Necessary for What? Auditing Demographic Disclosure to LLM Synthetic Respondents
Organizations: Institute of Information Science, Academia Sinica, Taipei, Taiwan
Abstract
Large language models (LLMs) increasingly answer surveys on behalf of real panelists, and each panelist's recorded demographic attributes (e.g., age, education, ethnicity) are often placed in the prompt on the premise that richer profiles make more faithful synthetic respondents. While data minimization limits personal data to what a purpose requires, a synthetic survey serves several purposes at once, from predicting individual answers to estimating how opinions are distributed and comparing groups. We ask which of these purposes require the respondent's record, and whether improved estimates support sound conclusions about groups. We vary whose profile a model receives, where it comes from, and how many fields it contains, in two U.S. surveys with eleven open-weight and three hosted models, and compare each model with a statistical model fit to the same information. We make four contributions. First, we show that the record adds little to individual predictions and that, for the open-weight models, most of its gain in population estimates remains when it is replaced by census data or a synthetic persona. Second, we find that group comparisons improve only with the respondent's own record, and that even then the models' group differences rest more on the demographic label in the prompt than on the individual respondent, which the usual check of reproducing known group differences does not reveal. Third, we find that the models make groups look more alike than they are, keeping differences in circumstances while flattening or reversing differences in attitudes, most sharply for religious affiliation and ethnic origin, both special categories of personal data. Fourth, we outline an audit that a survey organization can run on questions it has already fielded to decide which estimates may use the record.
Figures & tables
| Condition | Profile the model receives | Compared with | Question |
| Value and linkage | |||
| History | None (country only) | – | Baseline |
| Own | The respondent’s 15 recorded fields | History | Does the respondent’s record help? |
| Shuffled | A random respondent’s 15 fields | Own | Does correct linkage matter? |
| Substitution | |||
| Shuffled-10 | A random respondent’s same 10 fields | Own-10 | Does linkage matter on these fields? |
Appendix figures & tables12 assets
Supplementary material from the paper’s appendix.
Appendix
| ID | Short wording | Counted response side ( ) |
|---|---|---|
| Pew panel | ||
| DEBTa_W54 | Credit card debt | No, do not have this |
| DEBTb_W54 | Car loan | No, do not have this |
| DEBTc_W54 | Student loans | No, do not have this |
| DEBTd_W54 | Mortgage or home loan | No, do not have this |
| DEBTe_W54 | Medical-bill debt | No, do not have this |
| Field | Own | Own-10 | Own-7 | Own-4 |
|---|---|---|---|---|
| Census region | ||||
| Sex | ||||
| Age group | ||||
| Education | ||||
| Race or ethnicity | ||||
| Citizenship |
| Open-weight | Hosted |
|---|---|
| google/gemma-4-12B-it | gpt-4.1-2025-04-14 |
| google/gemma-4-31B-it | gpt-6.1-terra |
| allenai/Olmo-3-7B-Instruct | claude-haiku-5-5 |
| allenai/Olmo-3.1-32B-Instruct | |
| Qwen/Qwen2.5-7B-Instruct | |
| Qwen/Qwen2.5-72B-Instruct |
| : reproduced (%) | : reproduced (%) | : TVD gain (pp) | Own overlap (%) | ||||||
| Survey (pairs) | History | Own | Shuffled | History | Own | Shuffled | Own | Shuffled | |
| Pew (1,281) | 0.0 | 40.4 | 1.2 | 16.5 | 41.8 | 9.0 | 10.24 | 10.21 | 76.3 |
| ISSP (1,268) | 0.0 | 33.5 | 0.6 | 15.8 | 34.9 | 11.8 | 7.93 | 8.08 | 64.3 |
| Both surveys | — | 71.0 | |||||||
| Individual Brier ( ) | Population TVD | Subgroup TVD | Contrast error | Reproduced share | ||||||
| Condition | Models | Ref. | Models | Ref. | Models | Ref. | Models | Ref. | Models | Ref. |
| History | — | — | — | — | — | — | — | — | 22.6% | 41.8% |
| Own | 2.56 (8/1/2) | 1.36 | 0.74 (6/3/2) | 0.42 | 0.37 (6/3/2) | 1.41 | 1.13 (7/1/3) | 6.66 | 41.8% | 82.2% |
| Reduced own profile | ||||||||||
| Own-10 | 2.50 (9/2/0) | 1.23 | 1.07 (6/2/3) | 0.39 | 0.66 (6/3/2) | 1.10 | 0.43 (6/4/1) | 5.32 | 40.0% | 72.8% |
| Own-7 | 1.77 (8/2/1) | 1.03 | 0.43 (6/3/2) | 0.41 | 0.24 (5/4/2) | 0.96 | 0.53 (7/1/3) | 4.03 | 34.5% | 68.1% |
| Model | Comparison | Individual Brier | Population TVD | Subgroup TVD | Contrast error |
|---|---|---|---|---|---|
| GPT-4.1 | Own vs. History | 5.38 | 1.22 | 0.97 | 1.24 |
| Shuffled vs. History | 0.91 | 0.59 | 0.73 | 1.39 | |
| PUMS-10 vs. History | 0.82 | 0.77 | 0.61 | 0.79 | |
| Own-10 vs. History | 3.32 | 0.69 | 1.03 | 0.23 | |
| Own-4 vs. History | 2.42 | 0.42 | 0.40 | 0.55 | |
| PUMS-10 vs. Own-10 | 4.14 | 0.08 | 0.42 | 1.01 |
| Pew panel (1,281 prompted pairs), Own | ISSP (1,268 prompted pairs), Own | |||||||||||
| Model | Reprod. (%) | Missed (%) | Median | New | Corr. | Net | Reprod. (%) | Missed (%) | Median | New | Corr. | Net |
| Gemma 4 12B | 53.0 | 44.7 | 0.596 | 21 | 23 | -2 | 45.0 | 54.3 | 0.380 | 6 | 21 | -15 |
| Gemma 4 31B | 60.2 | 38.0 | 0.853 | 13 | 30 | -17 | 52.4 | 46.6 | 0.635 | 12 | 10 | +2 |
| OLMo 3 7B | 29.7 | 45.1 | 0.017 | 254 | 35 | +219 | 15.1 | 75.6 | 0.012 | 82 | 29 | +53 |
| OLMo 3.1 32B | 37.5 | 57.7 | 0.148 | 44 | 18 | +26 | 38.2 | 58.4 | 0.203 | 27 | 23 | +4 |
| Qwen2.5 7B | 32.6 | 65.2 | 0.014 | 25 | 20 | +5 | 22.6 | 73.3 | 0.050 | 49 | 17 | +32 |
| Attribute | Item class | Pairs | Model pooled (%) | Reference reproduced (%) |
|---|---|---|---|---|
| Age | Circumstances | 50 | 59.27 / 10.91 / 4.55 | 96.00 |
| Attitudes | 92 | 22.92 / 38.74 / 13.34 | 81.52 | |
| Education | Circumstances | 103 | 84.11 / 5.12 / 2.29 | 97.09 |
| Attitudes | 120 | 28.26 / 36.14 / 9.02 | 75.00 | |
| Marital status | Circumstances | 73 | 70.86 / 6.35 / 3.49 | 89.04 |
| Attitudes | 115 | 27.19 / 45.14 / 4.98 | 80.00 |
| Item class | Item identifiers |
|---|---|
| Circumstances (15) | DEBTa – DEBTe , FINANCEa – FINANCEc , FIN_SIT , FIN_SITFUT , FIN_SITGROWUP , WORRY2a , WORRY2c – WORRY2e . |
| Attitudes (48) | ECON1 , ECON1B , ECON3_a – ECON3_i , FIN_SITCOMM , FIN_SITMOST , GOVPRIORITYa – GOVPRIORITYf , GOVRESP_a – GOVRESP_h , INEQ3 , INEQ4_a – INEQ4_e , INEQ5_a – INEQ5_m , INEQ6 , INEQ7 . |
| Profile | Median ACS count | Below 1,000 adults (%) | Unique in Wave 54 profile (%) | Unique in Wave 54 profile + history (%) |
|---|---|---|---|---|
| History only | – | – | – | 18.0 |
| Own-4 | 1,509,304 | 0.0 | 0.1 | 90.4 |
| Own-7 | 776,456 | 0.5 | 6.5 | 93.6 |
| Own-10 | 21,211 | 13.1 | 46.6 | 99.5 |
| Own (15) | – | – | 94.5 | 100.0 |
LLM Auditing
Large language model auditing tests models and deployed systems for bias, safety, privacy, reliability, and provenance using behavioral and internal evidence.
See all 719 papers