Critical reasoning is an essential component of medical training, enabling students to develop complex cognitive abilities that directly contribute to improved decision-making, deeper clinical insight, and safer patient care [1]. This mode of thinking extends beyond the mere recall or application of knowledge; it encompasses problem analysis, evidence evaluation, and the formulation of well-informed clinical judgments [2]. As such, critical reasoning is widely recognized as a cornerstone of high-quality medical education and professional practice [3].

Over the past few decades, technological advancements have profoundly transformed educational ecosystems, including those within medical schools [4]. The advent of the internet and metasearch engines revolutionized access to scientific information, democratizing medical education on a global scale [4-6]. More recently, a new paradigm shift has emerged with the rise of generative artificial intelligence (GenAI) [7]. Tools such as ChatGPT, Claude, DeepSeek, Gemini, Perplexity, LLaMA, and Google Med-PaLM are becoming increasingly embedded in the academic routines of medical students [8,9]. From supporting clinical reasoning and diagnostic processes to assisting with critical essay writing and literature reviews, these platforms are rapidly reshaping the landscape of medical education [10].

This manuscript does not present original data but offers a critical synthesis of current evidence and theoretical perspectives on the role of GenAI in medical education. It explores key issues such as the balance between technological assistance and reflective thinking, the role of faculty guidance, and the ethical implications of artificial intelligence (AI) use. Rather than providing definitive answers, this paper aims to inform pedagogical discussion and support future curricular development.

The aim of this paper is to critically examine the dual impact of GenAI on critical thinking in medical education, exploring both its potential to enhance cognitive skills and the risks it poses to cognitive autonomy.

GenAI is a subfield of artificial intelligence that employs advanced machine learning models to generate humanlike language [10]. Large language models such as ChatGPT, Gemini, Claude, LLaMA, and Mistral are built on transformer architectures that leverage self-attention mechanisms to evaluate the contextual relevance of words within a sequence [11]. Trained on massive datasets, these models can generate coherent, context-sensitive, and semantically rich responses across a wide range of tasks [12].

Although the current literature remains limited, recent empirical studies offer a more nuanced understanding of GenAI’s measurable cognitive effects in real-world medical education settings. Roos et al [27] conducted a large-scale study among German medical students, comparing the performance of GPT-4, Bing, and GPT-3.5 on items from the 2022 German Medical State Examinations. GPT-4 and Bing significantly outperformed students, achieving correct response rates of 88.1% and 86%, respectively, compared to 74.6% among the students. These results suggest considerable potential for AI-assisted preparation tools, although the study did not assess higher-order cognitive outcomes [27].

Regarding learning outcomes, Sakelaris et al [28] found no statistically significant differences in the exam scores between students who used GenAI tools (primarily ChatGPT) for studying and those who did not, indicating a limited direct impact on knowledge acquisition and academic performance in preclinical settings. Nonetheless, broader concerns regarding cognitive offloading and critical thinking persist.

In a separate study, Güvel et al [29] compared the performance of GenAI tools—including ChatGPT-4o, Gemini, and Claude—with that of human experts in generating case-based rational pharmacotherapy questions. Although AI-generated items showed comparable discrimination indices and correct answer rates (with Claude producing the highest proportion of error-free items—12 out of 20—and ChatGPT generating the fewest unusable items—5 out of 20), expert validation remained essential to eliminate flawed or unsuitable content.

More instructively, 2 recent randomized controlled trials clarify how instructional design mediates GenAI’s impact on critical thinking and engagement. Shalong et al [30] demonstrated that LearnGuide, a ChatGPT-based facilitator, improved scores on the Cornell Critical Thinking Test (+7.11; P<.001), self-directed learning (+4.15; P=.01), and cognitive flow in problem-based learning environments. These gains were sustained at a 14-week follow-up, highlighting the potential of structured, reflective GenAI use [30].

Conversely, Çiçek et al [31] found that ChatGPT-generated feedback did not significantly improve critical thinking and was inferior to expert-written feedback in managing complex diagnostic tasks. Nonetheless, students who later learned of the AI involvement exhibited heightened critical awareness, suggesting that transparency and reflection can positively shape digital literacy attitudes (P<.001).

Taken together, these findings underscore the context-dependent nature of GenAI’s impact and reinforce the importance of pedagogical scaffolding, expert supervision, and reflective engagement to preserve cognitive autonomy in medical education settings (Table 2).

In response to the cognitive limitations of GenAI, reasoning AI has emerged as a complementary paradigm [22]. While GenAI is known for its linguistic fluency and pattern recognition capabilities, reasoning AI systems are designed to emulate structured, goal-directed thinking by integrating logical inference, sequential problem-solving, and cognitive reasoning frameworks into their architecture [32]. These systems are typically built on neurosymbolic or neurological models that combine deep learning with symbolic reasoning, enabling them to construct traceable logical chains rather than merely predicting the next plausible output [33].

Although GenAI can generate broad lists of potential diagnoses, it often fails to establish step-by-step reasoning that accurately links clinical signs and laboratory findings. In contrast, a reasoning AI system could model diagnostic probabilities by using Bayesian networks, such as correlating hyperkalemia, peaked T waves, and angiotensin-converting enzyme inhibitor use to support a diagnosis of hyperkalemic αβ insufficiency [34].

Reasoning AI tools such as neurosymbolic workstations or diagnostic decision-support systems based on graphical models can be integrated into emerging educational approaches such as problem-based learning. In this context, they guide students through structured reasoning processes. For example, “Given this patient’s vital signs and laboratory values, propagate the data through an acid–base physiological model, then adjust the anion gap calculation step by step.” This scaffolding reinforces explicit clinical inference across a range of instructional scenarios [34,35].

Although GenAI may stimulate early stages of critical thinking by proposing hypotheses or generating broad explanations, reasoning AI requires disciplined, verifiable reasoning pathways—bringing learners closer to the structured deductive logic required in clinical practice. In curricular terms, an integrated model might begin with GenAI-driven brainstorming (eg, list all possible causes of chest pain) and progress to decision-analysis modules powered by reasoning AI (eg, apply a decision-tree model to differentiate myocardial infarction from pulmonary embolism) [35]. This would ultimately support a complementary AI integration model aimed at enhancing logical, structured reasoning in favor of more reflective, accurate clinical decision-making.

This manuscript presents a theoretical perspective and does not include primary data collection, code generation, or statistical modeling. It is also important to note that the development of this manuscript did not follow a systematic search strategy, which introduces a potential risk of selection bias and the omission of relevant studies. Nevertheless, the analysis is grounded in a critical synthesis of publicly available scientific literature, drawing exclusively from peer-reviewed studies indexed in major academic databases such as PubMed, Scopus, and Web of Science, intentionally selected for their relevance and impact.

Another important limitation is that the cognitive effects of the GenAI tools analyzed are highly dependent on the educational contexts in which they were evaluated. Therefore, the conclusions should be interpreted with caution. The generalizability of the findings is constrained by the heterogeneity of learner populations, curricular frameworks, and levels of pedagogical oversight across the reviewed studies.

Consequently, we propose several future research directions. First, controlled comparative trials are needed to assess the differential impact of GenAI versus reasoning AI systems on the development of critical thinking. Second, it is essential to design and validate instruments capable of measuring cognitive autonomy in AI-mediated learning environments. Finally, ethical and digital literacy interventions should be developed and implemented for both students and educators, aiming to mitigate risks of overreliance and to promote reflective, context-sensitive use of these technologies in medical education.

The emergence of GenAI in medical education represents a profound inflection point—rich in potential, yet fraught with risk. These tools can support overwhelmed learners and educators, personalize instruction, and facilitate the development of cognitive skills. However, they are not pedagogically neutral. Without intentional design and critical engagement, GenAI may erode the very attributes of clinical judgment, ethical reasoning, and intellectual autonomy that define competent physicians.

The future of medical education lies not in rejection or blind adoption, but in thoughtful, ethically grounded integration. This requires digital literacy training, faculty-mediated scaffolding, and curricular frameworks that reinforce reflective reasoning. GenAI should be viewed not as a threat or a panacea but as a catalyst that exposes both the strengths and vulnerabilities of the current educational models. The challenge ahead is to prepare physicians who are not only technologically fluent but also critically empowered. That—more than any algorithm—will shape the future of care.