How Synthetic Polling Replaces Human Respondents With AI Personas

Out of every 100 people a researcher tries to contact for a community survey today, fewer than two will actually complete it—a response rate that has collapsed from 36 percent in 1997 to roughly 1 percent in 2026. To bridge that 99-person gap without spending millions of dollars on field operations, the polling industry is testing a radical replacement: synthetic respondents.[4]

Instead of calling human beings, researchers feed demographic and behavioral data into large language models to create what the industry calls "silicon samples." A prompt defines the persona—for example, a 45-year-old female nurse in Ohio who voted for a specific candidate and attends church weekly. The model then answers the survey questions as that persona, generating thousands of responses in minutes.[1]

The foundational logic for this shift was established in a September 2022 paper published in Political Analysis, titled "Out of One, Many." Researchers demonstrated that the algorithmic bias within language models is fine-grained and demographically correlated. When properly conditioned, the models exhibited "algorithmic fidelity," accurately emulating the response distributions of specific human subgroups.[1]

The commercial appeal is entirely structural. Traditional field polls take weeks to turn around and cost tens of thousands of dollars. Synthetic polling platforms, such as Civly, generate simulations of real registered voters anchored to certified election results, delivering a full survey in minutes without a single phone call.[2][3]

Market research firms are rapidly commercializing the concept. In August 2026, Breakthrough Research launched "Modeled Communities," a product that builds synthetic respondent panels from extensive interviews with real consumers. The company positions the tool as a way to pre-test ideas, vet hypotheses, and reach harder-to-recruit audiences before committing to a full human study.

Proponents argue that behaviorally grounded synthetic respondents show high correlation with real human answers. Civly, for instance, back-tests its models against certified election results, reporting that its average county-level simulation lands within 1.7 points of the actual 2024 presidential margin.[2]

But independent evaluations reveal severe limitations when the questions move beyond basic top-line preferences. A July 2026 study by the UK analytics group Strat7 tested synthetic respondents against a real 3,000-person sample to determine if the technology was genuinely ready for complex research.

The results were mixed. On top-line headline results, the artificial intelligence was only two to three percentage points away from the human baseline. However, when asked to track changes in opinion over time, the synthetic surveys had a 47 percent success rate—statistically worse than a coin toss.

The Strat7 evaluation also exposed a distortion in economic modeling. When the synthetic respondents were given a pricing exercise to determine their willingness to pay for a product, they generated prices 16 percent higher than the real human participants, a margin that would ruin a product launch if relied upon by a corporate board.

Academic researchers are finding similar boundaries in civic applications. Studies presented at the Massachusetts Institute of Technology in 2026 investigated whether language models accurately weight intersecting demographic variables when simulating local electoral behavior. The research found that the models default to broad geographic and racial heuristics, overwhelming the finer-grained local signals that drive actual municipal voters.[4]

The output is also highly sensitive to how the artificial intelligence is prompted. In one experiment, prompting a model to behave as a "senator" caused it to strictly follow party lines, while prompting it as a "citizen" produced greater willingness to compromise. The model does not have an opinion; it simply plays the role it is assigned.[4]

Critics, including prominent polling analysts, warn that synthetic polling risks creating a feedback loop. If synthetic data is based on past polls, and future polls incorporate synthetic data, actual shifts in public opinion will be lost. A synthetic respondent cannot change its mind based on a new event that occurred after its training data was collected.[3][4]

The debate ultimately centers on what a poll is supposed to be. Optimists view synthetic polling as a way to unlock public wisdom and overcome declining civic participation. Skeptics argue that polling draws its legitimacy from participation itself—the act of asking and hearing real people. Swapping in digital twins replaces the democratic social contract with computational guesswork.[4]

For now, the industry consensus treats synthetic respondents as an augmentation tool rather than a replacement. "Our approach has always been in building custom market research solutions for our clients' most challenging business questions," said Dan Braker, President of Breakthrough Research, when launching the company's synthetic panels. "Our modeled communities apply that same rigor to a faster, more flexible format. This is a complement to a strong people-informed research program, not a replacement for one."

As the technology scales into the 2026 midterm elections, the deciding factor will be transparency. The market will soon test whether campaigns and corporations are willing to disclose when their data comes from a silicon sample, or if the allure of instant, frictionless consensus will quietly overwrite the need to ask humans what they think.[4]