Prediction of Emic Responses via Generative-AI for Robust Information Environments (PEREGRINE)

In the fast-paced, massive modern information environment, IO Operators require an at-a-glance tool that offers detailed insights into the social and ideological perspectives of localized subpopulations. Prediction of Emic Responses via Generative-AI for Robust Information Environments (PEREGRINE) will encompass a suite of IO tools for LLM/AI-powered assessment, characterization, and forecasting of attitudes towards events and messages, grounded in a contextualized, emic analysis of the subjective perspectives of target populations. Based on user-specified parameters and queries that define a target location, population(s), and event/message, PEREGRINE will A) automatically source multi-modal, multi-linguistic, and multi-vector data to characterize the communities and populations of interest, then B) build group- and population-specific models at multiple scales (e.g., cultural, regional, national), informed both by theoretical models from contemporary social science and empirical evaluation of evolving narratives observed “in the wild”. The results of this analysis will be presented in a dashboard that includes a geographic heatmap, drill-down visualizations, and natural-language summaries. These tools will allow IO personnel to engage with PEREGRINE’s data and models, exploring stance, emotion, and attitudes towards events both quantitatively and qualitatively. In addition, PEREGRINE will feature an integrated conversational agent, allowing users to engage with AI-generated emic viewpoints in iterative testing, to assess how a typical population member is likely to respond to candidate narratives or possible events. PEREGRINE will deliver a fully capable software prototype that can collect the desired media, make the desired assessment (emotions/opinions tied to events) in selected geographical bounds/parameters, and provide a recommendation to an operator.Our Phase I effort will establish the feasibility of the prototype PEREGINE system via four core objectives that jointly will develop PEREGRINE’s analytic pipeline, build out a responsive user interface, and validate the results of emic models and summaries. We will (1) source multi-vector, multi-modal data at scale, segmented across diverse target populations defined by geography, demographic characteristics, and community membership. We will (2) develop multi-faceted, AI-powered psychographic models, integrating theoretical insights from the social sciences with bottom-up, empirical analysis of population perspectives and themes. These models will enable PEREGRINE to (3) present operators with quantitative visualizations and qualitative (natural language) summaries of target populations’ emic perspectives and attitudes on events and messages of interest. To evaluate the feasibility of our prototype system, we will (4) validate the system’s success at encapsulating target emic perspectives and evaluate the usability of our interface in internal testing and external demonstrations to AFRL.