The waiting room is full. Patients line the walls, some clutching ice packs while others sit quietly with symptoms that are not immediately visible. A middle-aged man experiencing chest discomfort waits beside a young woman with a high fever and an elderly patient struggling to catch his breath. Nurses move quickly between patients, balancing limited time and resources while attempting to determine who requires immediate attention and who can safely wait. As more patients arrive, wait times lengthen, information becomes harder to manage, and the risk of delays in identifying serious conditions increases. 

     Scenes like this are common in emergency departments (EDs) across the United States. Emergency room overcrowding and increasing patient volumes have placed significant strain on traditional triage systems, which rely heavily on human judgement under immense pressure. While healthcare professionals like triage nurses are highly trained to execute these decisions, the complexity and pace of modern emergency medicine creates opportunities for delays, inconsistencies, and preventable errors. As hospitals search for ways to improve triage efficiency and ultimately patient safety, artificial intelligence (AI) has emerged as a potential tool to assist the triage process. Supporters argue that AI could help identify high-risk patients more quickly and streamline patient flow, while critics raise concerns about reliability, bias, and privacy in clinical decision-making. As healthcare systems continue to evolve, the question remains whether AI can meaningfully improve ED triage while avoiding unintended consequences. 

 

Introduction to Artificial Intelligence

 

     One form of artificial intelligence increasingly being explored in emergency triage is machine learning (ML), a type of AI that can analyze large amounts of patient data and identify patterns that may not be immediately apparent to healthcare providers. ML algorithms can evaluate information such as a patient’s age, symptoms, vital signs, medical history, and laboratory results to help predict the severity of illness and prioritize patients based on urgency. More advanced systems also incorporate natural language processing (NLP) within generative AI, which allows computers to interpret and analyze information documented in clinical notes or patient descriptions of symptoms, and then provide a new, written response [1]. For example, ChatGPT is a large language model (LLM), which describes an AI tool that is capable of analyzing human language. It utilizes NLP technology, as it has the ability to generate new, written responses through predictive ML pathways. It is important to note that LLMs that are being considered to be used in ER triage are not actually “seeing” patients like a clinician would, but rather are working based on data and observations collected by health professionals. Recognizing signs and symptoms that could be indicative of a specific disorder to prompt an LLM is still a skill that would have to be possessed by clinicians.

 

Increased efficiency

 

     One of the most consistently cited benefits of the use of AI in triage is improved efficiency in high-volume clinical settings. Emergency department triage nurses must obtain a large amount of information in a short period of time while simultaneously managing multiple patients. AI-powered systems, particularly those using natural language processing, may assist in gathering and organizing patient information in a structured and efficient format for healthcare providers, while eliminating some human error. This could allow for more rapid medical intervention and efficient emergency operations. Even small reductions in administrative or cognitive burden have been shown to translate into faster patient throughput and reduced wait times. A 2024 review utilizing NLP, machine learning, deep learning, and some hybrid AI models  reports that these forms of AI can assist in reducing the crowding of emergency departments, thus allowing patients to seek care sooner [2]. As Dr. Joshua Stilley states, “Using an AI tool to assist in taking a patient history in triage may help with obtaining the appropriate and pertinent information for more appropriate evaluation and triage prioritizing, while at the same time improving the workload and flow of the triage staff.” Large language models utilizing NLP also has the potential to present a comprehensive history of present illness and review of symptoms before physician evaluation, which may improve efficiency in emergency departments. Dr. Stilley writes, “[AI] may have [a] downstream benefit for physicians if the AI is able to obtain a full HPI (history of present illness) and ROS (review of systems) before the physician encounters the patient, thereby improving overall ED and physician-specific timeliness as well.”

 

Clinical reasoning support

 

     AI systems also have the potential to support more consistent initial patient categorization in triage. Traditional triage relies heavily on rapid clinical judgment, which can vary between providers depending on experience level, fatigue, and patient volume. AI-driven algorithms, particularly LLMs with the ability to understand complex language, by contrast, apply standardized criteria across cases, which may help reduce variability in how patients are prioritized [3]. This consistency could improve fairness in triage decisions, particularly in busy or resource-limited settings where subjective decision-making may unintentionally influence outcomes.

     Another potential advantage is early identification of high-risk conditions, such as sepsis warnings, to prompt earlier evaluation and intervention. Two 2024 systematic reviews of machine and deep learning AI applications in emergency and disaster triage found that intelligent triage systems improved patient care through real-time data analysis, continuous monitoring of vital signs, and more efficient patient prioritization. These systems were shown to support timely treatment decisions and improve resource allocation by identifying patients whose conditions required urgent intervention [4,5].  By serving as an additional layer of surveillance, AI with machine learning capabilities may help clinicians recognize deteriorating patients sooner, allowing for earlier evaluation and treatment of potentially life-threatening conditions, despite some current limitations [6].  Although these technologies are not intended to replace clinical judgment, their ability to process large amounts of data quickly may enhance the early detection of critical illnesses in busy emergency settings.

     However, LLMs are not always accurate in emergency triage situations. Dr. Cale Davis notes, “While these algorithms can be useful, they are often not very accurate. Oftentimes they are used as a way to be incredibly sensitive pulling in all possible Sepsis patients but have very low specificity, meaning most patients that are flagged will not actually have Sepsis. I anticipate that as AI is incorporated into medicine further, the accuracy will improve, but certainly needs some work at this point.” While these systems are often intentionally sensitive rather than specific, they can serve as a safety net by flagging patients who might otherwise be overlooked in a crowded emergency department. Additionally, one study published in February 2026 found that while ChatGPT was able to mostly accurately triage routine cases, it consistently over-triaged non-urgent cases while under-triaging life-threatening emergencies [8]. In practice, this would translate to not providing treatment for people who truly need it the most while bogging the healthcare system down even more than it already is by treating non-urgent patients with more urgency than is warranted. As AI models continue to evolve, there is optimism that they may improve upon these existing tools by better balancing sensitivity and specificity. However, until LLMs can be trained to triage accurately, they pose a risk for under- or over-triaging, both of which can be dangerous to patients and wasteful of precious medical resources.

 

AI and Bias

 

     Eliminating bias is always a priority in healthcare and bioethics to ensure fair and just allocation of resources to patients. If implemented in healthcare settings, AI has the potential to reduce human bias during clinical decision-making. Physicians, despite extensive training and expertise, can be influenced by implicit and other forms of bias that may affect patient assessment and treatment. AI systems, when trained with diverse and representative datasets, could evaluate patient information using more consistent criteria for every individual. This standardized approach may help minimize disparities in care by focusing on objective clinical data rather than subjective impressions. AI-assisted tools can identify patterns in symptoms, laboratory results, and medical histories without being influenced by a patient’s appearance, socioeconomic status, or personal characteristics. While AI itself is not immune to bias and can reflect biases in its training data, proponents argue that thoughtfully designed and monitored systems have the potential to provide more consistent and equitable healthcare decisions than those influenced by humans alone. For example, a 2024 Harvard review reported that black patients are half as likely to receive Opioid pain medications as white patients with a similar symptom profile, but an AI tool can assess pain objectively regardless of patient background [7].  Physicians are hopeful that the addition of AI can reduce patient disparities. Dr. Joshua Stilley notes, “[The] hope is that the AI would be able to be more objective than humans in regards to bias, but there can be barriers including language, ability and access to technical tools, and inadvertent statements by the patient that would bias the AI.”

     Other studies have shown that AI in ER triage could actually potentially introduce additional bias when diagnosing underrepresented patient populations. In a 2021 Nature Medicine study, Canadian researchers found that a diagnostic AI algorithm that used deep learning to assess chest radiographs was more likely to underdiagnose under-served patient populations, and was especially more likely to underdiagnose intersectional groups such as young female patients, young African American patients, and young patients with Medicaid insurance [9]. Additionally, even if an AI algorithm performs well on a data set similar to the one it was trained on, it may not perform as well when applied to a more diverse, realistic data set, highlighting the need for appropriate training for machine learning, deep learning, and LLMs [10]. AI has the potential to be biased in medicine due to inaccurate data, limited knowledge, and bias in the amount of data that is provided on certain conditions or patient populations [10]. Thus, to implement a non-biased AI system in healthcare, the AI must be trained with non-biased data. As stated by Dr. Cale Davis regarding AI use in ER triage, the “... outputs are only as good as the inputs. In other words, like all computer functions, data must be inputted correctly and in a way that the computer can analyze properly to create an accurate output.” A 2024 study done by researchers at the University of Michigan found that white patients were significantly more likely to have blood panel data than African American patients. This is certainly not the only example of a field of health data that contains insufficient data for minorities. The authors of this study argue that an imbalance in the amount or quality of data from under-served patient populations, such as the lack of blood draw data for African American patients in their study, may exacerbate existing biases in healthcare if data like these are applied to train an LLM for ER triage. Therefore, part of developing an accurate and non-biased LLM for ER triage is ensuring that accurate, non-biased, diverse information and patient data is used to train the LLM [12]. 

 

Patient privacy

 

     Patient privacy is also a major factor to consider regarding AI use in ER triage, as upholding privacy and confidentiality for patients plays a major role in upholding patient autonomy. Ensuring that a patient’s data remains confidential means they are in complete control of making decisions about their health and need not be swayed by the opinions or influence of others [13]. As LLM training and use requires accessing large amounts of patient data, a valid concern is that data may be able to be traced back to specific patients without the consent of those patients. HIPAA permits the disclosure of de-identified patient data as it is no longer considered HPI and therefore no longer falls under the umbrella of HIPPA [14]. While some types of data can be de-identified, others, like a picture of a patient’s face from a dermatology or eye clinic, cannot [15]. Additionally, even if data can be de-identified, the risk of still being able to trace it back to the patient is high through data leaks, matching of particular sequences of events and dates, or breaching the pseudo-encrypted deindividualized data [16]. Moreover, according to the National Institute of Standards and Technology, AI models specifically pose a risk for leaking patient data, as they “may leak, generate, or correctly infer sensitive information about individuals” [17].  

 

Heterogeneity in clinical benchmarks for AI

 

     Some studies on AI performance in ER triage may not provide the best answers regarding whether AI is safe to use in the ER setting. In April 2026, researchers at Harvard published a study in which one experiment compared the performance of two LLMs developed by OpenAI to the performance of two physicians attempting to provide diagnoses for 76 emergency room cases. The study found that when provided with information obtained during the “Initial ER triage” stage, one of the AI models performed slightly better than the two physicians and found no significant difference between the human physicians and AI models once the “Admission to Hospital or ICU” stage was reached. Yet, the researchers claimed in the abstract of this study that the “...study suggests that LLMs have eclipsed most benchmarks of clinical reasoning, motivating the urgent need for prospective trials” [18]. Most major news sites, including Harvard Magazine, The Guardian, and NPR, reported on this study with headlines such as “In real-world test, an AI model did better than doctors at diagnosing patients” [19]. Aside from the immediately apparent issue of having an extremely small sample size of physicians, these researchers also asked internal medicine physicians to perform ER reasoning tasks. Internal medicine physicians are focused on providing a long-term diagnosis and treatment plan and perform very different jobs than ER physicians, who are focused on delivering immediate life-saving care, as explained by ER physician Kristen Panthagani in her Substack article, Did AI really beat ER doctors at ER triage? [20]. 

     Dr. Panthagani voices concern regarding the “hype” of AI in medicine and ER triage, fearing that placing too much confidence in unproven LLMs may lead to further mistakes made down the road, endangering patients by putting them at risk of letting life-threatening diseases go untreated while also potentially discouraging them from seeing a physician at all. Dr. Panthangani is not alone in this concern [20]: the main issue many physicians have with AI use in ER triage and in medicine in general is the lack of testing behind it. 

     In another Substack article, physician scientist Eric Topol highlights a paradox emerging with regards to AI use in medicine and LLM use in medicine, specifically. He explains that AI in medicine for examining diagnostic images has been repeatedly and thoroughly tested over several years and has proved to be more effective than diagnosticians in recognizing some small nuances in diagnostic images. However, the medical field refuses to widely establish AI use for analyzing diagnostic images. At the same time, medicine has somewhat widely accepted relatively untested and unproven LLMs for use [21], even if not in an official capacity with between 32 and 73% of surveyed adults using AI chatbots for health purposes [22, 23] and over 80% of physicians using AI in a professional capacity [24]. 

     One reason so many may be willing to trust LLMs in making healthcare decisions is due to the humanness of engaging with a conversational LLM, as discussed in a 2026 study by researchers at Texas Tech University. This same study found that a tested LLM was ineffective and non-reliable at making triage decisions using triage-based logic in mass casualty events and that the LLM relied instead on textual cues (for example, prioritizing a screaming patient regardless of actual symptoms) [25]. 

     As can be seen, many articles publish different findings on the safety and efficacy of AI and LLM use in ER triage and in medicine in general. Our overall conclusion is that additional testing is needed before trusting LLMs with patient’s lives. Accordingly, an April 2026 Nature Medicine editorial highlights the need for providing standardized evidence for whether LLMs are safe and useful in clinical settings. Specifically, just proving that a medical innovation technically works, whether for AI or any other technology, is not equivalent to proving that it will be able to be feasibly used in a safe and effective manner in clinic [26]. If we ever want to consistently establish that AI use is safe and effective in healthcare, we must start enacting randomized trials that are consistent across the field, like we would with any other medical technology or drug, to assess the safety and efficacy of LLMs in making triage decisions. We must take steps to establish standardized benchmarks to prove the value and safety of AI to prioritize patient wellbeing [27].