In an era of growing social and political dissent, University of Arizona scientists have created an AI tool that can help governments and law enforcement more precisely forecast and prepare for violent unrest.
Investigations of domestic terrorism in the U.S. have more than quadrupled in a decade, according to the Government Accountability Office, with anti-authority
extremism being the second-largest category of incidents. U.S. Capitol Police report that in recent years, threats against federal lawmakers have risen 400%.
Government and researchers have long worked to track political unrest online to inform policy and security. But the process has been labor- and time-intensive, requiring human monitoring of voluminous information.
ConfliBERT (a portmanteau of “conflict” and “BERT,” an AI created by Google) was developed through a $1.5 million National Science Foundation grant. Co-led by Javier Osorio, assistant professor in the College of Social and Behavioral Sciences, researchers trained the AI on materials curated for conflict and violence content, sourced from international agencies, mainstream media, the United Nations and U.S. foreign relations.
ConfliBERT far outperforms general natural language processing tools in spotting and analyzing red flags for conflict and violence, with applications for both predicting unrest and learning from past incidents.
The researchers also developed a novel training method using less than 34 GB of data. That model and ConfliBERT’s publicly available code dramatically lower barriers for communities around the world where developing similar tools would be cost-prohibitive.