Artificial-intelligence personas are beginning to enter the polling industry as a low-cost way to simulate focus groups and surveys. But a Syracuse University researcher says the technology faces a basic test before it can be trusted as a stand-in for voters: whether it can reflect changing public opinion.
Jennifer Stromer-Galley, a professor in Syracuse’s School of Information Studies who studies polling methodology, is warning that an AI persona can be consistent in a way that real people are not. A voter’s answer in October may differ from the answer they give months later as circumstances change. Whether synthetic respondents can represent that movement depends on how they are built and updated.
That issue goes to the core of synthetic polling. The approach can create hundreds or thousands of large-language-model personas using existing polling data and demographic information, then deploy those personas in artificial surveys or focus groups. The attraction is clear: polling vendors could expand their offerings without the logistics and expense of recruiting real people to answer questions.
What synthetic polling is trying to do
Traditional polling seeks responses from people at a particular time. Synthetic polling instead constructs digital stand-ins meant to represent prospective voters or other populations. The personas can be supplied with demographic information relevant to a poll and, in some cases, supplemented with documents and other material intended to make their positions more detailed.
In principle, a vendor could build personas with detailed perspectives on issues and policies, then ask them to participate in interest-based surveys or artificial focus groups. The system’s output may appear useful for testing messages, exploring scenarios or expanding research products.
Stromer-Galley describes the field as being at the beginning of a “synthetic persona era.” For polling firms, the prospect is to deliver additional research at relatively low cost. But lower operating costs do not resolve the methodological question of whether generated answers are an adequate representation of public opinion.
Consistency can become a flaw
Large language models generally work from the instructions and information used to construct a persona. That can make their responses stable. For some applications, stable behaviour is desirable: it can make testing repeatable and make a simulated participant easier to characterize.
Polling, however, deals with people whose views can shift. The issues a respondent considers most important at the beginning of October may not be the issues they identify three months later. Changes in personal circumstances, new information and developments in public life can all affect how people answer.
The concern is therefore not simply that an AI system might be inaccurate in a single response. It is that a system may preserve an initial profile too faithfully. If the persona continues to reproduce its starting assumptions after the real-world conditions surrounding its supposed counterpart have changed, it could miss the movement that polling is intended to detect.
That problem becomes especially consequential when synthetic results are interpreted as a signal of changing sentiment. A collection of personas may generate many answers quickly, but volume does not by itself establish that the answers capture the way public attitudes evolve.
Why the gap matters
Public-opinion polling has historically been valued for reliability, predictability and consistency. Synthetic methods complicate that picture because they introduce a different source of uncertainty: the design of the modelled respondent.
With a conventional survey, researchers must still make choices about sampling, questions and analysis. Yet the answers come from people experiencing the present moment. With an AI persona, the instructions, source data and supplemental materials become part of the respondent itself. The system’s assumptions about a voter may shape the answers before a question is ever asked.
That makes refresh and validation central issues. If a polling firm constructs a persona from demographic data and a set of policy signals, it needs a way to determine whether later output reflects genuine changes in the public environment or merely variations generated from a fixed underlying profile.
The source material does not establish that synthetic polling cannot be useful. It identifies an unresolved concern about whether the technology can account for opinion shifts. That is a narrower but important challenge: a simulated voter should not be judged only on whether it can produce a plausible answer today, but also on whether it responds credibly when the conditions that shape real answers have changed.
A methodological question, not just an AI question
The debate is ultimately about evidence. Polling firms may see AI personas as an opportunity to create richer offerings without the cost and difficulty of reaching respondents. The technology can also allow researchers to test a large number of scenarios rapidly. But those efficiencies do not automatically show that synthetic outputs measure public opinion.
For synthetic polling to earn wider confidence, its creators will need to address the difference between a well-described persona and a living respondent. A profile that reliably repeats an assigned position may be internally coherent. It may still fail to represent a person whose priorities, views or circumstances have changed.
That distinction is likely to matter more as vendors explore detailed policy personas, artificial focus groups and interest-based surveys. The more consequential the decisions informed by the output, the more important it becomes to know what the AI respondents are actually modelling.
Stromer-Galley’s warning provides a useful standard for evaluating the field’s next claims. Synthetic polling should be assessed not only for speed, scale and apparent realism, but for its ability to handle time. If public opinion moves and AI personas do not move with it, the technology could produce orderly results while missing the public it is supposed to represent.
This article is based on reporting by Phys.org. Read the original article.
Originally published on phys.org
