to apparently well-structured answers. In a medical training environment, that can support rapid review and reduce friction during routine study tasks. The same qualities create risk. Because the output is fluent, users may not pause long enough to verify whether a response is accurate, current, or suitable for a specific question.
That tension is especially important in medicine, where small errors can be consequential. Even when the immediate task is only an assignment, the habits formed around source checking and critical reading carry over into clinical practice. If interns grow comfortable accepting machine-generated answers without verification, they may weaken the discipline that medical decision-making requires. The study therefore speaks to more than technology adoption. It points to the need for institutional guidance on verification of AI-generated information and on academic integrity before these tools become fully normalized.
A quiet but important message runs through the data: use is already happening, and it is happening in an uneven way across training backgrounds. That reality makes policy harder, not easier. Training programs will need to decide not only whether to allow AI tools, but also how to teach students to question them, cross-check them, and disclose when they are used.
A Cross-Sectional Snapshot With Clear Limits
This was a descriptive cross-sectional survey, so it cannot show how DeepSeek use changes over time or whether use improves educational performance. It also cannot determine whether interns who favor the tool differ in unmeasured ways from those who do not. The study captures behavior at a single point in one Chinese teaching hospital, which narrows how far the findings can be generalized.
Even so, the dataset offers a rare look at how a domestic large language model is being used by medical interns in day-to-day training. The sample included 634 interns from clinical medicine, preventive medicine, and nursing, giving the survey enough scale to detect meaningful differences across groups. The next step is not simply more adoption data. It is clearer instruction on responsible use, stronger attention to verification, and better-defined norms for what AI should and should not do inside medical education.