Residents, Faculty, and Artificial Intelligence: Brave New World or Utopia?

Since the late 2022 release of OpenAI’s ChatGPT, an open-source natural language processing tool that can generate human-like intellectual content and responses, journal editors, including those at the Journal of Graduate Medical Education (JGME), have intensified debates around fair, appropriate use of this technology for submitted manuscripts.^1-4 We have recently expanded JGME’s overall policy of transparency to include artificial intelligence (AI): “Authors must communicate any use of artificial intelligence technology and similar tools, whether for writing assistance, storage, interpretation of qualitative research data, statistical analysis of numerical data, creation of visual imagery, or other uses.”⁵ In addition, listed JGME authors must be human. Why? Only humans can be accountable for all aspects of the work, a critical authorship criterion.⁶ Beyond these considerations of authorship and transparency, we are curious about potential AI uses in graduate medical education (GME) for both the process and content of training.^7-9

In 2020, we asked 40 thought leaders in medical education what they believed would be the most significant transformation in GME over the next decade: AI was a major theme.¹⁰ Three years later, AI use in daily life is everywhere, from cell phones and credit checks to weather and climate predictions. AI examples include natural language processing, speech recognition, and creating written works (See if you can tell the difference between human- and AI-generated text in Box 1). The volume of health information has increased beyond the capacity of humans to sift and digest: in 1950 medical information doubled in 50 years; by 2010, it took 3.5 years, and the estimate for 2020 was 73 days.¹¹ However, only a few undergraduate medical schools have added AI experiences, and the Accreditation Council for Graduate Medical Education Milestones do not include AI competencies.¹¹ Currently AI is being introduced into learning and clinical diagnosis support, automated assessments of clinical skills, setting educational goals, and designing curriculum and performance assessments.^12,13

Scholars have hailed the potential benefits of AI, such as improving manuscripts for authors writing in a second language, saving researchers time when creating research documents or conducting literature searches, and freeing trainees from memorization to focus on reasoning, counseling, and shared decision-making with patients.^11,14 Other scholars are concerned that AI tools in GME present serious risks. These include loss of trainee and patient anonymity in confidential information, failure to develop key competencies as those tasks are outsourced to AI tools, erroneous evidence and assessment summaries, and “paper mills,” or fabrication of fake research to pad resumes and grant applications.^15-18 As AI is wholly dependent upon the quality and included biases of the available training data, it may generate inaccurate syntheses, which could be missed by inexperienced trainees and beleaguered faculty. In addition, AI could negatively affect trainees’ ability to develop mental models or learn key foundational skills, such as clinical reasoning in treating complex patients. Many patients, with their unique values, social supports, finances, and medical histories, may be hard to fit into AI’s derived diagnoses and management plans. In 2023, the optimal interactions between physicians and AI, patients and AI, and trainees and AI are unknown, yet graduating residents and fellows will increasingly manage these interactions and confront numerous ethical sequelae.

Medical education is notorious for diving headlong into the next innovation, often with scanty evidence. But as Dr Rachel Ellaway points out, “the genie is out of the bottle.”³ We need research that examines the best uses of AI in GME. This research should include exploration of both the use of AI during training and a greater understanding of the ethical, beneficial uses of AI in medical education research. Which research tasks remain uniquely human? Which AI educational functions require minimal human oversight? Will AI replace some physician skills, rendering them obsolete and no longer taught in GME? Will AI technology transform cognitive skills, similar to robotics in surgery? We invite reports of how GME trainees, programs, faculty, and institutions are learning about and using AI tools (Box 2).

Of substantial concern is whether for-profit entities, which have contributed to the growing disparities in US health care, will encourage problematic AI use in medical education and patient care. We prefer that ethics and evidence lead the way. The US health care system is increasingly a “flailing medical industry shaped by profit rather than ethics.”^19,20 Thus we need AI tools that can enhance health equity. AI is hyped as a way to generate faster, more accurate diagnoses; reduce physician error; reduce or eliminate repetitive, mundane physician tasks such as the electronic health record (EHR); and reduce the high cost of health care.¹¹ Some of us are wary of these claims. They are reminiscent of those touted for EHRs, now shown to produce increased physician documentation time and burnout, especially for the endangered species known as primary care physicians.²¹

What do most GME educators want from AI? Our guess is they desire more time for teaching and tools that will help more learners reach “mastery” competency levels. They also want their trainees well prepared for future AI uses, in a medical utopia rather than a dystopian Brave New World.²²