Artificial intelligence has been likened, in terms of importance, to the invention of the wheel. A striking image for describing a technology on which several countries are pouring unprecedented investments—hundreds of billions of dollars—chiefly in the United States and China. Europe, with fewer industrial resources, has chosen to govern its impact mainly through rules and safeguards, as demonstrated by the AI Act. The readiness of enormous quantities of data, images, and medical reports makes a tool capable of organizing complexity and supporting clinical judgment increasingly valuable, without replacing the physician’s responsibility. The FNOMCeO (Federation of the National Orders of Physicians, Surgeons and Dentists) has dedicated several conferences and updates to the topic, testifying to an interest that is no longer theoretical but deeply operational.
In recent meetings, specialists, jurists, and risk managers outlined already concrete applications: systems that support patient self-triage, monitor chronic conditions remotely, automate repetitive tasks, and assist in follow-up management. In radiology, artificial intelligence contributes to recognizing lesions in mammography screening, to the early diagnosis of stroke, and to the three-dimensional reconstruction of anatomical structures. Far from a “digital doctor,” the technology is shaping up as a decision-support tool capable of lightening low‑risk activities and restoring time to clinicians for more complex cases.
Evidence, challenges and prospects, what some studies say
A Microsoft AI–led collaboration used 304 complex clinical cases published in the New England Journal of Medicine to compare the performance of an AI system with that of 21 experienced physicians. The system, called MAI-DxO, achieved the correct diagnosis in 85.5% of cases, while the doctors, working without consulting colleagues or external tools, hovered around 20%. The finding drew significant attention and suggests that AI can excel in highly structured experimental contexts. The most delicate issue is patients’ direct access to these tools. A Jama Network Open study evaluated 21 general‑use chatbots. When given all clinical information, the top models exceed 90% accuracy in the final diagnosis; but they fail in about 80% of cases to formulate a correct differential diagnosis, that is, the initial and more clinical step of medical reasoning. In other words, they are very effective at recognizing a solution when all data are available, but still fragile in navigating uncertainty.
A similar conclusion emerges from a recent Nature Medicine study showing that conversational systems perform well in moderately complex cases but tend to underestimate emergencies and urgent situations in a significant share of scenarios, in some cases more than 50%. The risk is not only error but false reassurance: a seemingly plausible recommendation that delays access to appropriate care. In this context, one of the leading AI developers announced in January 2026 new health features, currently available only in some countries and not yet accessible in Italy. The signal is clear: the major tech players view medicine as one of the most strategic domains for development.
Interoperable data and artificial intelligence: the future of Italian health care
The Italian Institute for Artificial Intelligence (The Italian Institute for Artificial Intelligence AI4I) is working to build structural capacities: HPC infrastructure, technology transfer, AI safety, and the connection between research and industrial deployment. The strategic direction is to move artificial intelligence from research to concrete adoption in production systems and services. In healthcare, this means transforming AI from a theoretical promise into a tool embedded in daily clinical practice.
According to Antonio Calegari, director of the Italian Institute of Artificial Intelligence AI4I, the health system is entering a phase where AI is no longer seen as futuristic technology, but as a tool that will concretely impact how medicine is organized, how professionals work, and how patients are managed. The challenge today is to identify which applications are truly ready for clinical practice and which require further validation and consolidation. “The most mature applications today are those already integrated into regulated and measurable pathways, particularly in clinical imaging: radiology, quantitative image analysis, triage, and reporting support. It is there that AI has reached the highest level of industrialization and the greatest number of solutions that are actually authorized and used. More recently is the case of generative AI, which today finds application mainly as support to the physician’s work: summarizing documentation, drafting notes, summarizing patient records, supporting coding and literature review.” In Calegari’s view, the issue of clinical responsibility and the need to maintain human oversight of therapeutic decisions remains central, even as systems grow more advanced.
“The clinical responsibility remains human at all times, and this is explicitly reaffirmed by the recent Italian AI regulation, in line with the European AI Act.” Looking ahead to the next five years, the fastest development prospects concern technologies capable of reducing administrative burden, improving monitoring of chronic patients, and enabling more continuous care management. “The first is imaging, because it starts from standardized data, clear workflows, and stronger evidence. Then comes the automation of clinical documentation, which can free up physician time and reduce administrative load and professional stress. The third is remote monitoring of chronic patients: sensors and devices collect data at the patient’s home, and AI helps identify early signs of deterioration so we can intervene sooner and avoid hospitalizations. Finally, more strategic is the integration of clinical data, omics data—such as genetic data—and longitudinal data, but here the decisive factor will be the quality of the data infrastructure.” On the international stage the global debate on health AI is currently framed by models that differ greatly. “Europe, however, can build a different advantage: trust, regulatory quality and interoperability. It is precisely the availability of interoperable and regulated data that can become Europe’s true competitive edge.”
For Italy, the game is mostly about building a truly integrated national digital infrastructure capable of linking clinical data, research, and healthcare practice. The objective is to transform the National Health Service’s informational heritage into a strategic lever for innovation and personalized medicine. “Without data that talk to one another, AI stays confined to pilot projects. In this sense, the Electronic Health Record 2.0 (FSE 2.0) and the Health Data Ecosystem coordinated by Agenas are central, because they enable the collection, standardization, and interoperability of clinical data on a national scale.”
Among the projects AI4I is pursuing are laboratories dedicated to the safety of AI systems, validation of algorithms, and the development of technologies applied to smart devices and continuous patient monitoring. A research line that especially targets the management of chronic illnesses and primary care. “The key is for AI to be invisible within the workflow: not another tool to manage, but an ability that inserts itself into existing processes and makes them more efficient. If we can successfully integrate these technologies into the public health system, AI can become one of the most important levers to improve professionals’ work, the quality of care, and the sustainability of the health system.”
Article taken from the October 2026 issue of Karla Miller
In the second part: diagnosis, research and the relationship between algorithms and clinical judgment
Bibliography
1. Regulation (EU) 2024/1689 (Artificial Intelligence Act).
2. OpenAI. Introducing ChatGPT Health. January 7, 2026.
3. Nori, H., Daswani, M., Kelly, C., Lundberg, S., Ribeiro, M. T., Wilson, M., et al. (2025). Sequential diagnosis with language models. arXiv preprint arXiv:2506.22405.
4. Huo B, Boyle A, Marfo N, et al. Large Language Models for Chatbot Health Advice Studies: A Systematic Review. JAMA Netw Open. 2025;8(2):e2457879.
5. Rao AS, Esmail KP, Lee RS, et al. Large Language Model Performance and Clinical Reasoning Tasks. JAMA Netw Open. 2026;9(4):e264003.
6. Ramaswamy A, Tyagi A, Hugo H, et al. ChatGPT Health performance in a structured test of triage recommendations. Nat Med. 2026;32(5):1671-1675.
Abbonati a Karla Miller
