Forced or Voluntary? Using a Vision Language Model to Understand Sheriff-Enforced Evictions from Court Cases

Background and Purpose: Eviction is a significant challenge in the United States today, with approximately 3.6 million legal eviction cases filed annually. One major obstacle to comprehensively understanding this issue is that a big amount of information is in unstructured court records, in PDF format. While some courts may provide filing data or outcomes of cases—such as judgments or dismissals—the way these cases are executed remains still understudied. This is significant because a sheriff’s execution can involve a range of situations, such as “ousted and ejected” or “returned per plaintiff.” If a tenant is forcibly evicted by the sheriff, it may indicate immediate homelessness and exposure to potentially violent behavior in the process. Thus, this study aims to identify individual, community, and macro-level factors associated with whether tenants move voluntarily or remain in the property until the sheriff enforces the eviction after judgment. The AI method used here also illustrates how court PDFs can be transformed into data for large-scale social work research.

Methods: This study employs a vision language model to extract key information, like execution outcomes, from PDF files. The analysis covers 40,112 unique eviction cases with sheriff’s return PDFs (71.7% of all cases) filed between 2004 and 2022 in a western US county. The location data of the cases enables linkage to property, community, and county-level datasets. Two classification models—logistic regression and random forest—are used to identify key factors associated with the outcomes of sheriff-executed evictions.

Results: The proportion of sheriff-enforced evictions rose from 20% in 2013 to nearly 50% in 2021, despite an overall drop in filings after 2020. In 2020, the number of expired (unexecuted) cases spiked, likely due to the national eviction moratorium. By 2021, the rise in sheriff-enforced evictions may reflect tenants’ reduced ability to secure alternative housing. Key variables included racial identity, property sale records, property type (e.g., single vs. mobile home), landlord type (organization vs. individual), local unemployment rate, and fair housing value. At the individual level, eviction properties with sales in the same year had higher odds of sheriff enforcement (AOR = 1.26, p < .01). At the census tract level, higher percentages of White (AOR = 0.89, p < .01) and Hispanic (AOR = 0.92, p < .01) residents were linked to a lower likelihood of sheriff eviction. The random forest model revealed additional factors, such as lower property values, associated with higher odds of sheriff execution.

Conclusions and Implications: With the assistance of AI, this study demonstrates an innovative approach to collecting data from unstructured PDF-format court files. This makes it possible to include tenant and property characteristics in the analysis—features often missing from traditional eviction datasets, which are typically aggregated. This method can be expanded to other types of unstructured data, such as historical administrative data. The findings also have practical implications: they can inform interventions to prioritize tenant rights education and prevent violence during evictions, ultimately supporting affected families and mitigating the broader eviction crisis.