AI is becoming increasingly adept at diagnosing diseases, with some studies suggesting that advanced systems can now identify patterns that may escape the clinician’s eye and support more accurate diagnostic and treatment decisions.

However, can AI meaningfully contribute to diagnosis in clinical practice today?

To explore this question, Speaking with El Médico Interactivo, part of the Medscape Professional Network,Ramón Puchades Rincón de Arellano, MD, coordinator of the Digital Medicine Working Group at the Spanish Society of Internal Medicine, stressed that findings from AI studies should be interpreted carefully.

“Many studies are conducted in controlled settings and compare very specific tasks. The fact that an AI system matches or outperforms a physician on a specific test does not mean that it can replace comprehensive clinical reasoning, which considers context, uncertainty, and patient preferences. Furthermore, data are context dependent and may vary according to the characteristics of the population in which AI tools are applied.”

Clinical Reasoning

One of the most widely discussed studies in this field, published in May 2026, compared AI with physicians across a range of clinical reasoning tasks. Researchers from Harvard Medical School and Beth Israel Deaconess Medical Center, Boston, evaluated an advanced large language model (LLM) from OpenAI’s o1 series directly against hundreds of physicians at different levels of training and experience on a variety of clinical cases, ranging from published patient vignettes to evaluations of brand-new emergency room patients, as well as on clinical tasks, including both diagnosis and planning of clinical management.

According to a report published in Science, researchers concluded that LLMs are rapidly approaching human level clinical reasoning and, in some areas, may already surpass it. They also suggested that AI-assisted decision-making, when combined with physician assessment, could help reduce diagnostic errors, delays in care, and disparities in healthcare access.

“We tested the AI model against virtually every benchmark, and it outperformed both previous models and our groups of doctors,” noted Arjun Manrai, PhD, co-author of the study and professor.

Interest in AI-supported diagnosis is not new. In 2024, researchers at Harvard Medical School developed a versatile AI model capable of performing multiple diagnostic tasks across different cancer types. By analyzing digital pathology slides, the system identified cancer cells and predicted tumor molecular profiles from cellular features visible in tissue images, achieving greater accuracy than many AI systems available at the time.

At the same time, other evidence has highlighted important limitations. An article published in JAMA Network emphasized that “despite advances, current machine learning models remain limited in early diagnostic reasoning and cannot yet be relied upon for unsupervised clinical decision-making.”

A Middle Ground

According to Puchades, the most realistic view of AI in the near term is as a clinical support tool rather than a replacement for healthcare professionals.

“The key point is that, in the short term, AI should be viewed as a support tool rather than a replacement for healthcare professionals. Diagnostic and therapeutic responsibility will continue to rest with physicians, who must supervise and validate its recommendations. While AI may be more accurate in certain static settings, such as closed clinical cases or examination questions, real world clinical environments are far more complex.”

Rather than replacing physicians, AI may provide an additional support for clinical decision-making. That conclusion was echoed by a study published in Nature, which highlighted the potential of LLMs to reduce diagnostic errors. However, the researchers also warned that these systems may generate inaccurate information when used on their own. They emphasized the importance of AI literacy among healthcare professionals so that these tools can be used appropriately as support systems rather than substitutes for clinical expertise.

“AI in real world clinical practice can help identify differential diagnoses, refine treatment decisions, summarize clinical information, detect relevant findings, and prioritize risks. However, all of this requires evidence and regulation. AI remains an intervention that requires validation and quality standards, just like diagnostic tests and treatments,” Puchades said.

Areas of Greatest Potential

Despite these limitations, some specialties appear particularly well positioned to benefit from AI-assisted diagnosis.

“AI is already having a tangible impact in data and image-driven fields such as radiology, dermatology, ophthalmology, and electrocardiogram interpretation, where comparisons tend to be more objective. It is also improving risk stratification and the detection of complex patterns that may go unnoticed in certain diseases, although the evidence remains at an early stage. Another important area is task automation, which could allow clinicians to devote more time to patients and clinical decision-making,” he said.

However, he emphasized that healthcare professionals remain essential if AI is to achieve its full potential.

“Healthcare professionals should view AI as a tool that, when properly validated and used appropriately, can enhance capabilities, improve efficiency, and reduce repetitive tasks. It should not be viewed as an infallible or autonomous system. AI has limitations, biases, and risks, which makes structured training and a strong culture surrounding AI essential.”

Overall, available evidence suggests that AI is already adding value in specific diagnostic settings, particularly those involving medical imaging and large volumes of data. However, recent studies and expert opinion converge on the same conclusion: At least for now, AI’s role is not to replace physicians but to complement their ability to make decisions in an increasingly complex clinical environment.

This article was translated from El Médico Interactivo.