AI in medical school: study for PASS, LAS and the EDN safely

An AI can produce fifty cardiology multiple-choice questions in ten seconds. It can also invent a contraindication, a cut-off value or a journal reference with exactly the same confidence. In medicine, what you learn will one day be used to treat someone, so the real question isn’t “should I use AI?” but “how do I use it without learning mistakes?”. This guide covers the French system (PASS, LAS, EDN and OSCEs, known in France as ECOS), but the method and the red lines apply wherever you study medicine.

In short: use AI as an examiner, not as a textbook: it asks, you answer, then you check the answer against LiSA, your specialty college’s reference text, the HAS or the public drug database. Published studies show that AI can build on an error hidden in your question and can invent references. Never paste a case from your clinical placement, even one you “anonymized” by hand: it is covered by medical confidentiality, and pseudonymized data is still personal data. On the calendar side, the EDN run from 12 to 15 October 2026, and PASS/LAS enter their last intake before the single pathway announced for 2027.

Where the French reform stands in September 2026

First year: PASS and LAS, last intake before a single pathway

PASS and LAS are the two first-year routes into French health studies. On 17 April 2026, ministers Philippe Baptiste (Higher Education) and Stéphanie Rist (Health) announced that both will be replaced, from the 2027 intake, by a single national pathway on Parcoursup (L’Etudiant). The new first year would be shared by five programs (medicine, midwifery, dentistry, pharmacy, physiotherapy) and split into three blocks: health (24 to 30 ECTS), disciplinary (24 to 30 ECTS) and cross-disciplinary skills (6 to 12 ECTS), according to Service-public.fr. L’Etudiant reports that students would keep two chances to apply, in first and second year.

If you started PASS or LAS this month, your year follows the current rules; Service-public says a transition to the new program will happen the following year. The timeline is still disputed: on 1 July 2026, Vidal reported that the conference of medical deans wants a one-year delay, arguing the reform can’t be organized properly in time and isn’t sustainable without firm funding commitments. The government is sticking to 2027.

Second cycle: EDN, OSCEs and matching

The second-cycle reform (R2C) replaced the old national ranking exams: the first session of the national digital exams (EDN) took place from 16 to 18 October 2023 (CNG). For sixth-year students, here is 2026-2027 as described by the Centre national de gestion (CNG), the body that runs these exams. Students from other European medical schools can register too; the registration window this year ran through July.

Step	When	What to know
EDN, first session	12 to 15 October 2026	Four 3-hour papers on tablets, including critical appraisal of scientific articles (LCA) on Thursday 15th (decree of 23 March 2026)
Threshold	After the EDN	14/20 on core (“rank A”) knowledge to sit the OSCEs (CNG)
EDN, second session	Before the OSCEs	Run by medical schools for students below 14/20: two 3-hour half-days, 180 to 220 questions, no LCA
National OSCEs (ECOS)	2027 dates to watch on the CNG site (2026: 2 and 3 June)	Ten stations, in two sessions of five consecutive stations per day; at least 10/20 to enter matching (CNG)
Matching	After the OSCEs	EDN 60%, OSCEs 30%, training record 10% (CNG)

The logic to remember: the EDN weigh the most, but without 14/20 on core knowledge you don’t sit the OSCEs, and without 10/20 at the OSCEs you don’t enter matching. Core knowledge is exactly where active recall pays off.

Revising with AI: what actually works

Get quizzed instead of getting summaries

The classic trap: ask for a summary of a topic and reread it. Rereading feels like knowing; testing yourself makes you remember (the evidence is in our article on active recall). Instead, ask the AI to quiz you from YOUR notes or the LiSA sheet, one question at a time, revealing the answer only after you’ve committed to yours.

For flashcards, the flashcards-anki skill turns your notes into question-and-answer cards you can import into Anki, one idea per card. Check every card before it enters your deck: a wrong card is one you’ll rehearse a hundred times.

Generate MCQs, then check them line by line

An AI-generated MCQ is a draft. France’s national health authority, the HAS, tells professionals to treat every generated output “as a proposal that may contain errors to check” (our translation), especially values, units and drug names (HAS, October 2025). Three habits:

1. Ask the AI for the syllabus item and the reference passage behind each answer, then read that passage yourself.
2. Be suspicious of numbers (thresholds, scores, time frames): that’s where plausible errors hide.
3. Use real exam questions when they exist. That’s the idea behind the qcm-sante connector: it draws from MediQAl, a public corpus of questions from French national medical examinations (CC BY 4.0 license), and marks you against the corpus answers. Even then, a past paper reflects the guidelines of the year it was written: the current reference text has the final word.

Train your clinical reasoning

The EDN also test how you reason through a case. A fictional clinical case with step-by-step questions and feedback afterwards forces you to prioritize: that’s what cas-clinique-edn does. For the foundations, semiologie-fiches has you fill in clinical examination sheets by system, anatomie-memo builds mnemonics and localization quizzes, and fiche-pharmaco-edn structures a drug class. For critical appraisal, a full EDN paper in its own right, lca-lecture-critique applies an appraisal grid (bias, validity, level of evidence), and biostat-medecine trains you to interpret p-values, confidence intervals, RR and OR.

Rehearse an OSCE station

OSCEs assess how you act in front of a standardized patient, against the clock. In text, an AI can play the patient, keep time and give you a marking grid with detailed feedback: that’s the role of ecos-station. It’s useful for rehearsing history-taking, the structure of a consultation or breaking news. But a written simulation replaces neither the physical exam nor practicing out loud with a partner. The HAS even advises professionals to keep “a share of practice done without generative AI” (our translation). That’s all the more true while you’re learning.

The sources that settle it

AI suggests; official sources decide. Keep these open:

• LiSA (UNESS): the learning record for the French second cycle. It gives access to the 367 official R2C knowledge items, with objectives ranked A or B and the clinical starting situations used in the reform. The sheets are written by teachers from the relevant specialty colleges, and access is free with your university login (UNESS).
• Specialty college reference texts: they define what’s expected of you.
• The HAS (has-sante.fr) for clinical practice guidelines.
• The French public drug database (ANSM, HAS opinions) for composition, benefit ratings and generics: see our guide to using it for the EDN. The medicaments connector, coming soon, will query it and cite the date of the data.
• Past papers, to practice on real questions.

Simple rule: if an AI tells you something you can’t find in any of these sources, treat it as false until proven otherwise.

Hallucinations: what the studies measure

The figures below come from published studies, on specific models, at a specific date. Models improve, but none of these studies measures a zero error rate.

• Great at exams, not necessarily helpful. In a randomized study published in Nature Medicine in February 2026, 1,298 participants had to identify the condition and the right course of action in ten scenarios, helped either by a model (GPT-4o, Llama 3, Command R+) or a source of their choice. On their own, the models identified the right condition in 94.9% of cases; participants using them did so in fewer than 34.5% of cases, no better than the control group (Nature Medicine). As the authors put it, “LLMs now achieve nearly perfect scores on medical licensing exams, but this does not necessarily translate to accurate performance in real-world settings.”
• It builds on your mistakes. A Mount Sinai team led by Mahmud Omar planted a single invented detail (a lab test, a sign, a disease) in 300 physician-validated clinical vignettes and gave them to six models. Depending on the model and the instructions, 50% to 82% of answers repeated or elaborated on the error; a caution prompt lowered the average rate from 66% to 44% (Communications Medicine, 2025). Rerun on GPT-5, the test gave 65% with a standard prompt and 7.67% with the caution prompt (npj Digital Medicine, 2026). In practice: if your flashcard contains an error and you ask the AI to explain it, it may explain it beautifully.
• It invents references. A team from Nice led by Mikaël Chelli asked three models to retrieve the bibliography of systematic reviews: 39.6% of GPT-3.5’s references, 28.6% of GPT-4’s and 91.4% of Bard’s were fabricated (JMIR, 2024). For critical appraisal or your thesis, check every reference in PubMed before citing it.
• Even in an official publication. In Google’s paper introducing Med-Gemini (May 2024), a model-generated CT report mentions an “old left basilar ganglia infarct”: a structure that doesn’t exist, a blend of the basal ganglia and the basilar artery (arXiv).

Our take: the counterargument is serious. Recent models know a huge amount of medicine, and a good prompt cuts errors sharply, as the GPT-5 test shows. But AI remains most dangerous exactly where you are weakest: on what you don’t know yet. A senior doctor spots the plausible error; a fourth-year student learns it. Hence the rule: AI to quiz and explain, the reference text to validate.

Medical confidentiality and GDPR: the red line

On placement you see records, reports, results. It’s tempting to paste an “interesting” case into an AI so it can walk you through it. Don’t, even without the name.

• Confidentiality binds you. The official commentary on the French medical code of ethics states that people in training who are present during examinations and decisions must keep secret what they have seen and heard (Ordre des médecins). In France, disclosing secret information is punishable by one year in prison and a €15,000 fine (Penal Code, art. 226-13).
• Health data is sensitive data under the GDPR, with a broad definition: physical or mental health, past, present or future (CNIL).
• “Anonymized by hand” isn’t anonymous. For the CNIL, France’s data protection authority, anonymization must make identification impossible irreversibly; replacing a name with initials is pseudonymization, and such data remains personal data (CNIL). An age, a ward, an admission date and a rare disease are often enough to recognize someone.
• The HAS has seen it happen: among the bad practices of professionals using generative AI, it lists sharing confidential information (HAS).

The right method: start from a fictional case (built from a syllabus item, not a patient), a past paper or a published case. And if a relative asks what “the AI says” about their symptoms, the only right answer is to send them to a doctor: neither you nor a revision tool gives medical advice.

Your faculty’s rules

There’s no single national rule: each institution sets its own, sometimes each teacher (we compare several policies in this article). The University of Angers policy, which applies to all its learners, is typical: you need your teacher’s or supervisor’s agreement to use generative AI, you may not share personal or confidential data with it, and any use must be declared in the final document (Université d’Angers). For a thesis, declaring also means citing: see how to cite AI in a thesis.

And remember what exam day looks like: at the EDN you work alone, on a tablet, in an exam center. AI is only worth something if it has made you able to answer without it.

What this changes for you, concretely

• Build your revision on LiSA and the reference texts; AI comes second, to quiz you.
• Get asked questions, answer, then check the answer in the source.
• Always verify numbers, units, drug names and references.
• Never paste anything that comes from a real patient, even without a name.
• Keep AI-free sessions: timed past papers, OSCE stations out loud with a partner.
• Note the dates: EDN from 12 to 15 October 2026; 2027 OSCE dates to watch on the CNG site.

FAQ

Can ChatGPT replace the reference texts for the EDN?

No. What’s expected is set by the official R2C items and the specialty colleges’ reference texts, available notably through LiSA. An AI can quiz you and explain, but its answers must be checked against those sources: published studies show plausible errors and invented references.

Can I paste an anonymized placement report into an AI?

No. You are bound by medical confidentiality, and removing the name by hand doesn’t anonymize anything: for the CNIL, pseudonymized data is still personal data. Work from fictional cases, past papers or published cases.

When are the 2026 EDN?

The first session runs from 12 to 15 October 2026: four 3-hour papers, including critical appraisal of articles on 15 October. You need 14/20 on core knowledge to sit the OSCEs; medical schools hold a second session before the OSCEs.

I’m in PASS or LAS this year: does the reform affect me?

Your 2026-2027 year follows the current rules. The single pathway is announced for the 2027 intake, with a transition planned for students who started in September 2026. The deans dispute the timeline, so follow your faculty’s official updates.

Further reading

• Revising for the EDN with the public drug database: sourced pharmacology sheets, benefit ratings, generics and the limits to know.
• Revising with AI without cheating: active recall, the Feynman technique and spacing, with the evidence.
• The qcm-sante connector: real exam questions, marked against the corpus answers.
• AI and academic integrity: what university policies really say.