miklia/llm-synthetic-survey-respondents-stochastic-parrots

AI Parrots: synthetic survey respondents from leading LLMs

This dataset holds 10,592 synthetic survey respondents generated by consumer AI platforms and frontier large language models. Every respondent answered the same 32-item questionnaire on ethics, political ideology and workplace experience in technology firms. The synthetic samples were benchmarked against a survey of Silicon Valley coders and developers with 400 complete human responses.

The data accompany Miklian, Hoelscher and Katsos (2026), Stochastic Parrots or Singing in Harmony? Testing Leading AI LLMs' Ability to Replicate a Human Survey with Synthetic Data, White Paper, Centre for Global Sustainability, University of Oslo.

• Paper page and free PDF: https://miklian.org/papers/stochastic-parrots-or-singing-in-harmony-testing-five-leading-llms-for-their
• arXiv: https://arxiv.org/abs/2603.00059
• SSRN: https://doi.org/10.2139/ssrn.6210099
• Zenodo: https://doi.org/10.5281/zenodo.23087156

Configurations

Config	File	Respondents	What it is
study2 (default)	data/study2/responses.csv	6,419	Claude Fable 5, GPT-5.6 Sol and Kimi K3, July 2026, under protocol A (shared conversation) and protocol B (one fresh context per assigned persona)
study1	data/study1/responses.csv	2,173	Consumer chat platforms, October 2025 to January 2026: ChatGPT, Claude, Gemini (two files), DeepSeek and Claude Cowork
prompt_sensitivity	data/prompt_sensitivity/responses.csv	2,000	Five prompts of increasing specificity (S1 to S5) run against a single model, 400 respondents each
datasets	data/datasets.jsonl	29 rows	One row per synthetic dataset: requested and observed n, analysis role, excluded items and limitations
item_level_*	data/item_level/*.jsonl.gz	338,944 item records	The same responses in long form, one row per respondent and item, with raw value, canonical value, coding status and eligibility flags

The respondent-level CSVs have one row per synthetic respondent. Columns are dataset_id, record_id, study, analysis_role, then age_gender and Q1 to Q31, each holding the canonical answer text. Multi-select answers are joined with |. Question and answer wording is in codebook/Codebook.pdf and codebook/instrument_dictionary.json.

Study 2 runs

Model label	Protocol A runs	Protocol B runs
Claude Fable 5	420, 420, 420	285, 286, 288 (truncated by a platform session limit)
GPT-5.6 Sol	420, 420, 420	420, 420, 420
Kimi K3	420, 100 (interim), run 3 missing	420, 420, 420

Protocol B assigned the same 420 personas in each numbered run across all three models; the assignments are in materials/personas/. Kimi K3 ran its two protocols at different reasoning settings.

Read before using

• Every row is model output. None of it describes a real person.
• Check analysis_role and the datasets config before pooling anything. S1_GEMINI_B2 is a labelled derivative of S1_GEMINI_B1 and was excluded from the analysis; S1_GEMINI_B1 has unresolved numeric coding; S1_CLAUDE items Q9 to Q31 used altered answer menus; S2_kimik3_A_run2 is a 100-row interim file.
• Respondents generated inside one conversation are not independent observations, and repeated profiles remain in the data.
• Model labels are the labels recorded at generation. Where request metadata are missing, the exact served model version is not independently authenticated.
• The prompt-sensitivity archive carries conflicting model descriptions (Claude Sonnet 4.5 and Claude Opus 4.6).
• A sixth Study 1 platform was removed from the study because its provenance could not be documented; its data are not included here.

Human benchmark

The human survey records are not in this dataset. They are in the full replication package, together with the analysis code that reproduces the paper's numerical results: https://drive.google.com/file/d/1jyT2qAAfTbkPD8_Q9NGqZwmhQLUoB6Vb/view?usp=drive_link. Aggregate human rates used in the paper are in results/human_references.csv and in the codebook. Human survey data were provided by RIWI. The benchmark study is Miklian and Hoelscher (2026), A new digital divide? Coder worldviews, the 'Slop economy,' and democracy in the age of AI, Information, Communication & Society 29(8), 2412-2428, https://doi.org/10.1080/1369118X.2025.2566814.

Other files

• materials/: Study 1 and Study 2 prompt templates, issued user prompts where transcripts exist, assigned personas, generation settings, request-log metadata, the archived survey instrument and the provenance probe record.
• results/: per-run estimates and summaries, the manuscript tables and the nine manuscript figures.
• methods/analysis.md: outcome definitions, denominators, the consistency rule and the distance measures.
• codebook/: the PDF codebook, the instrument dictionary, outcome definitions and the raw-to-canonical coding rules for the synthetic files.

Citation

@techreport{miklian2026_stochastic_parrots_or_singing,
  title = {Stochastic Parrots or Singing in Harmony? Testing Leading AI LLMs' Ability to Replicate a Human Survey with Synthetic Data},
  author = {Miklian, Jason and Hoelscher, Kristian and Katsos, John E.},
  institution = {Centre for Global Sustainability, University of Oslo},
  type = {White Paper},
  year = {2026},
  doi = {10.2139/ssrn.6210099},
  url = {https://miklian.org/papers/stochastic-parrots-or-singing-in-harmony-testing-five-leading-llms-for-their},
  note = {Also available at arXiv:2603.00059}
}

License

CC BY-NC-SA 4.0, matching the paper's arXiv licence.