A Harvard study that compared the diagnostic accuracy of AI models with emergency doctors generated an immediate impact on the global medical and tech community. The results showed that the AI models achieved greater accuracy in the differential diagnosis of emergency cases than the human professionals, even in scenarios of high clinical complexity.

Beyond the impact on medicine, the study opens urgent questions for the tech world: what new roles emerge in digital health?, what really limits the massive adoption of medical AI?, and how do technology professionals prepare to work in this sector in Argentina and LATAM?

The study's results in detail

The study, reported by TechCrunch on May 3, 2026, evaluated large language models (LLMs) trained with clinical data against groups of doctors in real emergency situations. The AI models demonstrated greater accuracy in generating differential diagnoses —the list of possible diagnoses ordered by probability— compared to the groups of doctors under the same information conditions.

The Harvard researchers were careful to point out that context matters: the models had access to all the structured data of the case (history, symptoms, laboratory results), while the doctors sometimes worked with incomplete information or under additional time pressure. But the results are statistically significant and represent a milestone in the medical AI literature.

How does medical AI work?

AI systems applied to medical diagnosis combine two large families of techniques:

Natural language processing (NLP) on clinical text

Systems trained with millions of clinical records, medical studies, and scientific literature learn to recognize symptom patterns, relate them to known diagnoses, and generate responses in medical language. This is what allows an LLM like GPT-4 or specialized clinical models (like Google's Med-PaLM 2) to reason about medical cases described in text.

AI on clinical imaging

Systems specialized in computer vision analyze X-rays, CT scans, MRIs, and histopathology slides. In these tasks, AI has been surpassing human specialists for years in speed and, in some studies, also in accuracy for specific pathologies like skin cancer, diabetic retinopathy, and pulmonary nodules.

What limits the massive adoption of medical AI?

Despite the results of the Harvard study and decades of research in medical AI, adoption in real clinical practice remains slow. The obstacles are multiple and not all are technological:

• Regulation: AI clinical decision support systems require regulatory approval (FDA in the US, ANMAT in Argentina) which is a long and costly process. Few systems have approval for real clinical use beyond informational assistance.

• Legal responsibility: If an AI system recommends an erroneous diagnosis, who is responsible? This question has no clear legal answer in most countries, including Argentina, and generates resistance in hospitals and health institutions.

• Data interoperability: Hospital systems in LATAM are very heterogeneous. Most public hospitals in Argentina don't have a unified electronic health record, which makes it impossible to feed AI models with structured data at scale.

• Biases in the training data: Most of the medical datasets used to train LLMs are from patients in the US or Europe, which can generate biases when applying those models to Latin American populations with different genetic and epidemiological profiles.

• Cultural resistance: Doctors and patients have understandable reservations about delegating clinical decisions to algorithms. Medical AI is more successful when presented as a second opinion or assistant, not as a replacement.

These challenges are similar to those faced by large AI models in general, as described in the analysis of Claude Mythos 5 and the limits of scaling language models.

Tech roles emerging in digital health in Argentina and LATAM

The growth of AI in health is generating a demand for specialized tech profiles that the job market in Argentina and LATAM still doesn't have in sufficient quantity:

Health AI Engineer

A developer or ML engineer specialized in the health domain. They work on integrating AI models with hospital systems (HIS/EMR), on processing clinical data, and on validating models for regulated environments. It's a profile that combines software engineering skills with medical domain knowledge and health regulation.

Clinical Data Scientist

A data analyst specialized in clinical data. Their focus is transforming health data (diagnoses, prescriptions, laboratory results) into actionable insights to improve care, reduce costs, or identify epidemiological patterns. They work in hospitals, insurers, laboratories, and healthtech startups.

Health Informatics Specialist

A profile that connects hospital information systems with AI initiatives. They know both the clinical and the technological logic, and act as a translator between doctors and engineers. In Argentina, the Argentine Society of Medical Informatics (SAIM) works actively on the training of this profile.

According to the World Economic Forum's Future of Jobs Report 2025, AI roles in health are among those with the highest projected growth in the next decade globally.

Coderhouse courses to enter the world of AI in health

If you're interested in working at the intersection of AI and health, these courses are the starting point for building the skills the sector demands:

• Introduction to Artificial Intelligence: Understand how language and vision models work, the same ones applied in medical diagnosis. The theoretical base is essential before specializing.

• AI Engineering Course: To build applications with AI in production, including RAG systems and agents that can integrate with health systems.

• AI Automation Career: Ideal for learning to automate workflows in complex environments, which is exactly what hospitals and healthtech companies that seek to incorporate AI into their processes demand.

Frequently asked questions

Will AI replace doctors?

No. AI is a clinical decision support tool, not a substitute for the doctor-patient relationship or for comprehensive clinical judgment. The studies that show greater accuracy of AI refer to specific and limited tasks (differential diagnosis from structured data), not to medical practice in its entirety. The most likely and most desirable scenario is that of doctors who use AI as a second opinion, similar to how they use scientific literature today.

How is medical AI regulated in Argentina?

In Argentina, medical devices with AI are under the regulation of ANMAT (National Administration of Drugs, Foods, and Medical Technology). Software systems that influence diagnostic or therapeutic decisions are classified as software as a medical device (SaMD) and require health registration. The regulatory framework is in the process of being updated to adapt to the speed of AI advancement.

What data does an AI model need to diagnose?

It depends on the type of system. Clinical LLMs work with text: symptoms, history, records. Medical vision systems need images (X-rays, CT scans, dermatological photos). The most advanced systems integrate multiple modalities (text + image + laboratory data). In all cases, the quality and quantity of the training data is the most determining factor of the model's accuracy.

Are there medical AI startups in Argentina?

Yes, the healthtech ecosystem in Argentina is growing. Companies like Nuvita (telemedicine with AI), Medintt, and other startups are developing local solutions. There are also AI labs in public hospitals like the Hospital Italiano and the Garrahan working on applied research projects. The ecosystem is small but active and with a lot of growth potential.

Does medical AI work the same for Latin American populations?

This is one of the main points of attention for researchers. Models trained mainly with data from the US or Europe may have lower accuracy for Latin American populations due to genetic and epidemiological differences and differences in disease presentation patterns. The generation of local clinical datasets is one of the field's priorities to improve the equity of the models.