Political research firms and polling organizations are beginning to field artificial intelligence systems capable of conducting synthetic survey interviews with simulated voter populations, raising fundamental questions about whether AI could eventually replace or substantially augment traditional telephone and online polling — and whether it should. Proponents argue the technology could eliminate sampling bias and dramatically reduce the cost and time required to gauge public opinion, while critics warn it risks severing the link between polling and actual human sentiment.
The technique, known variously as synthetic polling or agent-based opinion modeling, uses large language models fine-tuned on demographic and attitudinal data to generate simulated responses from representative populations. Rather than calling 1,000 real voters, a research firm constructs 1,000 AI agents calibrated to match the age, education, geography, partisan lean, and media consumption patterns of a target electorate, then poses survey questions to those agents and aggregates the results.
DataPulse Analytics, a political research firm based in Washington D.C., said it has been running synthetic polls alongside traditional polls for 14 months and has found that its AI model outperformed conventional polling in predicting the outcomes of 11 of 14 state-level elections tested during that period, including two races where traditional polling missed by more than six points. “The model is not reading minds,” said chief research officer Dr. Angela Marsh. “It is constructing statistically coherent opinion distributions from a richer feature set than any telephone sample can practically achieve.”
Conventional polling has endured a difficult decade. Response rates to telephone surveys have fallen below four percent in many markets, forcing pollsters to rely on weighting adjustments of increasing complexity to correct for non-response bias. Online panels have their own coverage problems. High-profile polling misses in several major elections between 2016 and 2024 have eroded public confidence in the enterprise and triggered methodological soul-searching across the industry.
AI proponents argue that synthetic polling sidesteps the response-rate problem entirely. Because the model is drawing on behavioral and attitudinal patterns embedded in training data — including consumer data, social media signals, and prior survey archives — it does not depend on persuading a real person to spend 12 minutes answering questions. Costs are correspondingly lower; DataPulse said a full synthetic national poll costs roughly 4,000 dollars to run, compared with 80,000 to 120,000 dollars for a comparable live-caller survey.
Skeptics raise both technical and philosophical objections. “What you are measuring is the model’s beliefs about what a 52-year-old white woman in suburban Ohio thinks, not what she actually thinks,” said Prof. Richard Okafor, director of the Electoral Research Centre at University College London. “If the training data systematically underrepresents certain groups, the synthetic population will too, and the model will produce confident but biased outputs without any of the uncertainty flags that honest pollsters attach to their work.”
There are also concerns about the potential for manipulation. Because synthetic polling data can be generated cheaply and quickly, critics worry it could be used to fabricate public opinion rather than measure it — producing misleading numbers to be fed into media coverage or used in political advertising. Several polling industry bodies have begun drafting disclosure standards that would require synthetic poll results to be clearly labeled as AI-generated.
The American Association for Public Opinion Research said in a statement that it was forming a working group to examine AI polling methodologies and expected to publish preliminary guidelines before the end of the year. “We are not dismissing the technology,” the association said. “We are insisting that it be held to the same standards of transparency and methodological disclosure as any other research instrument.”
Some practitioners see a hybrid future in which AI supplements rather than replaces human surveys. The model could be used to identify which subgroups are most difficult to reach, guide weighting strategies, or generate rapid initial estimates that are subsequently validated by smaller targeted polls. “The two methods have different failure modes,” said Dr. Marsh. “Used together, they could catch each other’s errors.”
Whether voters and journalists will trust AI-generated poll numbers remains an open question. Early evidence from DataPulse’s published work suggests the media is cautious; several outlets declined to report its synthetic results without also commissioning a parallel live poll for comparison. The firm says it welcomes that scrutiny. “We think validation is exactly the right approach for now,” Dr. Marsh said. “The burden of proof is on us, and we accept that.”