Loading ...

Artificial intelligence tools are beginning to move from research experiments into real-world emergency department workflows, raising new questions about how algorithms should be used in high-stakes clinical decisions.

In a prospective feasibility study published January 2026 in the peer-reviewed journal Discover Artificial Intelligence, researchers compared ChatGPT-4 with emergency physicians evaluating walk-in emergency department patients. The study evaluated the model’s ability to generate histories, differential diagnoses, and management plans in real emergency department presentations. 

Researchers found that the AI system captured additional elements of medical history in 21% of cases and showed moderate agreement with physicians’ diagnoses, although it tended to recommend more diagnostic testing and hospital admissions than clinicians. 

Another study published May 2024 in The American Journal of Emergency Medicine compared GPT-3.5 and GPT-4 with experienced emergency physicians calculating common emergency department clinical scores such as the NIH Stroke Scale and HEART score. While GPT-4 showed moderate agreement with clinicians, physicians still achieved stronger overall predictive performance.  

The growing presence of AI in emergency medicine has even begun to appear in popular culture. In the current season of The Pitt, the fictional team at Pittsburgh Trauma Medical Center wrestles with an algorithmic helper that proves both useful and problematic. In one scene, the AI helps a physician cut through a backlog of chart documentation. In another, its ambient scribe function introduces a medication error in the record — underscoring the importance of careful clinician oversight when using automated tools.

That tension mirrors what clinicians are beginning to see in real emergency departments, said Shawn Griffin, MD, president and CEO of URAC, an independent, nonprofit healthcare accreditation organization. Griffin said his perspective is informed by his 8 years as the chief quality and informatics officer for MHMD Memorial Hermann Physician Network, where he served as the CMS liaison and quality contact for the Memorial Hermann Accountable Care Organization.

“The most common use of AI in healthcare today is trying to fix the last big implementation of technology in healthcare, which is the EMR and how we all got turned into sort of caring for keyboards, not caring for people,” Griffin said.

Where AI Is Showing Up in Emergency Care

AI tools are already embedded in several aspects of emergency medicine, though most are focused on specific clinical tasks.

“AI tools are already embedded in many parts of emergency care, including imaging and ECG interpretation support, predictive models that flag patients at risk of deterioration or sepsis, AI scribes for documentation, and increasingly generative AI systems clinicians use as information resources,” said Katherine Eisenberg, MD, PhD, senior medical director at Dyna AI.

Individual clinicians may adopt AI tools to support decision-making or save time, while health systems often focus on operational challenges such as patient throughput, triage prioritization, and clinician workload.

“In practice, AI tool adoption in the ED is often fragmented, similar to many other parts of our health system,” Eisenberg said.

Other experts say the most practical uses today involve narrow tasks where rapid analysis can improve triage and prioritization.

“AI is already being used in emergency departments, but mostly in narrow, high-value tasks where speed matters,” said Roger Boodoo, MD, medical director of AI at HOPPR, an AI company involved in medical imaging. Boodoo is a radiologist and a 24-year Navy veteran whose perspective on healthcare AI is shaped by experience in frontline combat medicine.

Those applications include imaging triage, sepsis alerts, and risk-stratification tools designed to help clinicians identify urgent cases sooner.

“The goal is not to replace physician judgment. It is to improve prioritization,” Boodoo said. “In a crowded ED, the real value of AI is helping surface the right case at the right moment.”

Some of the clearest examples involve time-sensitive conditions such as stroke.

“One of the most challenging decisions in an emergency department — especially in rural communities — is whether a patient needs transfer to a higher level of care,” said Andrew Ibrahim, MD, a practicing surgeon, associate professor of surgery at the University of Michigan, and chief clinical officer of Viz.ai in Ann Arbor, Michigan. 

AI-enabled stroke pathways can help accelerate those decisions by rapidly identifying possible large vessel occlusions on CT scans and immediately sending those images to stroke specialists.

“Within minutes, right from their phone, imaging can be reviewed and coordination can begin to transfer the patient,” Ibrahim said.

Adoption of these systems has been associated with faster treatment decisions and shorter hospital stays, he said.

Evidence So Far: Workflow Gains More Than Outcome Data

Experts say the strongest evidence supporting AI in emergency medicine so far relates to workflow improvements rather than broad clinical outcomes.

Ibrahim pointed to several studies evaluating AI-enabled stroke triage systems. A multicenter stepped-wedge randomized trial examining automated large vessel occlusion detection software found that AI-enabled stroke triage significantly reduced time to thrombectomy by accelerating stroke team notification and care coordination.

Other real-world studies evaluating AI-enabled stroke networks have also reported faster identification of large vessel occlusions and improved coordination across hospital systems.

Still, Eisenberg said the evidence base varies depending on the type of AI system being studied.

“It’s important to distinguish between predictive AI tools and generative AI tools,” she said.

Predictive models — such as sepsis alerts or deterioration-risk tools — have been in use longer and therefore have a larger body of supporting evidence.

“Even with predictive tools, demonstrating improvements in patient outcomes is still difficult,” Eisenberg said.

For generative AI systems, the evidence base is still emerging.

“A tool can perform well on a benchmark and still not improve patient outcomes,” she said.

Why Clinicians Remain Cautious

Despite promising early results in areas such as stroke triage and risk prediction, many clinicians say the rapid expansion of AI tools in emergency medicine raises important questions about reliability, oversight, and how these systems should be integrated into clinical decision-making.

Griffin noted that a basic rule of informatics still applies: “garbage in, garbage out.”

“Most medical data is not really clean, pure data,” he said.

“Another concern is overreliance on algorithmic recommendations without understanding how the system arrived at its conclusions,” said Carlo Hallak, MD, physician informatics executive at San Juan Regional Medical Center in Farmington, New Mexico.

AI systems may miss contextual information that clinicians detect during patient encounters. Hallak said these systems can miss context that experienced clinicians pick up at the bedside. 

Alert fatigue is another concern.

“If tools generate too many signals that aren’t clinically meaningful, clinicians start tuning them out,” Hallak said.

Eisenberg said that automation bias, the tendency to rely too heavily on automated recommendations, is also an issue. The FDA’s January 2026 clinical decision support software guidance specifically called out this risk for tools used in acute or time-critical care settings. For generative AI systems, additional concerns include hallucinations and a lack of contextual understanding.

Ibrahim said clinicians should approach AI tools with the same caution they apply to other medical technologies.

“As a surgeon, I do not perform an operation unless I understand how it works and where it can go wrong,” he said. “We should have a similar approach to AI tools.”

No reported disclosures.