Does asking in Greek change how AI "silicon samples" answer as Greek respondents?

Research question

When a language model answers a few European Social Survey Round 11 questions as Greek respondents of a given age, sex, education and region, do its answers come closer to or move further from those of the Greek respondents when the prompt is in Greek rather than in English, overall and across groups?

Why it matters

The study that named silicon samples found close agreement for US vote choice when the model was given respondents' own partisan backstories. LD-07 also records that matching averages can hide failures on variance and subgroups, that studies against Chinese survey data found only partial reproduction, and that proprietary models change over time. Greece is already covered. A 2026 cross-national study reports ESS results for Greece from English-language prompts, and a 2024 study that simulated 548 Greek participants in a populism framing experiment found that prompting in Greek greatly reduced the simulation's fidelity. No study SocioLens found tests whether prompting in Greek rather than English changes how well a model reproduces Greek answers to survey attitude items such as those in the ESS. A small, documented comparison of the two prompt languages on such items shows whether that finding carries over from an experiment to a survey, and is an undergraduate-sized, single-country slice of RO-LD07-08.

Evidence that motivates it

1. In the study that introduced "silicon samples", GPT-3 conditioned on the demographic and political backstories of real US survey respondents reproduced their two-party presidential vote closely (tetrachoric correlations of 0.90, 0.92 and 0.94 for 2012, 2016 and 2020) and the pattern of associations among 2016 survey answers; the authors called this "algorithmic fidelity" and said researchers can use such samples once it has been established for a given model and domain. (LD07-SL-C01 v1.0)
   GPT-3 given the backstories of US survey respondents reproduced their vote choice closely, and the authors said researchers can use such samples once this fidelity has been established for a given model and domain.
   Cites: Argyle et al. (2022), “Out of One, Many: Using Language…”
2. Synthetic respondents can match selected survey averages while failing to reproduce real-world variance, regression relationships, individual structure and subgroup heterogeneity. (LD07-HQ-C06 v1.0)
   Synthetic respondents can match selected survey averages while failing to reproduce variance, regression relationships and subgroup differences.
   Cites: Bisbee et al. (2024), “Synthetic Replacements for Human…”; Xie et al. (2026), “Evaluating the statistical realism of…”; Wang et al. NMI 2025; Gong Weigang & Huang Siyuan (2025), “整体事实还是有偏样本——基于大语言模型生成…”; Ling Wanying & Cui Sizhan (2025), “复现的限度——‘硅基样本’的可能与可为”
3. Chinese-language silicon-sample studies benchmarked against CGSS data find partial macro-level reproduction alongside representational bias, micro/macro/meso instability and cross-model differences; current systems do not warrant replacement of human respondents. (LD07-HQ-C08 v1.0)
   Chinese-language silicon-sample studies benchmarked against national survey data find partial reproduction alongside representational bias and differences between models.
   Cites: Gong Weigang & Huang Siyuan (2025), “整体事实还是有偏样本——基于大语言模型生成…”; Ling Wanying & Cui Sizhan (2025), “复现的限度——‘硅基样本’的可能与可为”
4. Model identity is time-indexed: proprietary services can change materially across deployed snapshots, so exact version and date are part of the measurement protocol. (LD07-HQ-C12 v1.0)
   Proprietary models change across deployed snapshots, so the model version and date are part of the measurement.
   Cites: Chen et al. (2024), “How is ChatGPT’s behavior changing…”

Claims are shown as they read on the day this pack was built; claim pages keep every earlier version.

A possible design

Choose five to eight ESS Round 11 attitude items with clear response scales (for example trust in parliament, satisfaction with democracy or views on immigration). Draw a random sample of about 300 Greek respondents and build a short profile for each from their age, sex, education and region only, never from their attitudes. Ask one openly available model, with its version and the date recorded, to answer all the chosen items in a single prompt per profile, in a fixed order, once with the prompt in Greek and once in English (about 600 queries in all). Compare the answer distributions and the differences between groups (for example by age) with the weighted Greek ESS data for each prompt language, and report where the Greek prompt brings the model closer to, or further from, the human pattern than the English prompt. The replication files of the 2026 cross-national study give English-prompt results for Greece from two other models, which can serve as a check on the English arm.

Data needed

ESS Round 11 data for Greece (free after registration on the ESS Data Portal); access to one language model through a free tier, a university licence or a small model run locally; a spreadsheet or a short R or Python script. Optionally, the replication files of the 2026 cross-national study (https://github.com/chuyao-wang/silicon-replication).

Feasibility

Undergraduate thesis. Low; free survey data and about 600 model queries, which a free or low-cost tier or a small local model can handle. Needs basic data handling and descriptive statistics at the level of a quantitative methods course, and a supervisor comfortable with a short script. The result describes one model at one date in one country; it cannot show whether silicon samples work in general.

Has this been done?

Partly done already: related work covers some of this question (first-pass screen of 18 October 2026; not yet checked in depth).

A 2026 preprint tests silicon sampling against ESS Round 11 in 30 countries with two open-weight models and finds moderate, uneven agreement at country level and negligible agreement for individuals. Its replication files include Greece, with English-language prompts. A 2024 study of 15 countries, Greece among them with 548 simulated participants, compared prompt languages in a populism framing experiment and found that Greek prompting greatly reduced fidelity for the framing effects. What remains open is narrower. No study was found that compares Greek with English prompting on ESS attitude items against Greek ESS respondents, or with a current model. The search was web-based and partial.

Where the search could not reach. Screened by three web searches and arXiv records on the date shown. On 19 October 2026, after the Owner's review pointed to them, the replication files of the 30-country study were read, showing that it covers Greece, and the 2024 study was read on its arXiv HTML page; the full text of the 30-country study was not read. Absence of a hit is not proof that the work does not exist.

Before building a project on it, run your own search in a scholarly database such as Scopus, Web of Science, Google Scholar or SSRN.

Nearest work found

• Wang, "Silicon sampling answers with country-level assumptions, not individual attitudes: Cross-national evidence from the European Social Survey", arXiv:2609.16395 (2026) Judged from an abstract, summary or search listing only
  Abstract and replication files read (https://github.com/chuyao-wang/silicon-replication, 19 October 2026); the paper itself was not read. ESS Round 11, 30 countries, 42 items, two open-weight models (Qwen 2.5-7B and Llama 3.1-8B); adding the country name drives country-level agreement, individual-level agreement is negligible, and averaging neighbouring countries without a model does as well. The country-level result files include Greece, with the Greek ESS mean and the model mean for every item, and the prompt templates in the generation script are in English; no Greek-language prompt was found in the files.
• Kwok, Bravansky and Griffin, "Evaluating Cultural Adaptability of a Large Language Model via Simulation of Synthetic Personas", arXiv:2408.06929 (2024)
  Read on the arXiv HTML page on 19 October 2026, after the Owner's review pointed to its Greek results. GPT-3.5 simulates the 7,286 participants of a 15-country populism framing experiment (Bos et al. 2020), including 548 in Greece. Naming the country of residence improved alignment with the human results; prompting in the persona's native language reduced it, and Greek and Hebrew prompting did not reach sign agreement above chance for the framing and relative deprivation effects. It is an experiment, not a survey of attitudes, and the model is GPT-3.5.

References

Sources behind the motivating claims:

• Argyle, Busby, Fulda, Gubler, Rytting & Wingate, “Out of One, Many: Using Language Models to Simulate Human Samples,” arXiv:2209.06899 v1, September 2022 (SRC-SL-ARGYLE-BUSBY-FULDA-GUBLER-RYTTING-WINGATE-ARXIV-2209-06899, original)
• Bisbee, Clinton, Dorff, Kenkel & Larson, “Synthetic Replacements for Human Survey Data? The Perils of Large Language Models,” Political Analysis 32(4):401–416 (2024) (SRC-LD07-BISBEE-SYNTHETIC-REPLACEMENTS-2024, original)
• Xie et al., “Evaluating the statistical realism of LLM-generated social science data,” PNAS 123(19):e2538145123 (2026) (SRC-LD07-XIE-SSDATABENCH-2026, original)
• Wang et al. NMI 2025 (SRC-R03-WANG-MORGENSTERN-DICKERSON-NMI-2025, no checked link to the original yet)
• Gong Weigang & Huang Siyuan, “整体事实还是有偏样本——基于大语言模型生成数据的测量,” Academic Monthly 57(6):123–136 (2025) (SRC-LD07-GONG-HUANG-CGSS-2025, original)
• Ling Wanying & Cui Sizhan, “复现的限度——‘硅基样本’的可能与可为,” Journal of Intelligent Society 4(2):179–201 (2025) (SRC-LD07-LING-CUI-SILICON-SAMPLES-2025, no checked link to the original yet)
• Chen, Zaharia & Zou, “How is ChatGPT’s behavior changing over time?,” Harvard Data Science Review 6(2) (2024) (SRC-LD07-CHEN-ZAHARIA-ZOU-DRIFT-2024, original)

BibTeX

The first entry cites this opportunity; then its nearest work; then the sources above. Nearest work and sources are recorded as reference strings, so each goes in whole as the title; add authors and year from the original before you cite it.

@misc{sociolens_ro_ld07_11_v1,
  author = {{SocioLens project}},
  title = {Does asking in Greek change how AI "silicon samples" answer as Greek respondents?},
  year = {2026},
  date = {2026-10-18},
  howpublished = {\url{https://sociolens-web.pages.dev/opportunities/ro-ld07-11/v1/}},
  note = {Research opportunity RO-LD07-11, version v1}
}

@misc{ro_ld07_11_nearest_1,
  title = {Wang, "Silicon sampling answers with country-level assumptions, not individual attitudes: Cross-national evidence from the European Social Survey", arXiv:2609.16395 (2026)},
  url = {https://arxiv.org/abs/2609.16395},
  note = {Reference as recorded by SocioLens; authors, title and year are not separated; Judged from an abstract, summary or search listing only}
}

@misc{ro_ld07_11_nearest_2,
  title = {Kwok, Bravansky and Griffin, "Evaluating Cultural Adaptability of a Large Language Model via Simulation of Synthetic Personas", arXiv:2408.06929 (2024)},
  url = {https://arxiv.org/abs/2408.06929},
  note = {Reference as recorded by SocioLens; authors, title and year are not separated}
}

@misc{src_sl_argyle_busby_fulda_gubler_rytting_wingate_arxiv_2209_06899,
  title = {Argyle, Busby, Fulda, Gubler, Rytting \& Wingate, “Out of One, Many: Using Language Models to Simulate Human Samples,” arXiv:2209.06899 v1, September 2022},
  url = {https://arxiv.org/abs/2209.06899v1},
  note = {Reference as recorded by SocioLens; authors, title and year are not separated}
}

@misc{src_ld07_bisbee_synthetic_replacements_2024,
  title = {Bisbee, Clinton, Dorff, Kenkel \& Larson, “Synthetic Replacements for Human Survey Data? The Perils of Large Language Models,” Political Analysis 32(4):401–416 (2024)},
  url = {https://doi.org/10.1017/pan.2024.5},
  note = {Reference as recorded by SocioLens; authors, title and year are not separated}
}

@misc{src_ld07_xie_ssdatabench_2026,
  title = {Xie et al., “Evaluating the statistical realism of LLM-generated social science data,” PNAS 123(19):e2538145123 (2026)},
  url = {https://doi.org/10.1073/pnas.2538145123},
  note = {Reference as recorded by SocioLens; authors, title and year are not separated}
}

@misc{src_r03_wang_morgenstern_dickerson_nmi_2025,
  title = {Wang et al. NMI 2025},
  note = {Reference as recorded by SocioLens; authors, title and year are not separated}
}

@misc{src_ld07_gong_huang_cgss_2025,
  title = {Gong Weigang \& Huang Siyuan, “整体事实还是有偏样本——基于大语言模型生成数据的测量,” Academic Monthly 57(6):123–136 (2025)},
  url = {https://www.xsyk021.com/article/id/4c1886eb-c385-4019-92ff-d1cd75547fa5},
  note = {Reference as recorded by SocioLens; authors, title and year are not separated}
}

@misc{src_ld07_ling_cui_silicon_samples_2025,
  title = {Ling Wanying \& Cui Sizhan, “复现的限度——‘硅基样本’的可能与可为,” Journal of Intelligent Society 4(2):179–201 (2025)},
  note = {Reference as recorded by SocioLens; authors, title and year are not separated}
}

@misc{src_ld07_chen_zaharia_zou_drift_2024,
  title = {Chen, Zaharia \& Zou, “How is ChatGPT’s behavior changing over time?,” Harvard Data Science Review 6(2) (2024)},
  url = {https://doi.org/10.1162/99608f92.5317da47},
  note = {Reference as recorded by SocioLens; authors, title and year are not separated}
}