Informatic Displacement When AI Replaces Simulated Patients

1. Study information 1.1 Title Informatic displacement when AI replaces simulated patients: a literature-parameterized simulation and document analysis. 1.2 Background Medical schools increasingly offer AI-simulated patients alongside, or instead of, human simulated participants (SPs). The institutions that make this choice (here called operators) typically compare visible costs such as SP wages and license fees. We propose that the substitution also changes where an encounter's information goes, in three ways that operators do not usually count: Return flow. Human SPs report gaining health knowledge from their work and passing some of it to family and friends (Plaksin et al. 2016; Thompson et al. 2026). The evidence is mostly self-report from small samples, so its size is uncertain. Capture. Encounters with an AI patient may generate data that a vendor retains or uses for training, depending on its terms. Channel change. Learners may receive less nonverbal information from an AI patient than from a person, depending on modality. Benefits that are not valued are, in effect, given zero weight in a decision (National Research Council 2005). This study builds a transparent model of these flows so that an operator can see what substitution may cost. All outputs are properties of the model and its stated assumptions, not measurements. 1.3 Research questions What information flows does a human SP encounter generate, and which may be displaced, captured or substituted when AI replaces the SP? Under published parameter ranges, how large might the community story reach of an SP program be, and how does it change across substitution scenarios? What do publicly available terms of AI patient platforms permit regarding retention and training use of encounter data? How might an operator compare investing in AI simulation with investing in human SPs once displaced flows are counted, and which uncertain parameters most affect that comparison? 1.4 Expectations This is an estimation and sensitivity study, not a hypothesis test, and no p-values will be computed. We state four expectations in advance so that readers can check them against the results. Each names the output, the criterion and how a result is classified. Checks that hold by construction are listed separately as verification tests (§5.4), not as expectations. Expectation Output and criterion Classification E1 First-generation sharing and confidentiality drive the size of what is at stake. Output: baseline fidelity-weighted reach per program-year, F(0), which equals the reach lost under full substitution. Criterion: the sharing group (p1, n1 or their reported product, κ) has the largest group total-order Sobol index, and at least two of the three inputs with the highest individual total-order index belong to that group, and PAWN ranks at least two of the same inputs in its top three. Supported if all three conditions hold and the sharing group's bootstrap 95% interval does not overlap the next group's. Inconclusive if the conditions hold but the intervals overlap. Not supported otherwise. E2 Community-reach outputs carry the least certainty. Output: GRADE certainty ratings after the cap rule (§5.6). Criterion: F(0) and DR(s) are rated low or very low, and at least one of platform capture or the change in checklist-scored practice share is rated moderate or higher. Supported or not supported, with each rating and the parameter that set any cap reported. E3 More AI substitution increases learners' exposure to item-level checklist feedback. Output: change in checklist-scored practice share per learner from s = 0 to s = 1. Criterion: the 90% uncertainty interval excludes zero. Supported, not supported or inconclusive (interval includes zero). If the input ranges for AI and human checklist shares do not overlap, the direction follows from the inputs alone; we will report it as a consequence of the inputs, not as a finding. E4 Substituting formative practice costs less community reach per dollar saved than substituting everything. Output: DR for Hybrid A (AI for formative practice only) and for All AI, computed in each uncertainty run. Criterion: DR(Hybrid A) > DR(All AI) in more than half of the runs. Rationale (provisional): repeated formative encounters are likely less intense and less novel, and intensity and novelty predict sharing (Luminet et al. 2000). Supported if the share of runs is above 0.5 and its bootstrap 95% interval lies above 0.5. Inconclusive if the interval includes 0.5. Not supported if the interval lies below 0.5. This registration is deposited before any pedigree scoring, dossier completion, expert elicitation or model run. Its purpose is to fix the rules for sourcing, scoring and converting evidence, and for analyzing and reporting results, before anyone sees how those rules affect the results. What had been done before deposit, disclosed in full: The literature was identified, read and annotated (76 annotated entries). While reading, we noted candidate values for several parameters, for example the sharing rates and reach anchor in Harber & Cohen 2005 and Curci & Bellelli 2004, retention per retelling and drift in Kashima 2000, unbiased hit rates by channel in Paulmann & Pell 2011, and the change in core discussion networks in McPherson et al. 2006. These are candidates only. None has been scored, synthesized or converted into a distribution, and the final values may differ. No model code has been written and no model has been run. No platform documents have been collected or coded. Because candidate values were seen before deposit, this registration cannot be fully blind. Its protection lies in fixing the conversion and analysis rules (§5) so that the final values and conclusions follow mechanically from the evidence and the elicitation record.