Artificial Intelligence in Medical School 2026 and Why Doctors Are Still Needed

How much artificial intelligence there really is in medical education today

The gap between public debate and teaching reality is still palpable. A cross-sectional survey by Fitzek and Choi, published in BMC Medical Education in 2024, covered 409 medical and dental students from Germany, Austria and Switzerland. Data were collected between April and July 2023. Only 18.2 per cent of participants had received formal training on artificial intelligence at that point. At the same time there was a strong correlation between technical affinity and familiarity with AI, which in practice means: those who took a private interest in the subject were informed, and those who did not were not made so by the curriculum.

How fast the field has moved since is shown by an updated scoping review by Simoni and colleagues in 2025, likewise in BMC Medical Education. The group screened 3,238 publications and included 310. Fully 161 publications, that is 52 per cent of the included material, fell within an eight-month window and dealt exclusively with undergraduate medical education. The authors equally record, however, that there is neither a standardised approach nor a consensus on AI competences or ethical frameworks.

From elective to cross-cutting task

In 2025 Marc Triola and Adam Rodman described the central bottleneck in Academic Medicine: it is not the students but the teaching staff who so far predominantly have neither the competences nor binding guidelines to integrate generative AI into the core curriculum. Their recommendation comes down to three elements: clear rules of use that protect institutional and patient data, a governance structure with ethical principles, and articulated competences for students and teachers alike. That is why AI is currently starting at many faculties as an elective, a block seminar or a longitudinal curriculum, and not as a state examination subject of its own.

What the learning-objective catalogues require

The National Competence-Based Learning Objectives Catalogue for Medicine, NKLM for short, is the reference work of the medical faculties in Germany. By its own account it describes not only subject and reasoning knowledge together with practical skills and attitudes, but expressly also overarching competences, among them interprofessional, digital and scientific competences. Digital competence is therefore established in the German reference framework, not optional.

The catalogue currently exists as the interim version NKLM 2.1, published in June 2026 according to the German Medical Faculty Association. It is maintained in two application variants, one following the 2023 draft bill and one following the medical licensing regulations in force. In parallel, work is under way on version 3.0, whose declared aim is to reduce content and focus on material relevant to practice at the start of a career. Important for context: the NKLM is a framework of orientation set by the faculties, not a legal ordinance. The reformed medical licensing regulations have not yet entered into force; legally, the 2002 regulations in their 2023 version remain authoritative.

Four areas of competence can be distilled from the catalogues and the international professional discussion that future doctors will have to master:

• Interpreting data. Sensitivity, specificity, positive and negative predictive value, pre-test probability and the calibration of a model are not statistical folklore; they determine whether an algorithmic finding means anything in an individual case.
• Critical assessment of algorithms. Which population was it trained on, which was it validated on, how does the model behave outside that population, who approved it and for what purpose.
• Data protection and data sovereignty. Which data may leave a hospital information system, and what does that mean for freely accessible language models in daily work on the ward.
• Responsibility and liability. The medical decision remains attributable. A model's suggestion is a contribution to the decision, not the decision itself.

What language models actually achieve in medical examinations

Here the evidence is now reasonable, and it is more differentiated than headlines suggest. A meta-analysis by Nouri and colleagues, published in 2025 in the Journal of Educational Evaluation for Health Professions, analysed 36 studies of language models in medical licensing examinations. Pooled overall accuracy was 72 per cent, with a 95 per cent confidence interval of 70.0 to 75.0 per cent. GPT-4 reached 81 per cent, Claude 74 per cent, Gemini or Bard 70 per cent, and GPT-3.5 by contrast 60 per cent. Across languages the range ran from 62 per cent to 77 per cent, with the differences not statistically significant.

An analysis by Kasagga and colleagues published in Cureus in 2025 looked at 120 individual evaluations across ten examination systems and nine languages. The newer model generations scored considerably higher: GPT-o1 at 95.4 per cent, DeepSeek-R1 at 92.0 per cent, GPT-4o at 89.4 per cent. Thirteen of the 16 models examined crossed the 60 per cent mark that serves in many places as the pass threshold. Model type, examination system and language together explained around 89 per cent of the heterogeneity between studies.

Against that stands a more broadly framed meta-analysis by Waldock and colleagues in the Journal of Medical Internet Research from 2024. Across 32 studies, overall accuracy in medical examinations was 0.61, with a confidence interval of 0.58 to 0.64, and for the US USMLE only 0.51. The difference from the later analyses is explained mainly by the model generation and the choice of examination formats. For dentistry the picture was more restrained: a meta-analysis in the International Dental Journal in 2025 found, across eleven studies from eight countries, 72 per cent for GPT-4, 56 per cent for Bard and 54 per cent for GPT-3.5, with the express note that accuracy remains below the threshold required for clinical application.

Where the evidence shows its limits

An examination question is a closed problem with pre-formulated answer options. Clinical work is the opposite. It is exactly here that the research becomes revealing. In 2024, Ethan Goh and colleagues randomised 50 doctors for JAMA Network Open, 26 specialists and 24 residents, to work through diagnostic cases with or without access to a language model. Median diagnostic performance was 76 per cent with model access against 74 per cent without, a difference of two percentage points and not statistically significant. The model alone performed 16 points better than either group of doctors. The mere availability of a powerful tool therefore does not automatically improve results.

A follow-up study by the same group, published in Nature Medicine in 2025, randomised 92 practising doctors to management decisions with GPT-4 plus conventional resources against conventional resources alone. Here a significant advantage of 6.5 percentage points emerged, though with more time spent per case and no significant difference between the model-supported doctors and the model alone. Taken together, both papers yield a clear lesson: the benefit does not come from the model but from the way a doctor works with it. That is a competence, and competences are taught and examined.

Then there is the question of reliability. A systematic analysis by Sallam, published in Healthcare in 2023, found concerns about inaccurate content and the risk of hallucinations, that is plausibly worded but false statements, in 96.7 per cent of the papers analysed. Nor is the educational benefit automatic: a meta-analysis by Li and colleagues in BMC Medical Education in 2025 pooled eleven randomised controlled trials with 786 medical students and found no statistically significant difference in knowledge acquisition between AI-supported and traditional teaching, with a standardised mean difference of 0.27. Advantages showed up in subgroups, particularly with longer learning phases, practice-oriented courses, the development of practical skills and student satisfaction.

Radiology and pathology between headline and evidence

No specialty is cited more often as an example of the supposed replacement of doctors than radiology. The largest randomised evidence to date comes from the Swedish MASAI trial on mammography screening. The first interim analysis by Lång and colleagues appeared in Lancet Oncology in 2023 and covered 80,033 women. The detection rate was 6.1 per 1,000 in the AI-supported group against 5.1 per 1,000 with conventional double reading, the recall rate 2.2 against 2.0 per cent, and the reading workload fell by 44.3 per cent.

The follow-up analysis by Hernström and colleagues in Lancet Digital Health in 2025 confirmed the picture across 105,934 women with 6.4 against 5.0 detections per 1,000 and a workload reduction of 44.2 per cent. The decisive endpoint followed in The Lancet in 2026: Gommers and colleagues reported, for the same cohort, 1.55 interval cancers per 1,000 in the AI group against 1.76 per 1,000 in the control group, with a sensitivity of 80.5 against 73.8 per cent and an identical specificity of 98.5 per cent in both groups. A systematic review in BMJ Open in 2025 covering 31 studies and more than two million examinations arrived at a reduction in reading volume of 40 to 90 per cent with comparable diagnostic performance.

What is notable is how these programmes are built: in MASAI, AI does not replace reporting, it replaces part of the double reading and prioritises. The medical decision remains the endpoint of the chain. Pathology shows exactly the same pattern. A multicentre diagnostic study by Wu and colleagues, published in Lancet Oncology in 2023, examined 20,954 lymph nodes on 7,991 whole-slide images from 998 patients with bladder carcinoma. The model achieved areas under the curve of 0.978 to 0.998. The genuinely informative finding, though, lies in the combination: the sensitivity of junior pathologists rose with AI support from 0.906 to 0.953, and that of experienced colleagues from 0.947 to 0.986. That is amplification, not replacement. And the amplification is greatest for those who already master the field.

Law, liability and data protection make AI competence binding

Regulation (EU) 2024/1689 on artificial intelligence entered into force on 1 August 2024, according to the European Commission, and has been generally applicable since 2 August 2026. The provisions take effect in stages: the prohibitions and the obligations on AI literacy have applied since 2 February 2025, and the governance rules and the requirements for general-purpose models since 2 August 2025. For high-risk systems in particularly sensitive areas the Commission names 2 December 2027, and for high-risk AI in products 2 August 2028.

For medicine this is no marginal matter. In its recitals the regulation expressly states that diagnostic systems in healthcare must be particularly reliable and accurate because of the risk to life and limb, and it counts medical devices and in vitro diagnostics among the products subject to third-party conformity assessment where high-risk AI is embedded. In practice that means: anyone practising medicine in future works with regulated systems whose limits they must know in order to meet their own duty of care.

That this question also occupies the profession itself is shown by the 2026 AMA survey. Eighty-eight per cent of respondents consider validation of safety and effectiveness important, 86 per cent data protection guarantees, and 31 per cent named clear liability frameworks as the very highest priority. At the same time, 88 per cent reported concern about losing their own skills. That is remarkable, because it comes from the same group that uses the technology most heavily.

This section reflects the state of the publicly available rules and does not replace legal advice in an individual case.

International comparison of the requirements

A look across borders is worthwhile, because the three systems have chosen three different routes: a professional-body one in the United States, a regulatory one in the United Kingdom and a competence-based one in Switzerland.

In the United States, the American Medical Association has deliberately opted for the term augmented intelligence, in order to stress the supporting role relative to medical judgement. Its 2026 survey of 1,692 doctors shows an average of 2.3 use cases per person against 1.1 in 2023. The most frequently named were summarising research literature and guidelines at 39 per cent, drafting discharge reports and care plans at 30 per cent, and documentation and coding at 28 per cent.

In the United Kingdom, the General Medical Council works with the Outcomes for Graduates, published in 2018 and last updated in November 2020. They contain no AI outcome of their own, but they do contain the load-bearing precursors: Outcome 19a requires graduates to be able to use decision and diagnostic technologies effectively, Outcome 19b the application of confidentiality and data protection law together with local information governance procedures, and Outcome 19e the application of the principles of health informatics in medical practice.

Switzerland steers its degree programmes through PROFILES, issued by the Joint Commission of the Swiss Medical Faculties. The current version, PROFILES 2023, links the CanMEDS roles with entrustable professional activities and situations as starting points. Framing things through entrustable activities is particularly well suited to the AI question, because it does not test bodies of knowledge but answers the question of which task a graduate can be entrusted with unsupervised.

Why this does not make the medical profession redundant

Set the studies named here side by side and a very stable pattern emerges. Models are strong where the problem is cleanly formulated and the data complete. They are weak, or untested, where both are missing. And that is precisely where the medical core lies.

• Taking the history is what creates the data in the first place. No model can analyse a symptom that was never described. The quality of any algorithmic analysis depends on the quality of the conversation that preceded it.
• The physical examination is not a dataset but an act of collection. Palpation, auscultation, gait and the overall clinical impression arise at the patient's side.
• Decisions are made under uncertainty. The MASAI figures show a sensitivity of 80.5 per cent, not 100. Someone has to carry every remaining uncertainty, weigh it and explain it to the patient.
• Responsibility attaches to a person. European regulation assigns high-risk systems in healthcare to a conformity assessment, but it does not replace the doctor's duty of care in the individual case.
• The combination wins, not the tool. In the pathology study, the sensitivity of experienced readers rose with AI support to 0.986. The gain arises among people who master the field.

For prospective students something very reassuring follows from this: the need is shifting, it is not disappearing. What changes is the share of training devoted to data competence, the critical assessment of tools and communication. Anyone beginning a medical degree today will work in a profession that is changing its tools and keeping its core task.

If you would like to know which study route fits your situation, you can arrange a free, no-obligation consultation with us.