Session: For the Common Good: Harnessing Cutting-Edge AI Methods to Drive Social Justice and Democratic Inquiry (Society for Social Work and Research 31st Annual Conference)

Methods: This symposium includes four presentations complementary in their data sources and AI methodologies. Focusing on housing evictions in the U.S., the first study used unstructured 40,112 eviction courts records filed between 2004 and 2024; it applied a vision language model for key data extraction and supervised machine learning models to predict sheriff-enforced evictions (Study 1). Aiming to better a voting bloc that was influential in the 2024 U.S. Presidential Election, the second study used survey data collected in 2025 from 367 politically conservative men; it applied random forest modeling with SHapley Additive exPlanations (SHAP) analysis to predict the most important predictors of fathers’ involvement with their children in politically conservative households (Study 2). Designed to help clinicians address chronic pain flares in military veterans, the third study leveraged multimodal data collected through daily diaries and wearable devices over 4,700 person-days; it used gradient-boosted mixed-effects models to identify same-day pain and predict next-day pain increases (Study 3). With the goal of poverty reduction in under-resourced communities, the fourth study used multimodal data from satellite imagery, social media, and crowd-sourced geographic information within 535 Zambian villages; it used deep learning modeling–unimodal, multimodal, stacked multimodal models–for high resolution poverty mapping, poverty prediction, and assessment of equitable resource allocation (Study 4).

Results: The first study showed the feasibility of using a vision language model for data extraction; moreover, it found that eviction properties with sales and the proportion of White residents were most predictive of sheriff-enforced evictions. The second study demonstrated the importance of adding SHAP analysis to random forest for model transparency and deeper insights into the contributions of top variables–such as political conservative men's beliefs about fathering, life satisfaction, family functioning–in predicting father involvement. In the third study, models using wearable data only performed just as well as models that also included daily diary data in classifying veterans’ same-day pain; affect, sleep, and activity were among the top predictors of next-day pain. The fourth study revealed that stacked multimodal modeling performed the best, with infrastructure features serving as the most important predictors of poverty; moreover, optimization-based budget reallocation reduced poverty severity most effectively, with more egalitarian approaches achieving broader geographic coverage.

Conclusion and Implications: This symposium highlights state-of-the art social work research that employs advanced AI methodologies to solve complex social problems and thus advance democracy. The symposium will also provide critical reflections on reducing bias and increasing transparency in AI models and introducing practical strategies to effectively disseminate evidence. An expert in AI and data science in social work will serve as the discussant.