Syracuse professor warns AI polling can miss changing views after Pew’s 12-point gap

AI voter personas may keep giving consistent answers while the people they represent change their minds, Syracuse professor Jennifer Stromer-Galley warned in comments published October 19. Her warning highlights a risk behind Pew Research Center’s September experiment, in which synthetic survey results differed from human answers by an average of 12 percentage points.

Synthetic polling uses AI-generated personas, built from polling data and demographic information, as substitutes for people in surveys and focus groups. Stromer-Galley says their appeal is lower cost and fewer recruitment hurdles. But matching a population’s demographics does not necessarily correct how a model answers for each profile.

Real profiles, different answers

Pew’s September 30 study used a “digital twins” approach: an AI model adopted the personas of actual American Trends Panel members. It received both their demographic details and their answers to Pew’s 2025 political typology survey, which asked about their political views.

• The model answered questions from three surveys conducted in the first half of 2026, following the same order, instructions and survey programming as human participants.
• Comparisons included only panelists who had completed the political typology survey, so human figures can differ slightly from previously published results.

Across nearly 300 questions, the average gap was 12 percentage points. Nearly half the questions also had at least one answer choice that no AI respondent selected, according to Syracuse’s description of the experiment.

The models’ strengths shifted between questions. Opus estimated dissatisfaction with the country’s direction at 70%, close to the human result of 69%; GPT estimated 100%. Yet GPT closely matched the share seeing clear solutions to major national issues—57%, against 56% among humans—while Opus estimated 25%.

Consistency can hide a failure to change

Stromer-Galley’s concern extends beyond reproducing one survey. A person’s priorities may change as their circumstances change. A persona anchored to fixed instructions may instead produce a narrower range of answers over time. She also warns that people with views atypical for their demographic group can disappear from synthetic results.

Pew says newer models and better synthetic-sample methods could reduce errors. Its experiment tested AI-generated survey responses, not every use of AI in polling. The center itself has no current or future plans to use AI models to generate survey results.