Clinical care will most likely undergo a paradigm shift in the coming years. Patients may find themselves facing not a doctor, but a new kind of assistant: an  Artificial Intelligence (AI) system .  

The implementation of this technology will not only expedite diagnosis but will also monitor your vital signs in the waiting room itself, performing diagnostic tests, such as blood pressure measurement, continuously and autonomously. Simultaneously, an AI application, much like a human healthcare professional, could be “listening” to the conversation between you and your doctor to detect potential human errors or suggest the next clinical steps. This vision of AI-assisted healthcare is moving from science fiction to an imminent reality. 

A recent study published in the prestigious  scientific journal Science  reveals that a specific type of AI, the so-called “large-scale language models” (LLM), often already outperform flesh-and-blood doctors in diagnosing complex and potentially life-threatening conditions. 

AI passes the triage test  

The research team, which included experts from Beth Israel Deaconess Medical Center and Harvard University, tested the diagnostic accuracy of the o1 model developed by OpenAI. To do this, they used real cases of emergency room patients. 

Emergency departments are high-pressure environments where healthcare professionals must act quickly, often managing multiple cases simultaneously with fragmented information. In the initial stages of triage, which involves classifying patients according to the type and severity of their illness or injury to determine the order and location of their care, an error can have immediate and catastrophic consequences, such as misdiagnosing a serious bacterial infection with a common cold and sending the patient home without antibiotics. 

Faced with this complex scenario, the results of applying AI were surprising. With the initial information, it managed to identify an accurate or very close diagnosis in approximately 67% of cases, surpassing the 50% to 55% success rate achieved by human physicians.  

Even more striking was the performance on the “clinical reasoning” task: the AI ​​scored a perfect 98% in clearly explaining its diagnosis and the steps to follow, compared to a meager 35% achieved by the doctors involved. 

Technical limitations: the machine is not infallible  

Despite the enthusiasm of experts from institutions such as Google or the Dana-Farber Cancer Institute to see these tools in real-world environments, the system has clear limitations.  

Adam Rodman, an internist and co-author of the study, points out that emergency room stays are relatively short. If AI were to deal with hospitalized patients whose medical records accumulate “days and days of information,” its technical performance would likely decrease considerably. Furthermore, in this study, the model only analyzed written text, remaining “blind” to vital non-textual data such as X-rays or scans, which are essential for diagnosing conditions like cancer or blood clots. 

Other tools in the AI ​​environment are already trained in the recognition and interpretation of diagnostic images, so this limitation could be overcome with their use. 

Perhaps more important is a doctor’s ability to interpret signs not recorded in the medical history or diagnostic tests, which they would obtain from personal contact with the patient, listening to them, or interpreting their attitude, appearance, or expression—something that is difficult for an AI algorithm to achieve. 

Bioethical assessment 

The introduction of AI into clinical practice poses an anthropological challenge. First, medicine is not merely a discipline of statistical calculation; illness is experienced by a suffering individual. As we reflected in our article  “Can Human Suffering Be Measured? An Uncomfortable Question for Medicine ,” suffering is “one of the most complex and difficult realities to define in contemporary medical and bioethical practice.” An algorithm can predict the decrease in blood flow to the heart in milliseconds, but it is fundamentally incapable of understanding, measuring, or accompanying the patient’s suffering. 

Secondly, delegating the burden of diagnosis to technology compels us to reaffirm the irreplaceable value of human care. As Amparo Aygües highlights in her analysis for the Bioethics Observatory,  “I Will Not Die of Love: Responsibility for the Other and the Humanizing Power of Care,” healthcare possesses a profound ethical component: “care ennobles and humanizes both those who receive help and those who provide it.” If we allow AI to interfere and weaken the doctor-patient relationship, we risk losing that connection and our ethical responsibility toward “the Other.”  

We can conclude that Artificial Intelligence can be a formidable tool for reducing medical errors and streamlining diagnostic steps and guidance on therapeutic options, both in emergency medicine and in any other clinical setting. However, algorithmic precision must always be a complementary tool at the service of the professional, and never a substitute for their moral agency. The greatest challenge for the medicine of the future will be embracing technological innovation without abandoning its primary vocation: the humanizing power of care, which relies on listening, empathy, and intuition. These qualities arise from close relationships and, although they can be imitated by AI algorithms, will never be more than an imitation, far removed from the authenticity required in human relationships, especially in situations of vulnerability. 

Julio Tudela. Cristina Castillo. Bioethics Observatory. Catholic University of Valencia