Content Benefit

Content Strategy | UX Writing | Content Design | Globalization | Communications

AI in care, not in your pocket

AI in care, not in your pocket

Where healthcare AI projects get stuck

On Agenda Digitale, Sergio Cuschera and Luigi De Angelis turn the usual question upside down. Rather than asking why so many hospital initiatives around artificial intelligence stall, they look at what allows a clinician to see one through. The answer, they argue, does not live on the wards alone: it starts further upstream, in the industrial world that builds those tools. Their model rests on five legs, where technology, processes and governance have to talk to each other all the way to the patient’s bed.

The framework holds up, and it deserves widening. Public debate swings between two rather sterile poles: software that will replace doctors, and software that will wreck everything. In between sits solid ground, made of laboratories, test results and operating theatres, where outcomes can already be counted.

Drug research has changed pace

The least discussed chapter is also the sturdiest. The AlphaFold database holds more than 200 million predicted protein structures and serves over three million researchers scattered across 190 countries, a million of them in low and middle income economies. Roughly a third of the related scientific literature deals with understanding disease, atherosclerosis included.

Prediction has given way to design: Isomorphic Labs, the pharmaceutical offshoot of DeepMind, is moving computer-drawn cancer molecules towards their first human trials, backed by agreements with Eli Lilly and Novartis and a 600 million dollar round closed in 2025. Here the algorithm shortens the slowest and priciest stretch of a pipeline that used to burn a decade before reaching a vial, without taking anyone’s job away.

Mammography screening and the numbers from the MASAI trial

Early detection finally has a large randomised trial behind it. MASAI enrolled more than 100,000 Swedish women between April 2021 and December 2022, and the definitive results appeared in The Lancet on 29 January 2026. Assisted reading picked up 81% of cancers against 74% for conventional double reading, with almost identical false positives (1.5% versus 1.4%). Carcinomas surfacing between one round and the next fell by 12%, the large ones by 21%, the most aggressive subtypes by 27%. The 2023 interim analysis had already recorded a 44% drop in the volume of images to examine.

Translated: fewer late diagnoses, and specialists relieved of a caseload no organisation can absorb any more.

Doctors and algorithms fail in different ways

A 2025 PNAS paper from the Max Planck Institute for Human Development, written with the Human Diagnosis Project and the Institute of Cognitive Sciences and Technologies of the Italian National Research Council, combed through 40,000 diagnoses across 2,100 simulated clinical cases, putting five systems to the test, GPT-4, Gemini and Claude 3 among them. Mixed groups, professionals alongside machines, beat both the all-human collectives and the purely artificial ones.

The interesting detail sits in the explanation: the two categories stumble over systematically different errors, so they cover for each other. Neither, taken on its own, performed as well.

In the operating theatre, supervision remains a craft

In July 2025 Science Robotics documented SRT-H, the system developed at Johns Hopkins, able to carry out a phase of gallbladder removal autonomously on an ex vivo model, correcting its trajectory from spoken instructions. A remarkable achievement and, it should be said, still a long way from a living patient.

Assisted robotics, meanwhile, has been at work for a quarter of a century: the first Italian da Vinci arrived at San Matteo in Pavia in July 1999, and over 300,000 people have since passed through a robotic theatre in the country. The mechanical arm dampens tremor, filters movement, lights up a tiny surgical field. Whoever decides where to cut still wears a gown.

When constant assistance costs a professional the knack

The flip side carries an unflattering name, deskilling. An observational study across four Polish centres, published in the Lancet Gastroenterology & Hepatology on 12 August 2025, followed 19 experienced endoscopists, each with at least 2,000 procedures behind them. In colonoscopies performed without computer support, adenoma detection slid from 28.4% to 22.4% once the software had arrived: a fifth lower in relative terms.

Whoever designs such systems should keep that in mind, much as interface designers weigh how much to hand over to the machine. Steady help that dulls the attention of the person watching yields an apparent gain and a deferred loss.

Self-diagnosis deserves a discussion of its own

This is where things fall apart, and it is the point closest to my heart. Between a properly equipped ward and a phone in a pocket runs a distance that the evidence measures rather well. In February 2026 Nature Medicine hosted research from the Oxford Internet Institute and the Nuffield Department of Primary Care Health Sciences, run with roughly 1,300 participants, pitting GPT, Llama and Command R+ against plausible clinical scenarios. Queried on their own, the models recognised the condition in 94.9% of cases. Handed to ordinary users, they led to the right identification in fewer than 34.5% of attempts, and to the correct course of action, family doctor or emergency room, in fewer than 44.2%. Against the group muddling through an online search, no advantage at all.

Three causes emerged: nobody knows which details to supply, the reply shifts when a comma shifts in the question, and a sensible tip shares a paragraph with nonsense. The mechanism echoes the automatic summaries of search engines and their cheerful approximation.

How many Italians ask chatbots about their health

The UniSalute Health Observatory, produced by Nomisma with a sample of 1,300 Italians aged between 18 and 70, captures a settled habit. 60% look up illnesses and medicines with digital tools, 49% turn to artificial intelligence at least once a week, 27% ask it for guidance on their own condition. 42% want to know whether a symptom warrants a visit, 41% try to decipher a test report.

The rest of the picture reassures: 75% rate the doctor more trustworthy, 73% more empathetic, 63% believe the technology can assist without replacing. Popular wisdom is holding, then, though that 27% deserves better company, above all when an automated answer contradicts the specialist and the patient chooses to believe the former. It is the updated version of the well-informed cousin who explains thyroids to everybody over Christmas dinner.

The skills missing between the model and the patient

Back to the five legs of Cuschera and De Angelis, with one more from my own trade. Between a validated system and a frightened person there is always a text: the report, the message in the app, the disclaimer, the screen stating how reliable a result happens to be. Writing those lines means deciding whether somebody books an appointment or puts it off.

It applies to lab results as it does to wrist sensors, which I covered when writing about wearables and health: a figure without context breeds anxiety or indifference, seldom a considered choice. Anyone working on clinical content should insist on a seat beside engineers and consultants, and learn the practical limits of these tools before describing them.

Artificial intelligence in medicine works when it shortens discovery, when it finds in an image what a tired eye lets slip, when it steadies a clamp. It stops working the moment somebody mistakes it for a free professional opinion, lit up on a screen at three in the morning and forever agreeing with whoever asks.

Related sources

Chiara Bonifazi