We survey citizens’ voting preferences to understand or explain their voting decision but also to predict election outcomes. Because we observe election outcomes on a regular basis, we are able to monitor the trends in the performance of our modeling efforts. As Jennings and Wlezien (Reference Jennings and Wlezien2018) pointed out, the overall prediction error in preelection national polls actually has declined somewhat, reflecting the increasing number of polls being produced and individuals polled. Conversely, particularly during the past decade, state-level polls and some national polling organizations have performed poorly, and the results of some presidential contests have been more difficult to predict (Clinton et al. Reference Clinton, Cohen, Lapinski and Trussler2021; Jackson and Lewis-Beck Reference Jackson and Lewis-Beck2022; Kennedy et al. Reference Kennedy, Blumenthal, Clement, Clinton, Durand and Franklin2018). Maintaining a low level of prediction error in preelection polling has become increasingly challenging. This article describes how we address this challenge with a method that combines recent advances in Large Language Models (LLMs) with the proliferation of social media content. To illustrate, we estimate the vote shares of 2024 US presidential candidates on a biweekly basis using our AI polling method: PoSSUM, a Protocol for Surveying Social Media Users with Multimodal LLMs.
Election polling has faced challenges on a number of fronts but three core elements of the polling enterprise have proved particularly challenging. Election polls are now almost entirely conducted either on the telephone or online. Response rates for traditional random digit dial polls are now much less than 10% (Keeter et al. Reference Keeter, Hatley, Kennedy and Lau2017; Kennedy and Hartig Reference Kennedy and Hartig2019). Similar low response rates have been reported for recruitment into online surveys (Mercer and Lau Reference Mercer and Lau2023; Wu et al. Reference Wu, Nagler, Tucker and Messing2023). Selection effects imply that these samples often are not representative of the broader population. The use of increasingly unrepresentative samples contributes to systematic bias in the predictions of public opinion polling (Kennedy et al. Reference Kennedy, Blumenthal, Clement, Clinton, Durand and Franklin2018; Sturgis et al. Reference Sturgis, Baker, Callegaro, Fisher, Green and Jennings2016).
The foundation of traditional polling is a survey instrument that poses questions to which interviewees respond. Critical assessments of the design of these questions, the timing of the interview, and how survey respondents answer these questions suggest that the survey and interview likely bias polling results. A second possible factor contributing to prediction performance of election polls is the sincerity of voting intentions expressed by survey respondents. For example, evidence suggests that social desirability affects survey reporting of voting intention (Claassen and Ryan Reference Claassen and Ryan2024) and likely voting turnout.
A third critical and increasingly challenging element of the polling exercise is weighting of the sampled respondents (Gelman Reference Gelman2007; Rothschild et al. Reference Rothschild, Goel, Shirani-Mehr and Gelman2018). Most important, nonresponse is not random, which has undermined efforts to weight survey data. This has affected the accuracy of election surveys (Clinton et al. Reference Clinton, Cohen, Lapinski and Trussler2021; Kennedy et al. Reference Kennedy, Blumenthal, Clement, Clinton, Durand and Franklin2018) as well as surveys conducted in other areas (Bradley et al. Reference Bradley, Kuriwaki, Isakov, Sejdinovic, Meng and Flaxman2021). As a result, scholars give increasing attention to the correlation between whether and how people respond to surveys and how this correlation interacts with population size (Bailey Reference Bailey2023, Reference Bailey2024).
This article introduces an alternative AI-driven approach to polling that significantly reduces the estimation biases associated with these three features of traditional polling. Our biweekly PoSSUM estimation of the 2024 US presidential vote share provides an opportunity to test this claim. The article first describes how AI polling is likely to reshape the future of election polling. The second section describes the methodology. Results of our first two biweekly estimates of 2024 presidential vote share, benchmarked against other polling organizations, are presented in the third section. The discussion concludes in the fourth section.
THE AI FUTURE OF POLLING?
In the not-to-distant future, the entire polling enterprise will be redefined by the value added that LLMs can bring to the design, implementation, and analysis of surveys. Our PoSSUM poll of the 2024 US presidential election illustrates one direction that this AI election polling can take. Our proposed AI polling method leverages the proliferation of social media content and recent developments in LLMs while retaining the core features of a classic public opinion poll.
Population
The “target” population of interest is likely voters in the 2024 US presidential election. Our data collection is guided by a stratification frame that represents the population of the United States. We populate the relevant cells of this stratification frame with population figures from the American Community Survey (US Census Bureau 2021). The vote probabilities in these cells are estimated using multilevel regression with post-stratification (MRP) along with the results from our AI survey—an estimation strategy that Cerina and Duch (Reference Cerina and Duch2023) and others (Lauderdale et al. Reference Lauderdale, Bailey, Blumenau and Rivers2020) have championed as a method for improving the precision of vote-share estimates.
Sampling
The classic data-collection strategy for election polling is a version of a random probability sample from the population of individuals who are eligible to vote in the US presidential election. As previously mentioned, these samples are increasingly unrepresentative and problematic. In many cases, the sample is not from the US population per se but rather a segment of the population. This is the case, for example, with online surveys that sample individuals who have Internet access or who have been recruited into a sample pool.
All of these methods have in common the fact that the individuals in their sample respond to interviews either in person, on the telephone, or online. Our AI polling does not require our sample of people to respond to questions. The LLMs will collect digital traces from members of the population of interest. These digital traces will come from diverse subscribers but they hardly represent the complete population. This sampling requires that social media platforms provide sufficient information to allow the LLM to match the account holder to a cell in our stratification frame. There also must be a sufficient regular volume of political content to allow the LLM to infer an opinion or preference—in our case, likely vote choice. The LLM will parse out the digital traces that are informative. The goal is to construct a representative sample of the population of interest. Few social media platforms meet these criteria; X (formerly Twitter), with all its imperfections, does satisfy these conditions and is the basis for our online social media panel. Pfeffer et al. (Reference Pfeffer, Matter, Jaidka, Varol, Mashhadi and Lasser2023) provide an informative overview of the X “population”: their complete 24-hour “audit” of tweets generated 375 million tweets sent by 40,199,195 accounts. During this 24-hour period, the United States accounted for 20%, or about 70 million tweets, generated by 8 million accounts. The authors’ analysis of hashtags suggests that approximately 5% had a political theme—ignoring Iranian protest hashtags that accounted for 15% at the time. For our 2024 presidential vote-share estimates, we sample from these US X accounts. Previous efforts to use X for election forecasting have failed in part because of how the X samples are constructed and subsequently deployed in forecast modeling (Huberty Reference Huberty2015). We address these limitations by adopting an innovative approach to sampling social media that harnesses the power of recent advances in LLMs along with MRP statistical modeling.
The AI polling method we propose can accommodate and should include diverse social media platforms such as Facebook, Instagram, and TikTok. Each of these platforms caters to distinct demographic profiles, and tapping into this diversity would reduce bias in our digital sampling frame. Progress in incorporating this diversity into our digital sample is hindered by access restrictions to the Application Programming Interfaces (APIs) of these social media platforms.
Interview
Public opinion surveys consist of a questionnaire with closed and open-ended questions that are administered by an interviewer either in person or on the telephone; alternatively, they are administered online. As discussed previously, the “interview” must be constructed and administered, and it is the source of significant measurement error (Krosnick, Presser, and Art-Sociology Building Reference Krosnick, Presser, Building, Wright and Marsden2009). This is problematic because the accuracy of election polling is reliant on interviewees expressing sincere preferences and opinions. We avoid this particular source of measurement error with our method because LLMs do not ask questions. They unobtrusively observe digital conversations and infer preferences and opinions from the conversations—they are, for example, instructed to infer vote choice from the digital traces that they “digest.”
Although AI polling is unlikely to experience these conventional sources of measurement error, other types of measurement may be prevalent. Of particular concern for our method, from a measurement perspective, is whether (1) individuals are misrepresenting their sincere political preferences; and (2) this misrepresentation goes undetected by the LLM. For example, social pressures might lead some individuals to express “conforming” opinions within their social media networks. Our ongoing research will explore the extent to which this is the case. Although there clearly is a hesitancy for individuals to express their political preferences on social media, our intuition is that misrepresentation of preferences is probably relatively rare (McClain Reference McClain2019).
Uncertainty
A broader challenge that encompasses measurement error is to associate a measure of uncertainty with the estimates generated by AI polling. We propose a number of strategies in this regard. First, the LLM associates a speculation score with the profile estimate it generates (e.g., the profile’s gender and likely vote).
Weighting
Of course, our method makes no claim to be a random probability sample. Our point of departure is quota sampling. The LLMs are instructed to identify sufficient digital information for each cell of a stratification frame. The occurrences of the cells in the population effectively “weight” the digital opinions that we collect. We recognize the limitations—we are not observing the counterfactual identical individuals with each of our sociopolitical stratification frame profiles who are not X users. These “counterfactual” individuals may not be “missing at random,” thereby introducing bias into our estimates of vote share (Bailey Reference Bailey2023, Reference Bailey2024).
THE METHOD
As with conventional polling, our data collection focuses on sampling and conducting interviews (Cerina and Duch Reference Cerina and Duch2024). Our approach is tailored to the X API, which uses the digital trace of X users as the mold for LLM generation. However, this general approach can be extended to any social media that allows querying of a user panel via user- and content-level queries. PoSSUM is composed of two principal LLM routines that create the digital panel and then conduct the digital interview.
Gathering a Digital Panel
To create a digital panel of X users, we rely on the tweets and search API endpoint. Users who have participated in conversations related to the query during the past seven days (as per the limits of X ’s Basic API tier) are gathered to build the digital subject pool. Listing 1 is an example query for the X API. This type of query is likely to yield users who explicitly express opinions about candidates—and therefore will yield highly informative digital traces—that the LLM can annotate with confidence. However, selection effects loom large with this type of query—that is, the type of user who frequently comments on politics on X is likely to be different from one who does not, ceteris paribus. To account for this selection, we complement this political query with a set of queries based on currently trending topics (see https://trends24.in/united-states). Trending topics may be related to politics—for example, during party conventions and televised debates—although they are more likely to be associated with events such as sports, concerts, marketing campaigns, famous people, and otherwise viral online content. Users engaging with this set of queries are far more likely to be normies, who pay relatively little attention to politics and therefore can balance the high-attention selection associated with the query in listing 1. Figure 1 is an illustration of the trending topics associated with users in our digital panel.
Listing 1 Search Terms for Tweets Related to Candidates Involved in the US 2024 Presidential Election
Figure 1 Word-Cloud Presenting Words from the “Trending” Queries to the X API, for PoSSUM Polls Fielded Between August 2015 and August 2023 and Between July 2009 and December 2009
The words are weighted by the number of users associated with each word.
The digital panel then is further filtered, according to a number of sequential exclusion criteria. This is done for two reasons: (1) it contributes to data quality by ensuring that the digital traces belong to real existing users within the population of interest; and (2) it improves the efficiency of the sampling by identifying hard-to-find users who are more “valuable” for the pool. We exclude from the sample users whose self-reported location information is missing and those for whom we already have gathered a digital trace within the last τ days (i.e., to avoid overreliance on frequently active users). Users who do not represent a real offline person—including accounts for organizations, services, and bots—are discarded. Users who reside outside of the United States are discarded. We again rely on the LLM’s judgment, using the profile as a whole to make a determination when the self-reported location is not exhaustive or otherwise uncertain. Given the user’s characteristics, we then match the user to a cell in the population, according to a stratification frame (table 1 presents an example). If the user belongs to a cell for which a given representation quota has been filled, the user is discarded.
Table 1 Example Implementation of a Stratification Frame with Quota Counter
Notes: This table shows a stratification frame for a target sample size Ω*=1,500. The snapshot was taken with 647 respondents still to be collected.
Digital Interview
Users who survive the inclusion criteria comprise our final survey sample. Using the users/:id/tweets endpoint of the X API, we collect the most recent m tweets for each user. We append these tweets to the profile information and pass this augmented mold to the LLM to generate plausible survey responses for a given user. m is a hyper-parameter to be tuned depending on the provenance of the subject pool. Users captured among those discussing trending topics are unlikely to frequently generate text associated with political preferences and, as such, a larger record of their digital behavior is necessary to reasonably inform the LLM’s judgment. The opposite is true for users sampled via explicitly political queries, leading to the following heuristic: mtrending=λ × mpolitics, ∀λ>1.
Listing 2 presents an extract from the feature-extraction prompt. A features-object (see listing 3) is appended to this prompt. The features-object is given a standard structure: it is composed of a set of elements—each element contains a title, which describes a survey question; a set of categories, which represent the potential responses; and each category is identified by a unique symbol.
Listing 2 Standardized Feature-Extraction Operation
The text is followed by a list of features to be extracted, such as those in Listing 3.
Listing 3 Example of a “Dependent Features” Object
The feature-extraction operation considers all features simultaneously and prompts the LLM to produce a joint set of imputed features for a given user. We find that for most tasks, simultaneous feature extraction is preferable to a set of independent prompts, one for each attribute of interest. Separating prompts is an intuitively attractive choice due to its preservation of full independence among extracted features. However, this is extremely inefficient in terms of tokens, given that each prompt must redescribe the background, the mold, and the operations of interest. Prompting the LLM to extract all features simultaneously, by including the full list of desired features in a single prompt, is generally a productive approach.
An important caveat specific to this type of joint extraction pertains to the order in which features are presented in the prompt. The auto-regressive nature of LLMs (LeCun Reference LeCun2023) implies that when multiple answers are presented in response to a given feature-extraction prompt, previous answers will affect the next-token probabilities downstream. To minimize the overall effects of auto-regression on the generated survey-object, we can randomize the order of all features in the feature-extraction prompt so that order effects on the overall sample cancel out with a sufficient number of observations. The auto-regressive nature of the LLM is another reason that we prompt an explanation before a given choice is made, as opposed to after—we want to avoid post hoc justification of the choice and instead induce the LLM to select a choice that follows from a given line of reasoning.
We innovate LLM feature extraction by prompting a speculation score. A classic critique of silicon samples (i.e., synthetic survey responses) is that the data-generating process of the LLM ultimately is unknown. More crucially for PoSSUM, it is uncomfortable to not know how much of the LLM’s “own” knowledge—which it has acquired during its training phase—is responsible for a given estimate and how much is simply evident in the X profile and tweets.
To address this concern, we provide the LLM with instructions to generate a speculation score S ∈ [0,100] associated with each imputed characteristic. The wording of the prompt makes explicit that speculation refers to the amount of information in the observable data (e.g., the text of the tweets or the pixels of the profile image), which is directly useful to the imputation task and distinguishes this from other types of knowledge that the LLM might leverage. The score has a categorical interpretation that identifies “highly speculative” imputations at S>80.
Model-Based Weighting
As suggested previously, some quotas will be difficult to fill given the highly unrepresentative sampling medium (i.e., the X platform). The weighting method of choice is MRP (Gelman and Little Reference Gelman and Little1997; Lauderdale et al. Reference Lauderdale, Bailey, Blumenau and Rivers2020; Park, Gelman, and Bafumi Reference Park, Gelman and Bafumi2004). We consider this the obvious weighting choice given the sampling method: the explicit knowledge of unfilled quotas prompts a treatment of these cells as having missing dependent variables. We then can use a hierarchical model, under the ignorability assumption (Van Buuren Reference Van Buuren2018), to estimate the dependent values for the incomplete cells and stratify these estimates to obtain national- and state-level estimates. This also allows a comprehensive treatment of uncertainty at the cell level, which is liable to provide more realistic intervals on the poll’s national vote-share estimates than traditional adjustments.
The target stratification frame, which is derived from the 2021 American Community Survey (US Census Bureau 2021), is extended according to the MRP procedure (Leemann and Wasserfallen Reference Leemann and Wasserfallen2017) to extend the stratification frame and to include the joint distribution of 2020 Vote Choice as derived from the 2022 Cooperative Election Study (Schaffner, Ansolabehere, and Shih Reference Schaffner, Ansolabehere and Shih2023) (see table 1).
The hierarchical model used to generate estimates of the dependent variable of interest imposes structure (Gao et al. Reference Gao, Kennedy, Simpson and Gelman2021) to smooth the learned effects of a model trained on AI-generated data in a sensible way. LLMs can leverage stereotypes in making their imputations (Choenni, Shutova, and van Rooij Reference Choenni, Shutova and van Rooij2021), which can translate to exaggerated relationships between covariates and dependent variables. Adding structured smoothing to the model allows us to correct for this phenomenon to some degree. We regress the dependent variable—which is assigned a categorical likelihood with SoftMax link—onto sex, age, ethnicity, household income, and 2020 vote. Sex and ethnicity effects are estimated as random effects; stateFootnote 1 effects are assigned an Intrinsic Conditional Auto-Regressive (ICAR) prior (Besag, York, and Mollié Reference Besag, York and Mollié1991; Donegan Reference Donegan2022; Morris Reference Morris2018); and date, income, and age effects are given random-walk priors. Separate area-level predictors are created for each dependent variable of interest. The covariates and parameters used in the model for 2024 vote choice are presented in table 2.
Table 2 Model Predictors and Parameters for the 2024 Vote-Choice Model
Notes: “iid” refers to fully independent parameters or “fixed” effects (see Gelman et al. Reference Gelman, Carlin, Stern, Dunson, Vehtari and Rubin2013). “Unstructured + shared variance” priors refer to classic random-intercepts. Random-walk and spatial correlation structures are explained in detail in the article text.
We have described the three broad features of our AI polling method: recruitment, sampling, and measurement. They correspond to similar core elements that define telephone and online polling methods. To put the elements of our AI method in context, figure 2 compares our AI approach to these three core activities with those undertaken for telephone and online polling.
Figure 2 Election Polling: Random Digit Dial, Online, and AI Polling
RESULTS
During the course of the 2024 US presidential election campaign, we are publishing biweekly vote-share estimates for the candidates. These include the national vote-share estimates for the presidential candidates as well as the vote-share breakouts at the state level, along with vote-share tables for our key sociodemographic profiles. Our national-level vote-share estimates from our August 15–23 and September 7–12, 2024, AI polls are presented in table 3. For our first August wave of the PoSSUM, we estimated that Harris had a national vote share of 46.4% compared to 47.2% for Trump. In the second wave, Harris scored 47.6% and Trump registered 46.8% (i.e., Harris is estimated to have 50.4% of the two-party vote). Table 4 breaks out these estimates by gender. As most election polling has been suggesting, Harris has a significant lead over Trump with women and Trump leads Harris among men. As indicated in table 5, race and ethnic differences between Harris and Trump supporters match those of other polling organizations: Trump has a lead over Harris with whites; Harris has a Black and Hispanic lead over Trump, and this appears to be growing. The PoSSUM national presidential vote-share estimates, along with demographic breakouts, align with similar estimates by the leading US polling organizations.
Table 3 PoSSUM Poll Estimates of National Presidential Candidates’ Vote Share
Table 4 PoSSUM Poll Estimates of 2024 Presidential Vote Choice by Sex
Table 5 PoSSUM Poll Estimates of 2024 Presidential Vote Choice by Race/Ethnicity
To benchmark our estimates against those of other major US presidential polls, we analyze the vote-share cross-tabulations produced by these polling organizations. This allows us to benchmark our estimates on a biweekly basis. The results for our first two polls are presented in figure 3. Each polling estimate includes a 95% confidence interval. Note that the line in each figure is the overall average for the vote-share estimates of all of the polling organizations. In the case of the Trump vote share, our PoSSUM MRP estimate is slightly higher than this average in the August poll and almost identical to this average in the September poll. Our vote-share estimate for Harris is lower than most other measurements in both the August and September polls.Footnote 2
Figure 3 Benchmarking PoSSUM 2024 US Presidential Vote-Share Estimates with Major Polling Houses
The dotted line represents the simple average of polls for each candidate (excluding PoSSUM).
As described previously, the PoSSUM 2024 presidential election study constructs a national sample of the US voting population. It is feasible, however, by using our MRP modeling strategy to generate state-level estimates of candidate vote share. Given that the sampling strategy was not designed to generate representative samples of individual state voting populations, we expect state-level vote-share estimates to be imprecise. Nevertheless, the state-level breakouts provide an additional indication of the robustness of our AI polling method. Figure 4 presents state-level vote-share differences for the two Republican and Democratic candidates (i.e., Republican vote share minus Democratic vote share). Posterior distributions are shown for states where polls have been fielded in a comparable period and are published on the FiveThirtyEight state-level polling database. There are some states in which the estimates are implausible—Maine, in particular, although its estimates are based on a total of four users across both samples and, as such, should be discounted. We aim to aggregate samples from our biweekly polls, accounting for temporal dynamics in the MRP, to improve state-level coverage. For the important swing states—with the possible exception of Wisconsin—the results track those of other major polling organizations. The dotted vertical line in the state figures represent these simple polling averages for the state. If we consider Arizona, for example, the polling organization average difference between Republicans and Democrats essentially is zero. We are estimating a 2.2% lead for the Republicans and a probability of a Republican win of 0.80. Although the AI sampling strategy was not designed for estimating vote share at the state level, our state breakouts are generally reasonable, providing further evidence of the robustness of the AI polling method. Our Electoral College vote-share estimates based on the state forecasts from the September poll are 301 for Trump versus 237 for Harris.
Figure 4 Benchmarking PoSSUM 2024 US Presidential Vote-Share Estimates State Breakouts
The dotted line represents the simple polling average for that state. The x-axis presents the Republican lead in the district. States are ordered alphabetically.
CONCLUSION
The PoSSUM 2024 US presidential election vote project explores the feasibility of replacing conventional election polling estimates with an AI survey application. Our goal is to provide the only detailed and open-sourced AI polling estimates of the 2024 US presidential election candidate vote shares. On a biweekly basis during the US presidential campaign, we publish our vote-share estimates at the national and state levels. Additionally, we harmonize estimates being generated by other polling organizations and benchmark them against our detailed estimates.
This article identifies a number of the most serious challenges currently facing election polling. We make the case that LLMs combined with rapidly growing social media content are the solution to the serious challenges facing conventional polling today. Increasingly unrepresentative samples are a serious challenge for election polling. We address this challenge with a sampling method that leverages voluminous social media content with the rapidly increasing capabilities of LLMs. Of growing concern for election polling is the declining quality of the data generated from a conventional survey interview with humans. There are no humans interviewed in our AI polls. LLMs unobtrusively observe, collect, and analyze human opinions that are expressed by human subjects in social media conversations. Conventional election predictions require a strategy for weighting the data that are generated from increasingly unrepresentative samples. Weighting is accomplished transparently by our PoSSUM method because vote probabilities are estimated using MRP with a stratification frame that guides the LLM in creating our digital sample.
Our initial predictions confirm that presidential candidate vote-share estimates based on AI polling are broadly exchangeable with those of other polling organizations. We present our first two biweekly vote-share estimates for the 2024 US presidential election and benchmark them against those being generated by other polling organizations. Our post–Democratic National Convention presidential vote-share estimates for Trump (47.2%) and Harris (46.4%) closely track results generated by other polls during the month of August. The subsequent early September (post-debate) PoSSUM vote-share estimates for Trump (46.8%) and Harris (47.6%) again closely track with other national polling being conducted in the United States. An ultimate test for the PoSSUM polling method will be the final preelection vote-share results that we publish before Election Day on November 5, 2024.
LLMs will have an increasingly important role in how we conduct preelection polling. The methods we describe in this article and the open-sourced code being made available to readers are an important foundation for facilitating the integration of AI into our election polling strategies.
ACKNOWLEDGMENTS
We are grateful for the generous research funding support provided by the University of Oxford “Talking to Machines Project,” the Swiss National Science Foundation (Grant No. 100018M-215519), and Nuffield College Oxford.
DATA AVAILABILITY STATEMENT
The editors have granted an exception to the data policy for this manuscript. In this case, replication code and data are available to reproduce its figures and tables, but there are substantively small differences between the replication and the printed results. This exception was granted because the authors affirmed that these differences are attributable to randomness in the sampling procedure that generates draws from Bayesian posterior distributions and do not change the conclusions of the manuscript.
CONFLICTS OF INTEREST
The authors declare that there are no ethical issues or conflicts of interest in this research.
Open access
