A groundbreaking Harvard study has found that artificial intelligence outperformed human doctors in emergency room triage, diagnosing patients more accurately during the high-pressure moments when people first arrive at the hospital. Researchers tested OpenAI's o1 reasoning model, a type of large language model (LLM) that uses advanced reasoning to process information, against hundreds of physicians in real clinical scenarios. The results suggest AI could reshape how emergency medicine works, though experts emphasize the technology is meant to assist doctors, not replace them.

How Did Researchers Test AI Against Emergency Doctors?

The Harvard team conducted a rigorous trial at Boston's Beth Israel Deaconess Medical Center, comparing AI performance directly to human physicians in several diagnostic scenarios. In one key experiment, 76 patients who arrived at the emergency room were evaluated by both an AI system and pairs of human doctors. All were given the same standard electronic health record, which typically includes vital signs, patient demographics, and a brief note from a nurse explaining why the patient came to the hospital.

The results showed a clear advantage for the AI in fast-paced triage situations. When working with minimal information, the AI identified the exact diagnosis or a very close match in 67% of cases, compared with 50% to 55% accuracy for the human doctors. The AI's accuracy improved to 82% when more detailed patient information was available, though human doctors also improved to 70% to 79% accuracy at that level. The AI also excelled at longer-term planning tasks, scoring 89% when asked to develop treatment plans like antibiotic regimens or end-of-life care strategies, compared with just 34% for human doctors using standard resources like search engines.

Why Does AI Perform Better in Emergency Situations?

The AI's advantage appears strongest in situations that require rapid decisions with incomplete information, which is exactly what emergency medicine demands. One striking example from the study illustrates this strength: a patient presented with a blood clot in the lungs and worsening symptoms. Human doctors assumed the anticoagulant medications were failing, but the AI noticed something the physicians missed. The AI reviewed the patient's medical history, identified a diagnosis of lupus, and correctly reasoned that lupus was causing the lung inflammation rather than medication failure. The AI's diagnosis proved correct.

This type of pattern recognition across large amounts of medical data is where large language models excel. These AI systems are trained on vast amounts of text, including medical literature and clinical records, allowing them to consider a wider range of possible diagnoses than a human doctor might think of under time pressure. However, researchers stress that the AI was essentially performing like a clinician writing a second opinion based on paperwork, not evaluating the patient in person.

What Are the Key Limitations and Concerns?

Despite the promising results, experts and researchers identified several important limitations that prevent AI from being ready for independent clinical use. The study only tested AI on text-based patient data; the system never evaluated a patient's physical appearance, level of distress, or other visual cues that human doctors use constantly in emergency settings. This is a significant gap, since a patient's demeanor, skin color, and physical presentation often provide crucial diagnostic information.

Additional concerns about AI in emergency medicine include:

• Accountability Gaps: There is currently no formal legal or regulatory framework for determining who is responsible when an AI system makes a diagnostic error that harms a patient.
• Unconscious Deference: Research suggests doctors may unconsciously defer to the AI's answer rather than thinking independently, potentially reducing the quality of clinical judgment over time.
• Demographic Performance Gaps: The study did not provide detailed information about which types of patients the AI performed worse on, or whether it struggled more with elderly patients or non-English speakers.
• Safety for Routine Use: Experts cautioned that the study does not demonstrate the AI is safe for routine clinical use or that the public should use freely available AI tools as substitutes for medical advice.

Dr. Wei Xing, an assistant professor at the University of Sheffield's School of Mathematical and Physical Sciences, emphasized these concerns: "This tendency could grow more significant as AI becomes more routinely used in clinical settings," he noted, referring to the risk that doctors might over-rely on AI recommendations without independent verification.

How Could AI Worsen Health Disparities?

While the Harvard study focused on diagnostic accuracy, a parallel concern is emerging about whether AI healthcare tools could amplify existing inequalities in medical care. A 2024 systematic review analyzing 30 studies over a decade found a significant association between AI use in healthcare and worsening of racial and ethnic health disparities.

The research documented several troubling patterns. One machine learning algorithm used for scheduling patient appointments led to Black patients experiencing 33% longer wait times than other patients, because the model relied on socioeconomic indicators like employment status, zip code, and insurance type, which are correlated with race. Another widely used algorithm assigned Black patients the same risk level as White patients even though Black patients were sicker, because the algorithm used healthcare costs as a proxy for illness severity, and less money is typically spent on Black patients due to existing inequities in care access.

In diagnostic AI, models often underperform on patients with darker skin because training datasets contain more data from lighter-skinned patients. Language-based AI models also showed worse performance predicting depression severity for Black patients compared to White patients, since the two groups use different language patterns to describe depression symptoms, and most AI training data comes from White populations.

What Do Experts Say About AI's Future Role in Medicine?

Rather than replacing physicians, researchers envision a collaborative future. Dr. Arjun Manrai, one of the lead authors of the Harvard study and head of an AI lab at Harvard Medical School, stated: "I don't think our findings mean that AI replaces doctors. I think it does mean that we're witnessing a really profound change in technology that will reshape medicine".

"I don't think our findings mean that AI replaces doctors. I think it does mean that we're witnessing a really profound change in technology that will reshape medicine," said Arjun Manrai.

Arjun Manrai, Lead Author and AI Lab Director at Harvard Medical School

Dr. Adam Rodman, another lead author and a physician at Beth Israel Deaconess Medical Center, described large language models as "among the most impactful technologies in decades." He proposed a new model for clinical care over the next decade, calling it a "triadic care model" that would include the doctor, the patient, and an artificial intelligence system working together.

"Over the next decade, AI would not replace physicians but join them in a new triadic care model: the doctor, the patient, and an artificial intelligence system," explained Dr. Adam Rodman.

Dr. Adam Rodman, Physician at Beth Israel Deaconess Medical Center

Prof. Ewen Harrison, co-director of the University of Edinburgh's Centre for Medical Informatics, called the study important and noted that "these systems are no longer just passing medical exams or solving artificial test cases. They are starting to look like useful second-opinion tools for clinicians, particularly when it is important to consider a wider range of possible diagnoses and avoid missing something important".

How Widely Is AI Already Being Used in Healthcare?

The Harvard study arrives as AI adoption in healthcare is already accelerating. Nearly one in five US physicians, or approximately 20%, are already using AI to assist with diagnosis, according to research published in 2026. In the United Kingdom, the adoption rate is even higher: 16% of doctors use AI daily and another 15% use it weekly, with clinical decision-making being one of the most common applications.

Beyond diagnosis, AI is being deployed across healthcare systems for administrative and operational tasks. A 2025 survey found that eight in ten health insurers, or 84%, report using AI or machine learning for fraud detection, utilization management, and prior authorization decisions. Health systems are also using AI to limit claim denials and streamline prior authorization processes.

The public is also turning to AI for health advice at scale. More than 40 million people globally use ChatGPT daily for health information, according to OpenAI data from 2026. AI chatbots are becoming a significant source of information about health insurance and billing, with users asking between 1.6 and 1.9 million questions per week about plan comparisons, claims, billing, and coverage.

However, public trust in AI for health information remains mixed. A 2026 KFF survey found that about one-third of adults, or 32%, use AI chatbots for health information or advice. But two-thirds of adults, or 67%, say they trust AI tools "not too much" or "not at all" to provide reliable health information. Confidence is even lower for mental health advice, with 77% of adults expressing low trust in AI for mental health and emotional well-being information.

The Harvard study and growing body of research suggest that AI's role in medicine is evolving from experimental technology to practical clinical tool. Yet the path forward requires careful attention to safety, equity, and the preservation of human judgment in life-and-death decisions. As more hospitals and doctors adopt these systems, the questions of accountability, bias, and appropriate use will become increasingly urgent.