Government Social Media Monitoring with AI

Public posts can be searched manually or with tools that filter by keyword, account, date, location or topic. Some tools translate, classify or link posts. A model score should not be treated as proof of intent, identity or a threat. Entity resolution attempts to determine whether records or aliases refer to the same entity; a false link can attach another person’s speech to the wrong file. The U.S. State Department has requested social-media identifiers on visa application forms from most applicants worldwide since 2019, subject to limited exceptions. In 2025 it announced online-presence review and public-profile instructions for F, M and J student/exchange applicants. By September 2026, review also covered H-1B/H-4 and additional categories; State announced I, TN and TD would join effective October 1, after this review date. A 2019 identifier request is distinct from instructions to make a profile public for review. In 2016, the ACLU of California reported from records requests to 63 agencies that Geofeedia marketed monitoring of protests and activists, including Ferguson-related material. Platforms restricted its access; this was not proof every department used it. Automated classifiers can produce many false alerts when the underlying event is rare. If one post in a million is truly a threat and a classifier has 100% sensitivity plus a 1% false-positive rate, one million posts yield about one true positive and roughly 10,000 false positives. This illustrates base rates; it is not a measured agency system. In a 2019 study of several toxicity datasets, Sap and colleagues found AAE markers correlated with toxicity labels and models could label AAE or self-identified Black users’ tweets offensive up to twice as often in the evaluated setups; this result is limited to those datasets and tasks. Monitoring can also chill lawful expression when people fear that ordinary posts may be misinterpreted or linked to an official file.