Europe’s AI Reality: Practice Outpaces Policy and Reviews - Medscape

Ask hospital leadership in Europe about AI, and the answer usually involves a pilot program, an ethics review, and perhaps a newly formed AI lab. Ask the clinicians who see the patients, and a different picture emerges: Many are already using generative AI every day — quietly, informally, and often ahead of their institution’s ability to govern it.

That gap between institutional caution and clinical improvisation is, in many ways, the story of AI in European healthcare today.

Despite a growing supply of AI tools, their integration into European hospitals remains slow, weighed down by regulatory and organizational barriers. Many advanced tools remain stuck in trial stages. A recent survey of 35 hospital representatives across Europe found that the vast majority are currently testing at least one AI solution — testing, not yet deploying. Yet on the ground, healthcare professionals aren’t waiting. A recent UK study found that 29% of responding medical doctors reported using some form of AI in the past year — primarily for diagnostic support, such as identifying medical conditions, and generative AI tasks, such as for drafting responses to patient queries.

“What happens, of course, is that all of us right now are using large language models (LLMs) in our daily work to look things up — for example, to help with writing a short note,” said Martijn Tannemaat, MD, head of the Clinical Neurophysiology Section at Leiden University Medical Center (LUMC) in Leiden, Netherlands. Tannemaat specializes in AI applications for clinical neurophysiology. “I’m quite sure that, in an informal way, and by removing privacy-sensitive data, many doctors are using LLMs because it’s a very quick way to look something up and get the answer you’re looking for. But obviously, you always have to be aware of the risk that what the LLM is saying might not be true.”

AI is also becoming entrenched within medical education. “Part of my job involves grand rounds, where I teach junior doctors at patients’ bedsides,” said Thomas Tzimas, MD, MSc, director of the Department of Internal Medicine at General Hospital of Ioannina, Ioannina, Greece. “Each morning, I review the patients in the clinic and identify those I’ll use as teaching cases,” he said. “In the past, preparing teaching materials for those patients took a lot of time. With AI, I can take a dense, dry textbook and ask it to turn the material into multiple-choice questions. The process has become much easier.”

Still, Tzimas is proceeding with caution. “General-purpose AI outputs have to be taken with a pinch of salt, and you have to verify them against the gold-standard textbooks,” he added.

Patients Are Asking Too

The shift isn’t confined to clinicians and educators. Generative AI is also changing how patients access and interpret health information, reshaping the traditional doctor-patient relationship as patients increasingly bring AI-generated findings into consultations.

That means doctors have to give up being the “sole owners” of medical knowledge, said David Morquin, MD, PhD, head of AI strategy and data governance at CHU Montpellier, Montpellier, France. “Patients have always looked up symptoms online, but AI is often more accurate and less risky than random internet searches. Rather than resisting it, clinicians can use AI as an opportunity to help patients better understand their health, ask informed questions, and become more engaged in their care.”

Tzimas is more wary of what patients bring in. “When patients ask AI about their symptoms, the responses often include hallucinations and overstated issues,” he said. “Doctors must handle these expectations carefully.”

The evidence backs that caution. A study on the reliability of LLMs as medical assistants for the general public found that people who used them performed no better at clinical reasoning than those who didn’t — and were worse at recognizing relevant health issues. The core problem was communication: Users often gave incomplete information, and the models failed to adequately explain the correct answers back to them.

Are Hospitals Ready?

Institutionally, hospitals are moving forward, just not as fast.

Ambient listening — AI tools that record and transcribe conversations between clinicians and patients — is spreading fast, said Ilse Kant, MSc, head of the Department of AI and Digital Innovation at University Medical Center Utrecht (UMC Utrecht), Utrecht, Netherlands. “More and more physicians, but also nurses and different types of healthcare professionals, are using ambient listening to automatically transcribe and then summarize patients’ healthcare professional conversations,” she said.

Eric Wolters, MSc, a data scientist at UMC Utrecht, said the priority, now that healthcare professionals are using AI, is making sure a human stays in the loop as a safeguard. “I know that within our hospital, many of these ambient listening efforts are often initially conducted as pilots, with extensive evaluation of the outputs,” he added.

That same instinct to keep a person in the loop shapes what hospitals choose to automate in the first place.

Morquin agrees that supervision is essential to managing risk. “We avoid automating many tasks because we have to manage risk and ensure reliability,” he said. Still, some narrow use cases are mature enough for full automation. “We do use autonomous medical text categorization in specific areas,” he said. “For example, in France, there is a survey that hospitalized patients must complete. This survey includes an open-ended question, and every 6 months we review the results and generate a report. The reports contain verbatim responses, which we categorize using an LLM.”

Many hospitals are setting up AI labs. “Here at LUMC, we have our own AI group, and the hospital really encourages it because I think in general, hospital management sees the potential of AI. They are also appointing people to handle the regulatory side,” Tannemaat said. “So hospitals are preparing, but it’s still very early.”

Kant said hospitals are also recognizing that deploying AI responsibly in a live clinical setting takes a different skillset than building models on static research datasets. “We’re now beginning to exchange ideas on how to implement, deploy, and monitor these types of models in healthcare,” she said.

Can AI Save Hospital Manpower Challenges?

Hospitals are also faced with the challenge of proving the technology is worth their time and money.

AI is frequently pitched as a fix for staff shortages and administrative overload. According to a WHO report on the European region, 96% of EU states cite reducing pressure on the healthcare workforce as a major or moderate driver of AI innovation.

Morquin sees time savings as one of AI’s chief benefits, freeing clinicians to focus more on patients and less on clerical work.

Tannemaat is more cautious: He remains skeptical that AI is delivering meaningful time savings today, though he expects that to change once auto-scribes for transcribing patient conversations are fully rolled out.

“I think it’s very dangerous to say that AI will solve the manpower problems in healthcare in the near future because there will always be other drawbacks when you start using AI — such as having to check where things go wrong, or extra work needed to keep it running,” he said. “So right now, it’s a promise, but the problem hasn’t been solved yet.”

The data so far back a mixed verdict. A recent study found that introducing an AI medical scribe cut the average time clinicians spent on documentation by 29% per note, and clinicians reported feeling less stressed and more present with patients. But the median time spent editing those notes — about 93 seconds — didn’t change with continued use.

“We see that there are still a lot of promises about AI actually reducing administrative burden,” said Kant. “But it’s not so easy to scale up and adopt it in a day-to-day workflow.”

Much of the friction around AI in healthcare comes down to regulation. Europe’s AI rules are notably strict — designed to protect patient privacy and, in theory, to produce more reliable, trustworthy products — but that rigor slows things down early on, said Tannemaat.

“In a 4-year project we were working on, using electromyography recordings to train AI models, the first year and a half were mainly spent dealing with legal and regulatory matters, rather than actually working on the AI tool,” he said.

For Morquin, the path forward runs through evidence, not enthusiasm. “To prove the value of AI, you need to show that it can be professional, effective, and capable of improving patient quality of care or hospital efficiency,” he said. “Essentially, you must demonstrate that it works well, both to hospital authorities and to the general public.”

Tannemaat reported being paid for consultancies for Amgen, ArgenX, UCB Pharma, Johnson and Johnson, Peervoice and Medtalks and receiving research funding from ZonMW, NWO, ArgenX, UCB Pharma, NMD Pharma, and Huma. The NWO-funded project ARISE-NMD (ARtificial Intelligence for the analySis of Electromyography for precision diagnosis of NeuroMuscular Diseases) received in-cash and in-kind co-funding from Cadwell. All reimbursements were received by LUMC.

Morquin reported having no relevant financial relationships.

Tzimas, Kant, and Wolters did not respond to a question about relevant financial relationships.

Luca Arfini is an Italian science journalist and communications specialist based in Amsterdam, with expertise in health, sustainability, and EU policies.