Read the prompt Pew uses to make AI pretend to be a survey respondent

Pew Research Center has put its synthetic-respondent prompt online. The Appendix A: Model prompt text, posted on the Data Labs site on September 30, 2026, contains the full instruction set the organization uses to put a language model in the seat of a U.S. survey taker. The artifact is a roleplay wrapper: it directs the model to behave like a profiled person, in a web survey, with the limits of a real panelist. The prompt spells out those limits because a real person would not need to be told them.

The prompt runs several pages and is written in plain second person. It directs the model to maintain consistency with prior answers, accept that its knowledge may be limited or mistaken, recognize that it cannot perform mathematics quickly, decline sensitive items when appropriate, and consider that its responses may be shaped by cognitive bias. Each instruction is a constraint a human panelist would not need. A person already knows they do not have perfect recall. A person already knows they can refuse to answer. The prompt exists because a language model would not know those things unless told.

Pew's methodology paper makes the comparison explicit. In a January 2026 human survey of 6,700 American Trends Panel respondents, GPT-5.1's synthetic respondents produced an average absolute error of 13.3 percentage points against the human distribution. Claude Opus 4.6's synthetic respondents produced an average absolute error of 11.4 percentage points. The pattern is consistent across the comparison charts for the three waves Pew analyzed: a weighted Wave 185 subsample of 6,700 respondents, and two additional waves of 3,398 and 4,981 respondents who also completed the political-typology survey. Same direction, different magnitudes.

On the question of satisfaction with the country's direction, GPT-5.1's synthetic respondents put dissatisfaction at 100%. The human panel put it at 69%. Claude Opus 4.6's synthetic respondents put it at 70%, close to the human figure. The 100% reading is not noise: it is what a model produces when it is asked to roleplay a politically engaged American on a topic where the model's training already points one way, and then told to remain consistent with prior answers. The instruction to hold the line amplifies a strong prior into a uniform response.

Pew published it. The methodology page says the point of posting the full text is to let outside researchers reproduce the setup and stress-test the comparison. That transparency is the contribution, and it is also the constraint. A coherence test, which is what the prompt measures (does the model stay in role, accept its limits, decline appropriately, bias itself plausibly), is not a validity test (do the answers describe a population that exists). Pew's diversity analysis shows the gap. Across political attitudes and personal behaviors, Opus-based synthetic respondents produced compressed distributions relative to the human panel. The model answered the questions. The variance of a real population did not survive the translation.

Pew wrote four self-imposed limits into the prompt. The model "does not have … the ability to quickly conduct complex mathematics, or any other abilities that a human would not realistically have." It is "realistic for your role to not know about some topics, or for their knowledge about some topics to be mistaken." The model "may occasionally choose to provide no response." The model's "answer may be affected by cognitive bias." Each line keeps the model inside a person-shaped box. A reader who has never seen the document can use those four lines to decide what a synthetic respondent is: a coherence performance, not a stand-in for the distribution of opinion a real survey measures.

The work that follows from this artifact is methodological. Anyone who wants to cite a synthetic poll needs to know what its respondent was told to be, and what it was told it could not be. Pew has now made both lists public.