How Medical Education Has Changed Over 100-Year Cycles
For most of the 19th century, becoming a doctor in the UK or the US meant an apprenticeship — three years tagging along with a practising physician, paying a fee, doing menial chores and absorbing what you could. By 1800 only four medical schools existed in the United States; by 1876 there were 62, almost all proprietary, profit-driven institutions producing what Flexner would later describe as "indescribably foul" graduates. There were no standardised admission requirements, no laboratories worth the name and, crucially, no national exam.
Three forces ended that world. First, the 1876 founding of Johns Hopkins, modelled on German university medicine, demonstrated that clinical training could be integrated with laboratory science. Second, Abraham Flexner's 1910 Medical Education in the United States and Canada (the Flexner Report), commissioned by the Carnegie Foundation, recommended closing more than half of American medical schools and reorienting the rest around the scientific method. Within two decades, two-thirds of American medical schools had closed or merged. Third, in parallel, the National Board of Medical Examiners (founded 1915), the Federation Licensing Examination (1973) and finally the USMLE (1992) imposed a single national knowledge standard. The UK's PLAB (1970s) and now the UKMLA (rolled out 2024–2026) follow the same trajectory: from variable institutional finals to a single national licensing assessment.
The 2010 Lancet Commission on the Education of Health Professionals for the 21st Century, led by Julio Frenk and Lincoln Chen, explicitly framed itself as a centennial successor to Flexner. It diagnosed three "generations" of reform - Flexner's science-based curriculum (1910s), problem-based learning (1970s) and now systems-based, competency-led, interdependent education for the 21st century - and argued that medical schools were producing graduates "ill-equipped to provide health care in the current and future environment." A 2022 Lancet follow-up, written after the COVID-19 pandemic, doubled down: the next decade, it argued, would see "competency-based education, interprofessional education, and the large-scale application of information technology" finally realised at scale.
If history is any guide, the AKT of 2126 will bear roughly the same relationship to the AKT of 2026 as the AKT of 2026 bears to a viva voce examination conducted at Edinburgh in 1810 - recognisable in lineage, unrecognisable in form.
Where Medical Schools Are Heading in the Next 50–100 Years
The 2019 Topol Review, commissioned by the UK government and led by Eric Topol, is the most authoritative UK-specific futures document. Its core forecast is that genomics, digital medicine, AI and robotics will reshape every clinical role within 20 years, and that medical education must therefore be rebuilt around four pillars: genomic literacy for all clinicians, digital fluency, data interpretation, and ethical reasoning about machine-augmented decisions. Health Education England's 2023 progress report shows uneven but real implementation - DART-Ed (Digital, AI and Robotics Technologies in Education) programmes, NHS Genomic Medicine training, and digital fellowships are now embedded in UK postgraduate training.
The scholarly consensus on the content of future medical school curricula is now relatively settled across Academic Medicine, Medical Teacher, Medical Education and The Lancet Digital Health:
Genomic and precision medicine literacy. Whole-genome sequencing has fallen below US$1,000 and is forecast to be a routine part of primary care within 20 years. Yet a 2023 review in PMC found precision medicine featured in only ~21% of undergraduate curricula. Expect this to be a major area of curricular expansion.
AI and data science. A 2025 Lancet Digital Health viewpoint argued that "adapting to an AI-enabled future will necessitate dramatic changes in medical education, practice, regulation and technology." The AMA-Manatt report counted 692 FDA-approved AI/ML medical devices as of late 2024.
Simulation and virtual reality. A 2025 Advances in Simulation study found high-fidelity manikins and VR were comparable for assessing acute clinical skills; the ACGME has now formally recognised VR as a valid platform for procedural assessment.
Humanism, communication and uncertainty. As AI takes over rote diagnosis, scholars from Harden onwards predict the human-skill component will expand, not shrink. The UKMLA 2026 content map already adds "managing uncertainty" as a core competency.
Lifelong, time-variable, competency-based learning. The traditional five-year medical degree may give way to time-variable training in which trainees progress when they demonstrate competence (Royal College of Physicians and Surgeons of Canada's "Competence by Design" is the leading example).
The Future of the AKT and Knowledge-Based Licensing Exams
This is the question students preparing for the AKT actually want answered. The honest answer is: nobody knows, but the trend lines are visible.
The case that the AKT will survive (broadly) intact. Knowledge thresholds protect patients. Every previous "knowledge is now Googleable" moment - the printing press, the textbook, the internet, UpToDate, smartphones - was supposed to make rote licensing exams obsolete; none did. Regulators (the GMC, NBME, FSMB, MCC) have a deep institutional and legal incentive to maintain a defensible, standardised threshold. The 2024 USMLE Step 1 transition to pass/fail is instructive: the exam was reformed, not abolished, and the passing standard was actually raised by two points. The GMC has signalled the AKT will remain numerically scored (unlike Step 1) because that score is useful for medical schools and Foundation Programme allocation.
The case that the AKT will transform substantially. Several pressures are converging:
AI performance. GPT-4 scored ~76% on sample UK MLA AKT questions in 2024 (medRxiv pre-print) and 80–90% on USMLE Steps. A 2023 PMC commentary in response argued that “ChatGPT's success also reflects the rigidity in the way medicine is taught, wherein there is a right and wrong answer (that an AI chatbot could pick out), while the 'right' answer may be far more nuanced and contextually dependent.”
Cheating arms races. Generative AI browser extensions and microphone earpieces have made remote-proctored MCQ exams progressively harder to defend. The 2024 AIIMS Delhi ChatGPT cheating scandal is a likely harbinger.
Programmatic assessment. Cees van der Vleuten and colleagues have argued for two decades that the "big bang" end-of-course examination is psychometrically and pedagogically inferior to a longitudinal collection of low-stakes data points feeding a competency committee. The 2024 Perspectives on Medical Education "Next Era of Assessment" special issue, drawing on a 2022 invitational summit, codifies this as the emerging international consensus.
Computerised adaptive testing. The NCLEX (US nursing licensure) has used CAT since 1994; the Korean Medical Licensing Examination is in active CAT pilot. CAT can deliver a defensible pass/fail decision in ~40–60% of the items of a fixed-form test, freeing time for richer assessment formats.
Entrustable Professional Activities (EPAs). Pioneered by Olle ten Cate in 2005, EPAs are now embedded in AAMC core competencies, the Canadian CanMEDS framework, and increasingly in UK foundation training. They reframe assessment around "what can this trainee be trusted to do?" rather than “what facts can they recall?”
Most plausible 50-year trajectory for the AKT. Synthesising scholarly commentary, the most likely future is not abolition but integration. By 2076, the AKT (or its successor) will probably be:
Shorter and adaptive. CAT-delivered, perhaps 60–90 items rather than 180, calibrated to each candidate.
Multimodal. Embedded video, audio (e.g. heart sounds, patient interviews), and dynamic data (e.g. evolving observation charts) - building on what the GMC is already piloting.
AI-proctored and AI-authored. Items generated and quality-assured by LLMs, with behaviour-analytics flagging suspicious patterns.
One node in a programmatic web. Sitting alongside an e-portfolio of workplace-based assessments, EPA entrustment decisions, simulator performance data and possibly wearable-derived performance metrics during clinical placements.
More about reasoning under uncertainty than fact recall. Script-concordance items, very-short-answer questions, and clinical-prioritisation questions are all being trialled. The 2026 UKMLA content map already shifts in this direction with longer vignettes and a "manage uncertainty" competency.
By 2126? Speculation becomes harder. A serious minority view - represented by Bertalan Meskó, the "Medical Futurist," and by the more techno-deterministic strand of the Topol school - holds that knowledge-based gating exams will eventually be replaced entirely by continuous performance monitoring (anonymised consultation transcripts, ambient AI scribes, outcomes-attribution algorithms linking each clinician to their patients' actual outcomes). The 2020 Journal of Graduate Medical Education paper "Envisioning Graduate Medical Education in 2030" already articulates this: “By 2030, our current health care education management platforms will be relics… attribution algorithms will link individual residents to patients and outcome data.”
Will AI Make Clinical Knowledge Assessment Obsolete? The Debate
Arguments that AI changes everything:
LLMs already outperform the median medical student on USMLE/UKMLA-style exams, and performance is improving roughly an order of magnitude per generation.
A 2023 commentary in PMC argued that ChatGPT's success “is an indictment of what and how we train our doctors.”
If every clinician will work with an AI co-pilot at the bedside, why test what the co-pilot will know?
Bill Gates publicly predicted in 2025 that AI will be “as good as any doctor within the next decade.”
Arguments that knowledge assessment remains essential:
LLMs hallucinate, especially in out-of-distribution cases. A 2025 arXiv paper (Facts Fade Fast) showed that all eight tested LLMs systematically reproduced outdated medical knowledge from older training data - a serious patient-safety problem.
Clinicians need to recognise when AI output is wrong, which requires deep underlying knowledge. A 2025 PMC viewpoint titled "Not Replaced, but Reinvented" argued that doctors must lead AI integration, which requires fluency, not deference.
Liability, regulation and the doctor-patient relationship still require an accountable, qualified human.
The American Medical Association's now-canonical formulation, attributed to former AMA President Jesse Ehrenfeld: "AI is not going to replace doctors, but doctors using AI will replace doctors who aren't using AI."
A 2025 PMC narrative review concluded: “In the foreseeable future, AI will augment rather than replace physicians.”
The probable synthesis - and what we think any AKT candidate should internalise - is that the threshold of medical knowledge required for safe practice will not fall, even as the uses of that knowledge change radically. The AKT tests whether you have a safe baseline for unsupervised practice; that question will remain meaningful as long as patients want a human to be accountable.
Emerging Assessment Modalities to Watch
Entrustable Professional Activities (EPAs): Now embedded in undergraduate medical education in the US, Canada, the Netherlands and Sweden. The University of Minnesota now requires "indirect supervision"–level entrustment on 7 EPAs for graduation as of 2026.
Programmatic assessment: Multiple low-stakes data points → competency committee → high-stakes decision. Adopted at scale at Maastricht, Utrecht, and increasingly UK schools post-COVID.
Workplace-based assessment with mobile technology: AMEE Guide No. 154 (2023) provides explicit best-practice guidance for tablet/phone-based EPA capture in the clinical workplace.
Virtual reality OSCEs: A 2025 systematic review in PMC found 26 published studies using immersive VR for nursing/medical assessment, mostly in emergency scenarios, with the ACGME formally endorsing VR for procedural skills.
Script Concordance Tests (SCTs): Designed specifically to assess clinical reasoning under uncertainty. AI-generated SCTs have been shown to perform comparably to human-expert-authored ones (Medical Teacher, 2025).
AI-generated and AI-graded items: LLMs are now being used for automatic item generation (AIG) for licensing exams in pharmacotherapy, surgery and the Korean MLE, dramatically reducing faculty workload.
Continuous-monitoring assessment via wearables and digital portfolios: Still aspirational, but signalled clearly in the J Graduate Medical Education 2030 vision.
AI-proctored exams: Now standard for remote testing, but with serious documented bias (a 2022 arXiv paper demonstrated that facial-recognition proctoring software disproportionately flags darker-skinned candidates).
Specifically on the UKMLA/AKT
The GMC's official position, expressed through its Vision for the future of medical education and training and the 2026 MLA content-map revision, is incrementalist: keep the two-part structure (AKT + CPSA), make the content map a living document, integrate genomics and uncertainty, raise the standard for clinical reasoning over recall. The 2026 update expanded core conditions from ~311 to ~430.
UK-specific commentators (the iatroX 2026 guide; the Medical Schools Council; the BMC Medical Education 2026 study on UKMLA preparation resources) anticipate three near-term changes:
International CPSA test centres (likely in major IMG source countries by 2030).
More frequent content-map updates (annual or biennial rather than every 5–10 years).
AI-augmented exam authoring and item-bank curation drawn from the MSCAA bank.
The likelihood of the AKT being abolished outright in the next 30 years is, in the judgement of most scholarly commentators, low. The likelihood of its question style, length, and surrounding ecosystem being unrecognisable in 50 years is high.
What the Doctor of 2126 Might Actually Look Like
Synthesising Ronald Harden's Ten key features of the future medical school (2018), the Lancet Digital Health 2025 viewpoint, the Topol Review, the 2025 Cureus University-of-Nottingham 2050 study, and the AMA's "physician of the future" articles, a composite portrait emerges:
A specialist generalist. With AI handling much rote diagnosis, the physician's value lies in integration, judgement, communication, and accountability across systems.
A genomic interpreter. Routine whole-genome sequencing will make pharmacogenomics and polygenic risk scoring part of every consultation.
A data scientist, partly. Comfortable with the limitations and outputs of multiple AI systems, able to audit them.
A humanist. As the technical floor rises, the differentiating skills are empathy, ethics, end-of-life conversations, advocacy, and trust.
A team-based clinician. The solo practitioner is gone; doctors of 2126 will work in interprofessional, geographically distributed teams.
- A lifelong learner whose "education" never formally ends. Continuous programmatic re-assessment will replace the discrete medical school → residency → consultancy → retirement model.
The melancholy version of this picture, articulated by the physician-blogger Bryan Vartabedian in 2012 and echoed widely since, is that "the physician of 2050 will have workflows and ways of thinking that are critical but unrecognizable to today's physician. But we're not prepared for the future. The next generation is not ready for the changes that are coming." That warning is now 14 years old, and arguably more urgent.
What This Means for AKT Candidates Today
Two practical points fall out of all this for a student sitting the 2026 AKT:
The fundamentals you are revising - pathophysiology, NICE guidelines, safe prescribing, recognition of the deteriorating patient - are not what AI is going to make obsolete. They are what makes you a safe day-one doctor and what makes you a competent supervisor of AI when you're a year-five registrar. Revise them properly.
- The style of question that will define your career-long assessment is already shifting under your feet. The 2026 content map's emphasis on clinical reasoning, uncertainty, and 430-condition breadth (versus 311) is a small preview of a much bigger transformation. Cultivate the habit of reasoning aloud, of saying "I don't know but here's how I'd find out," and of using AI tools critically. That habit will outlast any one exam format.