Given their long work hours and stressful work environment, GME trainees are particularly susceptible to burnout, with rates higher than their age-matched peers in non-medical careers and higher than early-career attending physicians (21). Burnout among the academic physicians who comprise most GME faculty also occurs, and may impact the quality of training they are able to deliver (22, 23). Thus, innovations that prevent overwork and burnout have the potential to benefit GME trainees and faculty.

One unintended consequence of the adoption of electronic health records (EHRs) has been a dramatic increase in time spent in documenting clinical encounters. Many physicians now spend as much time documenting in the EHR as they do in patient-facing activities (24). This documentation burden can result in medical errors, threats to patient safety, poor quality documentation, and attrition, and is a major cause of physician burnout (25). Various strategies have been tried to reduce physician documentation burden, including medical scribes and various educational interventions, workflow improvements, and other strategies (26). Given its ability to summarize, translate and generate text, GenAI demonstrates clear potential as a technological aid to alleviate the burden of clinical documentation. The most notable current application is ambient listening tools that use GenAI to transcribe and analyze patient-doctor conversations, converting them into structured draft clinical notes that the physician would theoretically only then need to review for accuracy. Numerous organizations are piloting such technology as of the time of this writing (27), though the few results published so far about real-world performance have been mixed (10, 28, 29). Examples of other less commercially mature concepts for how GenAI could reduce clinical documentation burden include tools to improve medical coding accuracy (30), to generate clinical summary documentation like discharge summaries (31), and to draft GME faculty supervisory notes (32).

In addition to documenting clinical encounters, physicians (including GME trainees) spend large amounts of time in the EHR managing inbox messages, including patient messages, information about tests results, requests for refills, requests to sign clinical orders, and various administrative messages (33). As another major contributor to workload, EHR inbox management is also a cause of burnout (34, 35). This problem came to be of particular importance during the COVID-19 pandemic, where patient messaging increased by 157% compared to pre-pandemic levels (35). LLMs have shown the ability to draft high-quality, “empathetic” responses to patient questions (36). Early efforts to use LLMs for drafting replies to patient inbox messages have shown promising results, with multiple studies showing that LLMs can draft responses of good quality (37, 38) and at least one study showing good provider adoption with significant reductions in provider assessment of multiple burnout-related metrics (11). Multiple health information technology companies, including the largest United States EHR vendor, have already brought GenAI functionality for EHR inbox management to market (39–41).

Simulation-based medical education (SBME) has evolved significantly since the early use of mannequins for basic life support training 60 years ago, and simulation using high-fidelity mannequins and virtual and augmented reality tools are now a vital component of GME. There is a substantial body of evidence confirming the benefits of simulation-based training and the successful transfer of these skills to real patients (42, 43). Simulations are used both to educate and to assess performance in GME. For example, the American Board of Anesthesiology incorporates an Objective Structured Clinical Examination (OSCE) meant to assess communication and professionalism, as well as technical skills, into the board examination process for anesthesiology residents (44). Many of the current applications of SBME in GME are targeted at procedural skills like complex surgical techniques, bridging the gap for trainees’ experiential learning on invasive, uncommon, or high-acuity procedures (45). The integration of artificial intelligence into clinical simulations would theoretically allow for the customization of scenarios based on a trainee’s skill level and performance data, providing a personalized learning experience and potentially opening the door to new types of patient simulation (43). Accordingly, there has been interest in using conversational GenAI to simulate patient encounters to practice cognitive and communication skills, though this application is more often focused on undergraduate medical education (15, 33, 46–49).

Among the most interesting potential applications of GenAI in GME is the concept of using synthetic data as training material for visual diagnosis. GANs and diffusion models have shown promise in generating realistic images of pathology findings (50, 51), skin lesions (52–54), chest X-rays (55), genetic syndromes (56), and ophthalmological conditions (57). The synthetic data approach may ultimately address important limitations in image-based training data sets, such as underrepresentation of certain patient demographics and adequate demonstration of rare findings.

Individualized tutoring produces better academic outcomes than learning in a traditional classroom setting (58). Skilled teachers can guide learners at different levels through complex topics, offering tailored and accessible explanations. One-on-one tutoring delivered by humans is costly, and skilled teachers are not available everywhere, but GenAI tools may have some of the same benefits at a fraction of the cost. LLMs show promise as a tool for explaining challenging concepts to graduate medical trainees in a manner tailored to the learner’s level (18), and LLMs could be configured to act as personalized tutors (59). In one study, an LLM successfully reviewed trainee-generated radiology reports and generated relevant educational feedback, a concept which could be extended to other types of clinical documentation (60).

GME trainees preparing for board examinations often use question banks to study, and GME programs use board-exam style questions to assess trainee progress. Question generation can be a costly and labor-intensive proposition (61). Authors report mixed success with using LLMs to generate board exam-style questions (61, 62), but as the technology matures, it seems likely that LLMs will be used by trainees and educators alike to create high-quality self-directed study materials and test questions.

LLMs are powerful tools for academic research and writing, and can assist in idea generation, processing complex background information, and proposing testable hypotheses (63, 64). LLMs readily summarize complex academic papers and draft academic text, abilities that can accelerate academic productivity (65). When paired with reliable academic databases and search engines and/or when fine-tuned with specific knowledge, LLMs do a serviceable job of conducting literature reviews (66, 67), synthesizing findings from existing literature, and drafting new scientific text with accurate literature citations (68). LLMs have great utility in assisting non-native English speakers with academic writing, representing a cost-effective and always-available alternative to commercial editing and proofreading services or to searching for native English-speaking collaborators (69).

Among the competencies listed in the ACGME’s Common Program Requirements is the ability to “systematically analyze practice using quality improvement (QI) methods” (20). GME trainees are required to participate in QI projects, which are typically require quantitative data analysis. Trainees are often underrepresented in organizational quality improvement activities, with one potential reason being the substantial time and effort needed for data collection and analysis (70). LLMs have some ability to facilitate straightforward data analysis and can generate serviceable code for statistical and programming tasks (71). LLMs are also adept at natural language processing tasks like extracting structured data from unstructured medical text (72).

Computer-based clinical decision support (CDS) systems are among the most effective tools for guiding good clinical decision-making (73). For GME trainees, CDS that provides authoritative, evidence-based guidance has both great practical clinical and educational utility (73). CDS that delivers evidence-based clinical guidance based on relevant patient data is a required feature for EHR systems certified by the United States government (74). A widely accepted CDS framework explains that CDS should be delivered according to the “five rights”: the right information, to the right person, in the right format, through the right channel, at the right time (75). Most current CDS consists of rule-based expert systems that display alerts to providers. While such systems are effective, rule-based alerts often suffer from practical problems such as a lack of specificity, poor timing, and incomplete characterization of clinical context (76).

The potential for intelligent, interactive, authoritative, LLMs to serve as always-available clinical consultants and educators has generated compelling speculation (77). LLMs can provide context-sensitive and specific guidance incorporating clinical context and patient data, they can be accessed through readily available communication channels, and--in contrast to rule-based alerts--they are interactive. However, studies done to evaluate the potential of LLMs for clinical decision support in various clinical contexts (78–83) have shown mixed results so far, with limitations in their ability to handle nuanced judgment and highly specialized decision-making. Thus, while GenAI for CDS is an area of great potential and ways to improve performance are under development, GME faculty and trainees cannot yet rely on LLMs to directly guide clinical care.

In essence, LLMs are statistical models that predict the most likely continuation of a given input sequence, based on their training data. Sometimes this approach results in plausible sounding but factually incorrect outputs, a phenomenon called “hallucination.” This problem can be especially difficult when dealing with topics requiring nuanced understanding of context or specialized knowledge, conditions very common in healthcare and specialized academic settings. For example, a biomedical researcher recently reported a cautionary tale in which ChatGPT generated incorrect information about tick ecology, complete with an entirely fake but plausible-appearing source document citation (87). In clinical settings, LLMs have been found to occasionally add fabricated information to clinical documentation (88) and to provide incorrect clinical recommendations (89).

In GME, trainees learn in a real clinical environment where accuracy and context are critical. There is a risk that overreliance on LLMs can result in an incomplete or incorrect understanding of complex topics, contributing to a poor educational outcomes, loss of critical thinking skills, and/or to suboptimal care and patient harm (15, 90, 91). Techniques like retrieval augmented generation, fine-tuning and prompt engineering show promise in reducing or eliminating the problems of inaccuracy and hallucination (92–94), but at present, reliance on GenAI as a source of factual information in any important clinical or academic context is risky. In our view, assertions made by GenAI should be validated by the user to avoid misinformation, and GME trainees should not use GenAI to directly guide patient care decisions outside of a controlled research context. GenAI users should be aware of automation bias, a cognitive bias in which people tend to excessively trust automated systems (95).

In reviewing applications for GME positions, personal statements are one of the most important elements that program directors review (96), especially in modern era where in-person residency and fellowship interviews are less common. Personal statements allow program directors to assess an applicant’s interest in their program and the clarity, organization and effectiveness of their written communication (97). There have long been concerns about plagiarism in personal statements (98, 99), and these concerns are magnified by GenAI tools that can readily produce writing that is clear, well-structured and compelling but that lacks an applicant’s unique voice, style and values (100, 101). Similarly, through letters of recommendation (LORs), faculty advocate for applicants by highlighting qualities observed in longitudinal relationships; using GenAI to draft LORs may have benefits but raises similar concerns about authenticity (102). GenAI-written content can be difficult to detect, even with software assistance (103). Some authors recommend that program draft policies for the use of GenAI in personal statements and LORs, with a common recommendation being that the use of GenAI be disclosed by the writer (97, 101).

As noted above, GME trainees are also expected to participate in research, academic writing, quality improvement summaries, creation of educational curricula, and similar academic activities. There are currently no consensus standards for using GenAI in academic medicine, but a recent review synthesized existing papers into a proposed set of guidelines, paraphrased as: (1) LLMs should not be cited as coauthors in academic works, (2) LLMs should not be used to produce the entirety of manuscript text, (3) authors should have an understanding of how LLMs work, (4) humans are accountable for content created by the LLM, and (5) use of an LLM should be clearly acknowledged in any resulting manuscripts (104).

GenAI tools are typically trained on huge corpora of data from the internet such as informational web sites, public forums, books, research literature, and other digitized media. Given the uncontrolled nature of the training data, it is unsurprising that they can exhibit social bias and stereotypes in their output (105). If unmanaged, these biases have the potential to reinforce detrimental beliefs and behaviors (106). In healthcare, GenAI may overrepresent, underrepresent or mis-characterize certain groups of people or certain medical conditions (18).

GenAI is computationally intensive and expensive to operate. Thus, many resource-limited healthcare organizations or individual physicians may rely on third-party, external GenAI tools (107). Given the great utility of GenAI, knowledge workers may be sorely tempted to upload confidential information, despite significant risks (108). In healthcare, such risks are legal as well as ethical in nature, and transgressions can have implications for professional development.