Using AI in Medical School: What Works, What Doesn't, and the Rules in Europe

Most medical students already use large language models (LLMs). In a 2024 survey at Philipps University Marburg in Germany, 76.2% of respondents said they used ChatGPT for their studies, and usage went up significantly around exams. Students mostly used it to summarise material (60.7%), search the literature (57.7%) and clarify concepts (47.1%). Yet 73.3% were worried about misinformation, and only 15.5% thought it helped them retain knowledge in the long run. With 84 complete responses from one university the survey is small, but most teachers will recognise the pattern.

This guide looks honestly at where AI helps you learn, where it can hurt you, and which rules apply in Europe.

Where AI genuinely helps

1. Explaining mechanisms and answering follow-ups

LLMs are good at dialogue. You can ask for the Frank–Starling mechanism, the renin–angiotensin–aldosterone system or the action of SGLT2 inhibitors to be explained as many times and as simply as you need. You can push back with "so why does potassium rise?" and ask for analogies. That's valuable for understanding. Check the core facts afterwards against a textbook or guideline.

2. Generating questions, not just summaries

Learning science is clear on this: retrieval practice (testing yourself) and spaced repetition beat rereading and highlighting. Dunlosky and colleagues' major review rated practice testing and distributed practice as the two highest-utility techniques, and rereading and highlighting as low utility. The practical upshot is to have the AI quiz you rather than summarise for you:

• "Ask me 10 single-best-answer questions on heart failure, one at a time, and only explain after I answer."
• "Test me on the CURB-65 criteria and correct my mistakes."

You can check your score calculations with our clinical calculators, for example the CURB-65 calculator. For more on study technique, see our exam study guide for Germany, Austria and Switzerland.

3. Navigating guidelines, with sources you actually open

A full ESC guideline or a German S3 guideline can run to hundreds of pages. AI can help you find the relevant section, structure the recommendations, or spot what changed between two versions. This only works when the tool points to specific sources and you open them. An answer that comes as fluent prose with no checkable reference isn't suitable for this.

4. Writing, structure and language

LLMs are useful for outlining an essay, translating an abstract or tidying up a clumsy paragraph. For theses and assessed work, your university's rules apply (see below).

Where the limits are

Fabricated and inaccurate references

The best-known failure is the "hallucinated" reference: a citation that looks plausible but doesn't exist. Two studies put numbers on it:

Newer models do better, but not perfectly. In the Cureus study, an incorrect PMID was the most common single error and appeared in 93% of the generated papers. The rule is simple: never cite a reference you haven't opened yourself. Check the title, authors and year, and above all whether the source actually supports the claim.

Outdated knowledge

Guidelines change. The 2024 ESC atrial fibrillation guideline, for instance, replaced CHA₂DS₂-VASc with CHA₂DS₂-VA. A model trained on older data and not retrieving current sources may give you the old version, just as confidently.

Automation bias and deskilling

In 2025 the German Medical Association (Bundesärztekammer) brought together its positions on AI in medicine. It warns of a gradual loss of competence and control, and of automation bias, the tendency to over-trust machine suggestions. It also stresses that doctors must stay able to sanity-check AI output, and that understanding AI includes being able to appraise studies. For students the lesson is that AI can support your clinical reasoning but shouldn't replace it. If you only ever retrieve answers, you aren't training the skill that finals and the wards will test.

The same document notes that language models used for documentation can omit or invent content, so AI-generated text needs a final human check.

Patient data: keep it out of chatbots

Once you start clinical placements you'll handle real patient information. Two rules apply in Europe:

• GDPR: under Article 9(1), data concerning health is a special category of personal data. Processing it is prohibited unless one of a narrow set of exceptions applies.
• Professional confidentiality: national law protects patient secrets, and it usually also binds people working alongside doctors. In Germany, for example, section 203 of the Criminal Code (StGB) also covers those working with doctors as part of their professional training.

In practice, don't paste discharge letters, results or case notes with identifying details into general-purpose chatbots. Even "anonymous" details can identify someone when combined, such as a rare diagnosis, a date or a small town. If you want to think a case through, rewrite it as an abstract teaching question, and ask your hospital which tools are approved.

The EU AI Act: AI literacy is now an obligation for institutions

Article 4 of the EU AI Act has applied since 2 February 2025. It requires providers and deployers of AI systems to take measures to ensure, to their best extent, a sufficient level of AI literacy among their staff and others operating AI systems on their behalf. The obligation isn't aimed at students directly, but it does apply to hospitals and universities that deploy AI. Expect more training and more internal policies on AI use.

The German Medical Association also calls for digital skills, ethical and legal aspects, and the critical appraisal of AI tools to be taught more systematically in medical school and specialty training.

Exams, coursework and theses

Your university decides whether and how you may use AI for assessed work, not the tool. Check:

• your exam regulations and any faculty guidance on AI,
• your supervisor's expectations for a thesis or doctoral project,
• your university's good research practice rules, particularly on declaring the tools you used.

When in doubt, document what you used AI for, and ask first.

A practical workflow

Bottom line

AI can be a patient tutor and a fast search aid in medical school. It helps most when you use it to test yourself rather than to think for you. Every claim that matters needs a source you have opened, and patient data stays out of general-purpose chatbots.

Medelix is built to answer clinical questions with checkable citations from current literature and guidelines, and it is hosted in the EU. The same rule applies there: read the linked source before you rely on a statement.

Sources

1. Fußhöller A, Lechner F, Schlicker N, et al. Perceptions, Usage, and Educational Impact of ChatGPT Among Medical Students in Germany: Cross-Sectional Mixed Methods Survey. JMIR Form Res. 2025. doi:10.2196/81484
2. Bhattacharyya M, Miller VM, Bhattacharyya D, Miller LE. High Rates of Fabricated and Inaccurate References in ChatGPT-Generated Medical Content. Cureus. 2023;15:e39238. doi:10.7759/cureus.39238
3. Walters WH, Wilder EI. Fabrication and errors in the bibliographic citations generated by ChatGPT. Sci Rep. 2023;13. doi:10.1038/s41598-023-41032-5
4. Dunlosky J, Rawson KA, Marsh EJ, Nathan MJ, Willingham DT. Improving Students' Learning With Effective Learning Techniques. Psychol Sci Public Interest. 2013;14:4–58. doi:10.1177/1529100612453266
5. Bundesärztekammer (German Medical Association). Von ärztlicher Kunst mit Künstlicher Intelligenz (2025, in German). PDF
6. Regulation (EU) 2024/1689 (Artificial Intelligence Act). EUR-Lex; Article 4, AI literacy: artificialintelligenceact.eu
7. Regulation (EU) 2016/679 (General Data Protection Regulation), Article 9. EUR-Lex; Art. 9 text
8. German Criminal Code, section 203 (in German). gesetze-im-internet.de