Unlimited repetition is virtual patients' headline gift and their quietest hazard, and both halves are true at once. Repetition builds skill: consultation structure, questioning fluency and phrase availability all improve with volume, which is why unlimited private practice genuinely levels access. And repetition of an identical case eventually stops teaching the skill and starts teaching the script: the student learns this patient's disclosures, this scenario's turning points, this model's response patterns, and scores rise while transferable consultation ability stands still, memorisation wearing competence's costume. The problem has a detection method and a design solution, and both fit in this article.
How script-learning happens, mechanically
Three loops close around the repeating student. Content memory: the specific facts, the medication list, the hidden concern, the family history, surface earlier each run because they are remembered, not elicited, so the questioning skill that would find them in a new patient goes unexercised. Sequence memory: the scenario's productive path, which question unlocked which disclosure, gets replayed rather than rediscovered, and consultation flexibility, the actual skill, atrophies in the groove. And model-pattern memory: students learn the simulator's tells, the phrasings it rewards, the rubric's trigger words, which is optimisation of the measurement rather than the ability, and it transfers to real patients not at all. None of this requires laziness; it is what practice on a static target produces in any domain, and the signature is diagnostic: rising scores with falling effort and no felt uncertainty, the state real consultations never produce.
The scenario-mutation test for platforms
The design solution is variation, and buyers can test for it directly rather than trusting the word "unlimited". Run the same case three times and audit what actually varied: did the patient's presenting concern, emotional state, comorbidities, health literacy or disclosures change, or did the surface wording regenerate around an identical hidden state? Genuine mutation means the third run still requires elicitation, the questions must be asked because the answers moved; surface regeneration means the third run rewards recall, and the platform's repetition is a memorisation engine with a conversation interface. The question belongs in every procurement conversation and every student trial, phrased exactly: when I repeat a case, what varies, concerns, state, disclosures, or wording? Platforms engineering real variation, different ages, altered priorities, new complications, shifted emotional registers, are building the transfer machine; the ten-standards benchmark's consistency and disclosure tests overlap here deliberately: /blog/ten-standards-realistic-ai-patient-simulation.
The self-checklist: catching memorisation in your own practice
Five signs, checkable weekly. You predict disclosures before asking, content memory has replaced elicitation. Your question order has fossilised, sequence memory has replaced clinical responsiveness, and the fix is deliberate scrambling, open somewhere new. Scores rise while the case feels easier, the effort signature inverted; learning feels harder before it feels better, and comfort at ceiling is the script's smell. You would struggle to handle the same complaint in a different patient, the transfer test, runnable immediately by switching to an adjacent scenario and noticing the difficulty return, which is the healthy signal. And you have stopped being surprised, real patients surprise constantly, and a practice regime that never does has drifted from rehearsal into replay. The repair is always the same: perturb, new case, mutated parameters, or manual variation, choose the version of the complaint you have not met, and let the difficulty come back, because the difficulty is the practice. Repetition remains the gift; the discipline is repeating the skill, not the scenario, and platforms plus habits that enforce that distinction convert unlimited practice from hazard back into the access revolution it should be.
Frequently asked questions
How many runs of one scenario are useful before mutation?
Two or three: the first for the encounter, the second for structure with feedback applied, a third at most for fluency; beyond that, variation or a new case, and the itch to re-run for a higher score is precisely the signal to move.
Does this problem apply to question banks too?
Directly, as second-pass score inflation: repeated items reward recognition, which is why unseen-item performance is the only percentage worth tracking, the calibration argument made at /blog/why-qbank-percentages-are-not-comparable.
Can platforms fix this without harming beginners?
Yes, by staging: stable scenarios early, when structure is being built, mutation increasing with demonstrated competence; difficulty that grows with the learner is the design, and vendors describing exactly that staging deserve the benefit of a trial.
Is deliberate re-running ever the right tool?
For skill isolation, yes: repeating a case specifically to fix one behaviour, the safety-netting close, the drug history, is targeted practice; the discipline is naming the isolated skill beforehand, which script-drift never does.
Do OSCE examiners see script-learning in candidates?
Its signature, yes: fluent openings that survive contact with an unexpected answer poorly; mutation-trained candidates recover because recovery is what they practised, which is the transfer argument in one observable behaviour.