This 15-page preprint asks whether design preferences generated by LLM simulations match responses from real users. Its main contribution is comparison against material from professional practice rather than a purely artificial benchmark. It aggregates 29 preference-testing studies created by 29 organizations in UXtweak: 2,073 participants, 78 choice tasks, and 190 follow-up questions, 147 open and 43 closed, covering interfaces, components, layouts, navigation, notifications, illustrations, and copy across several domains. The strongest operational conclusion is that a synthetic sample should not replace a real audience without task-specific validation.
The system reproduces each study flow: it receives original messages and images, adopts a participant role, and answers preference tasks and follow-ups. The baseline uses GPT-4.1, temperature 1, top_p 1, and a mega-persona summarizing audience size, demographics, personality, and other available information. When such data are missing, the authors add general-population gender, age, education, and Big Five distributions. Mega-persona responses are generated in batches of 20 and merged. Five one-factor variants substitute GPT-5.2, temperature 0.2, top_p 0.2, individual personas, or a nondescript mega-persona. Fourteen studies were used for iterative prompt refinement and then remain in the reported evaluation; there is no separate holdout.
Choice metrics include per-task chi-squared difference, first-choice and ranking agreement, Jensen-Shannon distance, normalized entropy, and unique selection count. Text metrics include TF-IDF lexical similarity, all-MiniLM-L6-v2 semantic similarity and diversity, Yule's K, readability, and length, followed by qualitative examination. The baseline differs significantly from humans in 44% of 78 tasks, matches the most popular option in 53%, has mean rank agreement 0.53 and mean Jensen-Shannon distance 0.17. Synthetic entropy is higher, 0.93 versus 0.86, consistent with spreading support too evenly when humans have clearer favorites. Reverse cases also occur, with divided humans and deterministic simulations.
GPT-5.2 reduces significantly different tasks to 38% and reaches 65% first-choice agreement, but the paper does not find a statistically significant change from GPT-4.1. Lower temperature produces 41% different tasks and lower top_p 38%, with entropy around 0.94. Generic and detailed mega-personas behave similarly: generic personas differ in 46% of tasks. Individual personas are clearly worse: 91% different tasks, Jensen-Shannon distance 0.45, entropy 0.28, and 1.82 unique choices versus 2.59 for mega-personas. This shows that simulation structure matters and contradicts a reading of identical failure under every configuration.
For open justifications, simulations have mean lexical similarity near 0.25 and semantic similarity around 0.70 to human responses. The authors describe genericity, overpraise, fixation on isolated elements, lengthy explanations without depth, irrelevant inferences, and some outputs they deem nonsensical. Closed-ended answers differ in 53% of questions. The warning is plausible, but the qualitative component provides no codebook, coder count, double coding, agreement, sampling rule, saturation criterion, or auditable quotations. Images, responses, and study contents are proprietary, so readers cannot determine whether an interpretation was truly incoherent or compare it with the quality of human explanations.
Statistics constrain several conclusions. The same 78 tasks recur under every configuration, yet binary indicators are compared through chi-squared tests on 156 observations and the missing appendix is described as using Mann-Whitney for continuous measures. These are paired comparisons, calling for methods such as McNemar and paired Wilcoxon, plus a multilevel model because 78 tasks are nested in 29 studies and may share participants, prompts, and stimulus families. Seventy-eight goodness-of-fit tests per configuration receive no multiplicity correction or expected-cell diagnostics. Percent significant therefore mixes effect magnitude, sample size, option count, and false positives. Interpreting p>0.05 as proof that model, temperature, top_p, or specificity have no effect would require equivalence tests, power, and intervals that are not reported.
Generative uncertainty is also omitted. The baseline was reportedly run three times with little variation, but values, seeds, and tests are absent; each reported configuration then uses one run. Individual and mega-persona conditions also differ structurally: one conversation per person versus batches of 20 whose members may condition one another. This confounds persona type with context, length, and response independence. Only GPT-4.1 and GPT-5.2 are tested, without a dated snapshot, endpoint, GPT-5.2 reasoning effort, seeds, image settings, budget, retry policy, or failure handling. The study therefore does not show that chain-of-thought generally fails and does not cover multimodal LLM diversity.
Real studies increase ecological relevance but do not guarantee external validity. The paper withholds the sampling frame, criteria, dates, organization types, refusals, countries, languages, per-task n, participant overlap, quality checks, and distributions. All studies come from consenting customers of one commercial platform. Heterogeneity broadens scenarios while also adding uncontrolled recruitment, wording, order, stimulus-quality, and question differences. Calling human responses “ground truth” is reasonable for their own task, not as universal human preference. “Universal failure” is also too strong: 53% match the winner and 56% of baseline tasks show no significant distribution difference.
There is an institutional-transparency issue. All three authors list UXtweak Research / UXtweak j.s.a.; the company supplies proprietary studies and is thanked for technical and expert support. Yet the declaration says there are no relevant financial or non-financial interests. The affiliation is visible, but employment, data-access control, and product implications are relevant interests that should be characterized alongside roles in selection, analysis, and publication. The ethics statement provides no committee, approval number, or secondary-use basis, and does not distinguish study-owner consent from consent by each participant.
The results are not reproducible. The final prompt is assigned to Appendix A and complete analyses to Appendix B, but the PDF and complete TeX end at the references: neither appendix exists. There are no task-level aggregates, synthetic outputs, prompts, code, statistical scripts, API logs, or environment. GPT-4.1 and GPT-5.2 references list an impossible 2024-03-24 access date for a GPT-5.2 page published in December 2025. The faithful conclusion is two-sided: the paper offers important evidence that GPT synthetic samples can miss concrete visual preferences and create a false impression of balance; it does not prove universal failure, moderator equivalence, guaranteed high external validity, or a general theory that LLMs lack reasoning. The correct recommendation is not to replace user research with unvalidated simulation, while leaving room to evaluate calibrated, human-supervised auxiliary uses.