{"id":"e8665f68-c260-4c52-87bb-576c07a4012e","arxiv_id":"2501.06964","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"GPT-4 personas aligned with real patient answers 54.97% on average versus 26.7% random, but only when primed with education, and the reported 88% accuracy overstates what was measured.","lead":"This paper tests whether GPT-4 can impersonate patients with different backgrounds when answering questions about hospital discharge instructions, comparing the model's answers to those of 96 real survey participants. It matters because if LLM personas reliably mirror patient comprehension, they could cheaply pretest health materials, but the results show accuracy varies sharply by which personal trait is used.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The reported 'alignment rate' is undefined; without a no-persona control and a skew-aware metric, the 88% value and Table 2 comparisons do not establish persona simulation.","rationale":"The reader's weakest_assumption correctly identifies that the alignment metric is undefined and uncontrolled for answer skew; my independent reading of Table 2, Figure 4, and the abstract confirms this. The concern is load-bearing because every positive finding in the paper—especially the 88% headline and Finding 2's 77.50% average—is a number from this undefined metric. The internal contradiction with the abstract's 'below random chance' claim strengthens the case that the reported figures are not reliable. I agree with the CONDITIONAL verdict: the experimental questions are worth asking, but the paper must define the metric, add a no-persona control and a majority-class baseline, provide statistical uncertainty, and release the questionnaire and analysis code. No ad hominem is warranted; the issues are methodological and addressable. If the metric is clarified and the baselines confirm that the persona prompts add information beyond generic answer tendencies, the central claim could stand.","tokens_in":18571,"tokens_out":2772,"duration_ms":29196,"concrete_test":"Recompute Table 2 with an explicitly defined metric: for each question, compare the LLM response to the modal human answer of the target group, and report the same comparison for (a) a majority-class baseline that always chooses the modal human answer, (b) a no-persona control prompt identical but without the 'If you were...' prefix, and (c) a skew-aware agreement measure such as per-question Cohen's kappa or Brier score. Run at least 10 LLM samples per condition to estimate variance. If the persona-conditioned score does not significantly exceed both the majority-class and no-persona baselines, the 88% claim should be withdrawn.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that education-primed LLMs deliver 'accurate and actionable medical guidance 88% of the time' depends entirely on an 'alignment rate' that is never formally defined. Section 4.2 and Table 2 report similarity between LLM and human group responses, but the paper does not state whether the comparison is to the modal human answer, the full response distribution, or something else. If it is modal-answer matching, the High Edu. (0.88) and Male (0.97) scores can be explained by the LLM defaulting to the most common answer on questions where humans also cluster (e.g., 'A. Yes' on Q2-Q4), as Figure 4 illustrates: the LLM never selects 'I don't know' and concentrates on two options. The random guess baseline (0.278 for information tasks) does not correct for skewed answer distributions; a majority-class baseline would be far stronger. The abstract's additional claim that performance 'falls below random chance levels' when other information is provided is contradicted by Table 2, where every persona row exceeds the random guess baseline (e.g., Low ER_visit information rate 0.31 > 0.278). No confidence intervals, significance tests, or sampling details are provided, so the 88% headline is unsupported as stated. This concern is load-bearing because if the metric is merely modal-answer agreement, the paper's core conclusion about persona simulation collapses into a statement about the LLM's generic answer tendencies, a possibility the paper itself raises in Section 5 when noting that LLM responses are 'inherently identical or leptokurtic.'","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper evaluates whether GPT-4 can simulate the behavior of patient groups when answering questions about ICU discharge summaries. The authors recruited 96 Prolific participants, collected answers to eight information-based and two perception-based questions across four discharge summaries, and compared these with GPT-4 responses generated under personas defined by education, gender, doctor-visit frequency, and ER-visit frequency. They report alignment rates per persona (Table 2) and use them to conclude that education-primed personas achieve high alignment (88% for high-education information tasks), that performance drops for female and low-ER-visit personas, and that adding other information can reduce performance below random guessing. The paper frames this as a first step toward using LLMs as proxies for human subjects in personalized health communication.","tokens_in":18856,"tokens_out":6434,"duration_ms":55006,"significance":"If the quantitative claims were well supported, this would be a useful preliminary contribution to the growing literature on LLM role-playing and to debates about replacing human participants with LLM simulations. The external human-response data, IRB oversight, randomized discharge summaries, and transparent reporting of the per-persona percentages are genuine strengths, as is the honest qualitative discussion of overconfidence (Section 4.5) and of the risk of 'misportray[ing]' identity groups (Section 5). However, the paper's central quantitative claims currently rest on an undefined similarity metric, a weak random-guess baseline, and the absence of any no-persona control; until those are addressed, the alignment percentages cannot be interpreted as evidence for persona-simulation capability.","major_comments":[{"comment":"The paper never defines the 'similarity'/'alignment rate' between the LLM response and human responses. It is not stated whether the comparison is to the modal human answer for each question, to the full human response distribution (e.g., via a distributional metric), or to a per-question pooled statistic. Without this definition, the headline numbers (0.88, 0.97, 0.31) and the aggregate 'average alignment rate of 54.97%' are uninterpretable, and the central claim that the LLM 'simulates' these personas is not falsifiable from the table.","section":"§4.2, Table 2"},{"comment":"The abstract claims that education-primed LLMs 'deliver accurate and actionable medical guidance 88% of the time,' but Table 2 shows 0.88 under 'High Edu.' for the information-based task, which is an alignment rate with human answers, not a measure of medical accuracy or actionability. Moreover, the abstract's claim that performance 'falls below random chance levels' is not generally true: in the information tasks every persona row is above the reported random-guess value (0.278), and in the perception tasks only the Male persona (0.25) falls below 0.267. The wording must be corrected to describe alignment with modal human responses and to give the precise exception.","section":"Abstract and §4.2"},{"comment":"The random-guess baseline is not a meaningful control for skewed answer distributions. The reported baseline assumes uniform random choice among the answer options, but the human responses are highly skewed (e.g., Figure 4 shows a strong majority choosing 'A. Yes' for Q2-Q4). A majority-class baseline or a chance level conditional on the observed marginal answer distribution would be far higher than 0.278 for such questions. The paper also lacks a no-persona control condition: without asking the same questions without persona priming, or with a neutral persona, the high alignment of, say, the Male persona (0.97) may simply reflect the LLM's default tendency to choose common answers, a possibility the paper itself acknowledges in Section 5 when it notes that LLM responses are 'inherently identical or leptokurtic.'","section":"§4.2"},{"comment":"No confidence intervals, significance tests, or sampling details are provided for the LLM responses. Section 4.1 states only that GPT4_0613 was used via the Azure OpenAI API; the number of repeated samples per prompt, temperature, and number of questions per condition are not given. Section 4.2 uses the phrase 'significantly higher alignment rate' without any statistical test, and the differences between personas (e.g., High Edu. 0.88 vs. Low Edu. 0.72) are single point estimates from 96 human respondents split into multiple subgroups. These omissions make it impossible to judge whether any of the reported differences are robust.","section":"§4.1 and §4.2"},{"comment":"The claim that 'a straightforward query-response model could outperform a more tailored approach' is not supported by any experiment reported in the paper. The evaluation only compares persona-primed prompts to random guessing; there is no non-persona or 'straightforward' baseline in Table 2 or elsewhere. If the authors have such data, it should be reported; otherwise this conclusion should be removed or reframed as a hypothesis.","section":"Abstract and §7"}],"minor_comments":[{"comment":"The terms 'LMM' and 'LMMs' appear several times (e.g., 'failure modes of LMM'); these should be 'LLM'/'LLMs'.","section":"Section 5"},{"comment":"Q5 is labeled 'Do you know your diagnosis?' but the same label is used for Q3; the question text appears to be about other prescriptions, so the label is likely a copy-paste error.","section":"Appendix C.1"},{"comment":"Q8's option A reads 'A void fruit' and should read 'Avoid fruit'.","section":"Appendix C.1"},{"comment":"The terms 'similarity' and 'alignment rate' are used interchangeably; they should be unified and formally defined.","section":"Table 2 and §4.2"},{"comment":"The sentence 'Significant gender biases in LLM-generated content have been analyzed by Wang et al. in [47]' would benefit from a brief summary of the relevant finding, since the reference is to a general preprint.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The paper has a useful qualitative core and an interesting human dataset, but the quantitative claims need substantial revision. I recommend major revision rather than rejection because the missing metric and controls are fixable within the scope of the manuscript. The abstract overstates the results, which may be a framing issue; the authors should be asked to reconcile the abstract with Table 2 and to add the missing operational definitions and baselines."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper tests whether GPT-4 with persona prompts can reproduce how ICU discharge summaries are understood across demographic groups, comparing against 96 Prolific respondents. The real value is the design: external human responses as the reference, four discharge summary types, and a finding that clearly survives the paper's own limitations — that the LLM overestimates comprehension of long summaries. The education-level alignment pattern (77.5% average) is also interesting, though less solid than the authors claim.\n\nWhat it does well: comparing to real human answers is the right way to ask this question, and the random guess baseline is a reasonable floor. The paper is also candid in its broader-impact section, citing prior work on identity misportrayal and admitting that LLM responses are inherently leptokurtic. That self-awareness is not just decoration; it partially preempts the strongest objection.\n\nThe soft spots are real and load-bearing. The central metric in Table 2 is labeled 'similarity' but never defined. From context it appears to be per-question matching to the modal human answer, but the paper never says so. If that is the metric, the high alignment for High Edu. and Male is partly an artifact of answer skew — humans cluster on 'Yes' and 'Not provided', and the LLM never picks 'I don't know.' The random guess baseline of 0.278 does not correct for skewed marginals; a majority-class baseline would be far more informative. The abstract's claim that education-primed LLMs deliver 'accurate and actionable medical guidance 88% of the time' mischaracterizes the result: it is alignment with human answers, not medical correctness. And the claim that performance 'falls below random chance' is contradicted by Table 2, where every information row exceeds 0.278. No confidence intervals or significance tests are reported, and there is no no-persona control, so we cannot tell how much the persona prompt adds over the model's default answer distribution.\n\nThe stress-test note is correct on all these points. The undefined metric is not a minor omission; it is the hinge of the paper's main conclusion. That said, the flaws are addressable. Define the metric, add a majority-class baseline and a no-persona condition, report error bars, and correct the abstract, and the core comparison between LLM personas and human responders can stand as a pilot.\n\nWho is this for? Researchers interested in LLM simulation of human samples, and medical NLP folks thinking about low-cost patient comprehension testing. It will not change practice, but it is an honest, well-scoped pilot worth a serious referee. My recommendation: send it to review, with heavy revision expected. I would not cite it in my own work in the next year, but I would keep it on file.","headline":"A useful pilot on LLM persona simulation for health communication, but the headline 88% claim is not supported by the undefined alignment metric.","tokens_in":19422,"tokens_out":1603,"would_cite":false,"duration_ms":16771,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Persona-prompted GPT-4 matches high-education patients on discharge-summary questions, but not female or low-visit patients.","keywords":["large language models","role-playing","persona simulation","discharge summaries","patient comprehension","health communication","GPT-4","alignment evaluation"],"falsifier":"Re-score the survey using full answer distributions rather than modal agreement: for each of the ten questions, record the LLM's letter distribution under the \"high education\" prompt, the \"low education\" prompt, and a no-persona prompt, and compare them with the human group distributions. The claimed persona effect is falsified if the LLM's distribution barely moves across personas on questions where human groups differ, or if the no-persona prompt matches a group as well as that group's persona prompt does; the easiest check is Q2 and Q3, where most humans answer \"A. Yes\" and the model might do the same regardless of persona.","tokens_in":18390,"feed_emoji":"🩺","tokens_out":6992,"duration_ms":66250,"temperature":0.7,"pith_summary":"This paper asks whether a large language model can stand in for real patients when testing whether discharge summaries are understandable. The authors prompt GPT-4 with personas such as \"someone who has never received a college degree\" or \"someone whose gender is female,\" ask it the same ten comprehension questions given to 96 human participants, and compare answer patterns. They find that education-primed personas align with human answers at an average of 77.5%, with high-education information questions reaching 88%, while female, low-doctor-visit, and low-ER-visit personas align at 31–47%, near random chance. The point is to decide whether LLM role-play can cheaply replace user studies for patient-centered health communication, and the answer is: for some groups yes, for others no, and adding demographic context can make things worse.","feed_headline":"LLM personas match high-education patients 88% of the time","feed_subtitle":"Female and low-ER-visit personas fall to 31–44%, testing whether AI can stand in for patient studies.","key_machinery":"The intervention is in-context impersonation: the authors prepend \"If you were a {persona}\" to a fixed task instruction, and the model answers multiple-choice questions by letter only. Personas are built from four demographic attributes (education, gender, doctor-visit frequency, ER-visit frequency) split into high and low. The evaluation machinery is the reported alignment rate, which compares the persona-prompted LLM's letter choice with the dominant answer of the corresponding human group on the same questions, reported separately for eight information-based questions and two perception-based questions, with a random-guess baseline for comparison.","core_discovery":"The authors set out to establish that LLM-driven personas can reproduce how real patients answer questions about discharge summaries, and to map where that reproduction fails. Their central quantitative finding is that the model's alignment with human answers depends strongly on the persona and the task type: education-based personas average 77.5% alignment, high-education information questions hit 88%, and male information questions hit 97%, while female, low-doctor-visit, and low-ER-visit personas fall to 31–47%. The paper also finds that the model oversimplifies: it concentrates answers on one or two options, never chooses \"I don't know,\" and overestimates comprehension of long summaries. The authors read these results as evidence that in-context impersonation is promising for self-similar groups but not ready for diverse patient populations, and that adding patient-specific context can reduce rather than improve accuracy.","pith_inferences":["The alignment metric may be rewarding base-rate matching: on questions where almost all humans choose \"A. Yes,\" a model that prefers \"A\" scores high for every persona; comparing full answer distributions would separate persona-specific simulation from default answer tendencies.","A direct test of the mechanism would be to run the same prompts with no persona and with several personas; if the no-persona distribution matches high-education humans as well as the high-education persona does, the persona prefix is not doing the work.","The failure to ever answer \"I don't know\" is a plausible cause of both the inflated male and high-education scores and the depressed female and low-visit scores; adding a calibration step that permits uncertainty may change which groups look well simulated."],"forward_implications":["Education-primed LLM personas could serve as a low-cost first-pass screening tool for how well-educated patients understand discharge instructions.","Automated patient communication systems should not yet use LLM personas to represent female, low doctor-visit, or low ER-visit populations, because alignment there is near random.","Including demographic details beyond education in a prompt can hurt performance, so a simpler query-response format may be safer for generating patient-facing health text.","Because the model overestimates comprehension of long summaries and never expresses uncertainty, any LLM-generated discharge material needs human review before release."],"supporting_citations":[{"why":"Supplies the in-context learning mechanism that the persona-prompting method relies on.","marker":"[10]"},{"why":"Documents strengths and biases of in-context impersonation, the technique this paper applies to patient personas.","marker":"[38]"},{"why":"Provides role-play prompting as a zero-shot reasoning method that motivates the \"If you were a {persona}\" prompts.","marker":"[24]"},{"why":"Argues LLMs cannot portray identity groups, the limitation this paper's failure cases engage directly.","marker":"[48]"},{"why":"Shows an LLM discharge-summary generation system whose evaluation did not compare against human subjects, which this study does.","marker":"[52]"},{"why":"Analyzes gender biases in LLM-generated content, cited to explain the weak female-persona alignment.","marker":"[47]"}],"fun_headline_variants":["LLM patient personas: 88% accurate for some, 31% for others","AI patient sims: 97% for men, fall below chance with extra info","LLM role-play drops to random chance with extra patient info","Patient simulation: 88% for educated, 31% for others — not ready"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole evaluation assumes that agreeing with the most popular answer in a human group demonstrates that the model is simulating that group's perspective; if the agreement comes from the model's generic preference for common answers, the persona scores do not measure what they claim to.","fun_headline_variants_meta":{"raw":{"variants":["LLM patient personas: 88% accurate for some, 31% for others","AI patient sims: 97% for men, fall below chance with extra info","LLM role-play drops to random chance with extra patient info","Patient simulation: 88% for educated, 31% for others — not ready"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000716,"raw_usage":{"total_tokens":3249,"prompt_tokens":1005,"completion_tokens":2244,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":621,"completion_tokens_details":{"reasoning_tokens":2158}},"tokens_in":621,"tokens_out":2244,"duration_ms":16786,"temperature":1.0,"reasoning_tokens":2158,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T20:49:19.490462+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-score the survey using full answer distributions rather than modal agreement: for each of the ten questions, record the LLM's letter distribution under the \"high education\" prompt, the \"low education\" prompt, and a no-persona prompt, and compare them with the human group distributions. The claimed persona effect is falsified if the LLM's distribution barely moves across personas on questions where human groups differ, or if the no-persona prompt matches a group as well as that group's persona prompt does; the easiest check is Q2 and Q3, where most humans answer \"A. Yes\" and the model might do the same regardless of persona.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Documents strengths and biases of in-context impersonation, the technique this paper applies to patient personas."},{"cited_title":"Discharge Me!","cited_arxiv_id":null,"evidence_quote":"Shows an LLM discharge-summary generation system whose evaluation did not compare against human subjects, which this study does."}],"review_version":1}