Invited Commentary
April 28, 2023
How Chatbots and Large Language Model Artificial Intelligence Systems Will Reshape Modern Medicine: Fountain of Creativity or Pandora’s Box?
Ron Li, Andre Kumar, Jonathan H. Chen
JAMA Intern Med. Published online April 28, 2023. doi:10.1001/jamainternmed.2023.1835
In an era of clinicians being burnt out by electronic medical records and documentation burdens, we might all dream of having a personal scribe to draft progress notes, translate patient instructions, summarize the literature, complete insurance authorization paperwork, and respond to unending in-basket messages, as described in the Perspective in this issue of JAMA Internal Medicine.1 This would have sounded like a fantasy just a few years ago, but the release of rapidly developing chatbots now demonstrates the potential of large language model artificial intelligence (AI) systems with surprisingly adept language manipulation and knowledge processing capabilities. The underlying foundation model technology rides atop the peak of inflated expectations,2 reflecting a disruptive technology likely to change the way we work and live, even as we must be aware of substantial limitations. Good or bad, ready or not, Pandora’s box has already been opened. One such large language model, ChatGPT, is the fastest-growing internet application in history with more than 100 million users.3 This has shifted access to sophisticated AI capabilities away from concentrated pockets of technical experts to the masses, where all types of otherwise unimaginable (and unintended) use cases are being discovered. To ensure that the adoption of such tools into health care practice is done effectively and responsibly, physicians must lean in to understand and drive this conversation.
Large language models represent the underlying class of machine learning models trained in autocomplete tasks. Given the words “coronary artery,” these models may predict the next word to be “disease,” “bypass graft,” or “calcification” based on statistical parameters learned from prior training data text on how often those words appear together. These models have been growing increasingly larger, learning billions of parameters from many billions more books, articles, and conversations across the internet. This scale and fine-tuning by human examples4 have enabled the relatively simple autocomplete concept to exhibit surprising emergent properties of complex language capabilities including summarization, translation, and question answering, even without specific training for such tasks the way most other narrow AI systems work.5 Especially striking for the medical community is that these systems can now perform at a level that passes the US Medical Licensing Examination,6 while generating responses to patient questions posted on a social media forum with higher quality and empathy than responses by human physicians, as demonstrated in a cross-sectional study in this issue of JAMA Internal Medicine.7
The combination of these large language models with a familiar chat interface enables humans and AI systems to engage in a dynamic dialogue through the high-bandwidth yet relatable medium of human language. Given how language deeply affects how we think, behave, and communicate, this arguably makes these systems more dangerous when they are inaccurate or biased. Language models are prone to confabulation, assembling coherent strings of words into sentences that sound believable while being completely fabricated. Imagine a trainee who tells you when they are unsure vs another who confidently bluffs their way through rounds with made-up information—which one is a bigger liability for patient care? With both the capabilities and limitations of such systems in mind, we consider 3 levels of health care applications for large language model systems with increasing potential for disruption (and uncertainty).
Simplify (If Not Replace) Tasks Involving Text Analysis, Synthesis, and Generation
In an era when physicians regularly spend more time on the electronic medical record than with patients, language models could assist with clerical documentation activities, such as drafting notes and administrative letters, as well as perform the laborious “chart biopsy” tasks to create succinct summaries from dense patient medical records. Applied to medical information at large, these tools can analyze, synthesize, and summarize all of the published literature, textbooks, and internet content into an understandable and usable format. The risk of course is that this could just as easily propagate false, biased, or otherwise flawed information from such sources without regard for accuracy.
Enable New Workflows and Models of Care Delivery
Just as companies use AI chatbots for customer service, health systems may begin to use language models to facilitate patient communication. Language model-enabled patient portals may become the “front door” for health system information, relieving bottlenecks created by staffing call centers, in-basket pools, and overwhelmed clinics. For communication requiring clinician input, language models can draft responses that are selectively edited by the clinicians.
Language models will further empower patients to access and understand their own health information. Although clinicians must not directly enter individual protected health information into a third-party system due to patient privacy, patients can already easily copy and paste their own jargon-laden medical records into a chatbot and tell it to “explain this to me in plain English” (or any other language).
Patients are inevitably going to bypass health care systems and directly consult language models for medical advice, much as they already do with internet searches. While there must always be critical considerations around accuracy and safety, we must imagine new ways to support patients, as many will gladly reach out to an imperfect AI chatbot when it is available 24/7 for immediate response in a way that a constrained health care workforce cannot be.
Shift the Boundaries Between Human Expertise vs Artificial Intelligence Expertise
Language models are remarkably capable of answering complex medical examination questions with accuracy, concordance, and even insight beyond the question stem.6 While current systems are minimally passing medical knowledge examinations, the pace of progress indicates that we should be planning for future iterations that will likely surpass the test-taking skills of the average practicing physician in the foreseeable future.8
The (re)training of a generation of physicians must adapt to this new era where we can expect language models to be integrated into clinical decision-making, just as we expect any modern physician to know how to use online resources to support their care plans. Medical schools can use language models to aid trainees in honing clinical reasoning through the generation of illness script knowledge structures (eg, “generate a typical case presentation of pneumonia”), summarizing literature evidence, providing succinct yet insightful explanations for complex biological processes, and practicing effective (electronic) communication with patients.
The American Medical Association created a curriculum in 2018 surrounding AI in health care to aid physician understanding of how machine learning can affect their workflows.9 As the role of AI systems continues to expand in medicine, these curricula will need continuous revisions with competency tracked as important credentialing milestones. The emerging generation of physicians must be versed in the capabilities, limitations, and implications of these technologies to lead their effective use in health care while mitigating the dangers of inaccurate and biased results that will arise from both blind acceptance and rejection.
We are entering a new era amidst an abundance of information but a scarcity of time and human connection. The practice of medicine is much more than just processing information and associating words with concepts; it is ascribing meaning to those concepts while connecting with patients as a trusted partner to build healthier lives. We can hope that emerging AI systems may help tame laborious tasks that overwhelm modern medicine and empower physicians to return our focus to treating human patients.
