Short answer: A 2026 Nanyang Technological University study tested asking an LLM to answer a whole questionnaire in one prompt ("fast" silicon sampling). The synthetic survey data was at least as faithful as the usual one-question-at-a-time method, at about half the cost. Both methods still understated human opinion variance badly, keeping only about a quarter of real dispersion.
Last updated: October 27, 2026
Silicon sampling means prompting a large language model with a respondent's demographic profile and asking it to answer survey questions as that person would. Most studies do it slowly: one question (or a small block) per API call, with the session reset between calls. The preprint The conditional superiority of fast silicon sampling by Nickolas Hock Yuen Lam, Ji Xuan Voo and Xiangyu Ma (arXiv, August 2026) tests a practical question no earlier study had compared directly: does sending the whole questionnaire at once make the synthetic answers worse?
What is the difference between fast and slow silicon sampling?
Fast sampling puts every survey item into one large prompt and gets one complete response set back. Slow sampling splits the questionnaire into small units and makes a fresh API call for each. The authors note that slow sampling is the field's default. They hand-coded 75 arXiv preprints from 2023 to mid-2026 that use silicon sampling in a substantive way. Of these, 66 used slow sampling and 17 used fast sampling (eight used both).
Why do researchers default to slow? The authors infer three reasons. The first is precedent from the original Argyle et al. (2023) protocol. The second is that older models handled long prompts poorly. The third is a wish to avoid question-order effects by isolating each item. They also point out that human respondents answer surveys the "fast" way, carrying memory from one question to the next.
What did the study test?
The study built two synthetic versions of the Singapore sample of the World Values Survey Wave 7 (n = 2,012), one with each method, and compared both against the real respondents. According to Lam, Voo and Ma (2026), the setup was:
- Model: OpenAI GPT-5.4 (March 5, 2026 weights), temperature 1.0, up to 2,500 output tokens.
- Personas: each real respondent's profile was rebuilt from 14 demographic variables, including age, gender, citizenship, home language, education, religion and ethnicity. An introduction adapted from the WVS interviewer script came first.
- Items: 27 questions in two batteries. Eight cover the perceived effects of immigration and 19 ask whether ethical behaviors are justifiable.
- Scale: 30 synthetic surrogates per real respondent in each mode, or 60,360 synthetic response sets per mode.
- Fidelity checks: population means, respondent-level answer overlap (Jaccard distance), variance, pairwise associations (Cramer's V), and the overall "shape" of opinion space. That shape was mapped with multiple correspondence analysis and compared with Procrustes statistics.
Is fast silicon sampling cheaper?
Yes, substantially. Each synthetic respondent took about 20 seconds and 4,000 tokens with slow sampling, against 3 seconds and 1,450 tokens with fast sampling. For the full run, the authors estimate 335.3 hours and $603.60 for slow sampling, compared with 50.3 hours and $301.80 for fast.
| Measure | Slow sampling | Fast sampling |
|---|---|---|
| Runtime per respondent | ~20 s | ~3 s |
| Tokens per respondent | ~4,000 | ~1,450 |
| Estimated total cost | $603.60 | $301.80 |
| Mean standardized bias (27 items) | 0.40 SD | 0.37 SD |
| Real-world dispersion retained | 23.9% | 22.9% |
| Association match with WVS, immigration (r) | 0.28 | 0.42 |
| Association match with WVS, ethics (r) | 0.57 | 0.69 |
| Procrustes M², immigration (0 = identical) | 0.610 | 0.468 |
Source: Lam, Voo & Ma (2026), Tables 1, 2, 5 and 6 and accompanying text.
Is fast silicon sampling less accurate?
No. Fast sampling came as close to the human data as slow sampling on almost every metric, and was often slightly closer. Mean standardized bias across the 27 items was 0.37 SD for fast and 0.40 SD for slow. In either mode, a silicon respondent's answers overlapped with their real counterpart's only around 40% of the time, though fast sampling was marginally closer on both batteries. Fast sampling also reproduced the correlations between items better, for both immigration and ethics.
Fast sampling did not clearly win on the ethics opinion space. There the two modes landed within a few hundredths of each other (M² of 0.744 slow and 0.756 fast).
How accurate are silicon samples overall?
They are moderately accurate on averages and poor on everything underneath. This part of the paper applies to both methods, so researchers should weigh it most heavily:
- Variance collapses. Synthetic samples showed less dispersion than humans on every one of the 27 items. On average they kept only 24% (slow) and 23% (fast) of the real standard deviation. Seven ethics items and one immigration item had zero variance, meaning every synthetic respondent gave the same answer.
- Immigration is harder than ethics. Under slow sampling, immigration attitudes carried roughly twice the standardized bias of moral attitudes (0.60 vs 0.31 SD).
- The structure of opinion is distorted. Procrustes M² values ran from 0.468 to 0.756. A bootstrap resample of the real WVS data scored 0.0004 or lower, so the synthetic misfit is about three orders of magnitude larger than sampling noise.
The authors note that models have less training data on Singapore than on Anglo-European survey populations. They conclude that silicon samples there should be treated as exploratory tools at best. They are useful for generating hypotheses or stress-testing question wording, but they should not replace respondents.
Is this the same as batching many respondents in one prompt?
No, and the distinction matters. This study put all questions for one persona in a single prompt. A different 2026 study we covered found that generating many synthetic respondents in one request inflates correlations. Taken together, the two studies suggest a reasonable default: one persona per call, with the full questionnaire in that call. Packing several personas into one call is riskier.
What this means for researchers
If you already use synthetic respondents to pilot a survey, the cheaper single-prompt design looks safe to adopt in place of the slow default. It does not fix the core limits. Practical takeaways:
- Check variance, not just means. A synthetic mean can sit close to the human one while the distribution has only a quarter of the real spread.
- Validate structure. Before drawing conclusions about how attitudes relate, compare correlations between items, or the overall configuration, against a real benchmark.
- Use the savings for real data. Cutting generation cost in half is most valuable if the money goes to human responses that confirm what the simulation suggests.
The paper supports an exploratory role for synthetic panels. For teams testing questionnaires before fielding them, the Qualitati Digital Twin Panel fits that role, and AI Surveys handle the confirmatory step with real respondents.
Frequently asked questions
What is silicon sampling?
Silicon sampling means prompting a large language model with a person's demographic profile and collecting its survey answers as a synthetic stand-in for that person. The term comes from Argyle et al. (2023).
Does giving an LLM every survey question in one prompt make the answers less accurate?
Not in this study. The single-prompt approach matched or beat the one-question-per-call approach on means, respondent-level overlap and associations between items, and it cost about half as much.
Can silicon samples replace human survey respondents?
Not on this evidence. Both methods kept less than a quarter of real opinion variance and distorted how attitudes relate to each other. The authors therefore recommend exploratory use only.
Which model and data did the study use?
The study used OpenAI GPT-5.4 and the Singapore sample of the World Values Survey Wave 7 (2,012 respondents), with 27 items on immigration and ethical norms.
Last updated: October 27, 2026. This article is an independent editorial summary of third-party research by Lam, Voo and Ma (2026). It is not affiliated with or endorsed by the authors. See the original preprint for full methods and results.