A core feature of social science research is observing individuals and collecting their attitudinal and behavioral data. Advances in large language models (LLMs) will radically change the methods by which these data are collected and analyzed. Talking to Machines (T2M) is a collaborative project, developing the tools that facilitate the integration of AI and behavioral social science research. PoSSUM is one of these AI tools – it polls the public by inferring the attitudes and preferences of real-life social-media users with multimodal LLMs. It is being developed by Professor Roberta Cerina in collaboration with Ray Duch, a Director of the T2M project.

PoSSUM is a remarkably effective and cost-effective survey research tool for well-defined data collection environments. First, the tool requires a large corpus of identifiable individuals associated with the population of interest. For our 2024 U.S. Presidential election vote estimates, PoSSUM relies on the approximately 50 million active U.S. subscribers to X (formerly Twitter). But of course there are data collection applications with populations that are much more modest. Identifiable here implies basic socio-geographic demographics. Secondly, PoSSUM requires a large corpus of conversations, by these individuals, on “topic” that can be observed over time. Social media conversations are optimal because they are “unobtrusively observed” expressions of public opinions and preferences. Thirdly, PoSSUM adapts its LLM data collection and filtering strategies and its LLM forecasting algorithms with the assistance of a reliable stratification frame, i.e, a richly detailed census of the population of interest.

PoSSUM generates profiles of public opinion and preferences that are identical to conventional polling in terms of their rich detail. We believe that combining our PoSSUM AI methods with a large corpus of unobtrusive topical conversations generates highly precise profiles of these opinions and preferences – arguably more precise than conventional polling methods. The U.S. Presidential election campaign provides a unique opportunity to test these claims. Over the course of the 2024 campaign PoSSUM will be generating bi-weekly detailed Presidential vote estimates. They will be benchmarked against conventional polling estimates and of course will face the ultimate test on November 5th. Follow our regular forecast updates on this page.