Systematic evidence mapping shows artificial intelligence dependence claims outnumber corresponding assessments in health professions education
This systematic evidence mapping reveals that in health professions education, claims regarding artificial intelligence dependence significantly outnumber corresponding assessments capable of verifying such dependence, leaving the frequency, causal effects, and patient consequences of AI reliance unestablished.
Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
In the classrooms and training centers where future doctors, nurses, and pharmacists learn their craft, a new tool has arrived with the promise of transforming education. This tool is artificial intelligence, a technology capable of retrieving vast amounts of information, offering case-based practice, and providing instant feedback on complex decisions. For learners, this assistance can make difficult tasks feel easier and faster. However, a fundamental question has emerged that goes beyond simple convenience: if a student relies on this digital assistant to solve a problem, do they actually learn how to solve it themselves? The concern is not that the tool fails, but that the learner might become so accustomed to the support that their own independent ability fades. This is the core of the "dependence" question. It asks whether the skills and reasoning a student displays while using AI are truly their own, or if they are merely the result of a partnership that collapses the moment the machine is turned off. Understanding this distinction is vital for patient safety, because a medical professional must be able to verify information and make sound judgments even when technology is unavailable or, worse, when it provides incorrect advice.
To investigate this, a team of researchers at Sichuan University conducted a systematic search through the existing scientific literature to see what evidence actually exists regarding this issue. They did not simply ask if AI helps students learn; they looked specifically for studies that claimed AI might cause dependence and then checked whether those same studies had actually measured that dependence. The researchers focused on six specific ways dependence might show up: over-reliance on the machine, offloading mental work to it, trusting the machine too much even when it is wrong, handing over decision-making authority, a weakening of reasoning skills, and a degradation of practical abilities. They gathered nearly 14,000 records from four major scientific databases and, after a rigorous screening process to remove duplicates and irrelevant papers, narrowed their focus to 150 high-quality reports that involved human participants in health professions education. These reports formed the core of their analysis.
The researchers then performed a careful matching exercise. They looked at each of the 150 reports to see if the authors made an explicit claim about one of the six dependence phenomena. If a report claimed that students were becoming over-reliant on AI, for instance, the researchers checked if that same report included a specific test designed to measure over-reliance. They found a striking gap between what was claimed and what was tested. Out of the 150 reports, 82 made at least one explicit claim about dependence. However, only 13 of those reports included an assessment that met the strict design criteria needed to actually test for that specific phenomenon. Even more telling, only nine reports managed to align a claim with a proper assessment of the same phenomenon. This means that for the vast majority of studies discussing the risks of AI dependence, the authors did not provide the specific evidence required to prove that the risk was actually happening.
The situation was particularly clear when looking at specific types of dependence. The researchers found that claims about "over-reliance" appeared in 75 reports, but only five of those reports used a design capable of testing it. Claims about "weakened reasoning ability" appeared in 57 reports, yet only two included a proper test. For "skill degradation," 18 reports claimed it was happening, but none of them aligned the claim with a study design that could actually measure a decline in skill after the AI was removed. The few studies that did attempt to measure these effects often lacked a crucial element: a clear confirmation that the AI was unavailable during the final test. Without knowing for certain that the student was working alone, it is impossible to tell if their performance was independent or still influenced by the tool. Some studies did look at performance after a delay, but in most cases, the researchers could not verify if the students were still using the AI or if the test conditions truly isolated the learner from the machine.
The findings suggest that while the scientific community is actively discussing the potential dangers of AI dependence, the evidence base is not yet ready to confirm how often these dangers occur or what their true impact is. The researchers noted that some studies showed students performing better overall when AI was present, but this improved performance often masked a different reality: when the AI gave wrong advice, the students' ability to spot the error dropped significantly. This highlights that a learner's success with a tool is not the same as their independent capability. The study also found that very little research has looked at how these educational dynamics affect real patients, with only a handful of reports involving actual patient data or outcomes. The authors concluded that the current body of research cannot estimate the frequency of AI dependence, its causal effects, or its consequences for patient care.
This does not mean that AI has no place in health education, nor does it prove that dependence is inevitable. It simply means that the current scientific literature has not yet provided the specific, high-quality evidence needed to answer the question. The researchers emphasize that different types of dependence require different kinds of proof. For example, proving that a student has lost the ability to reason without help requires a test where the student works alone after a period of using the tool, with a clear baseline to compare against. Until more studies are designed with these specific requirements in mind, the conversation about AI dependence remains largely theoretical. The path forward involves building studies that can clearly separate the performance of the learner from the performance of the learner-plus-tool, ensuring that when the digital support is withdrawn, the human professional remains fully capable.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.