Researchers Combine AI and Humans to Detect Coded Antisemitism

The Conversation published an article on May 18, 2026 by researchers at American University reporting that automated methods combined with human research teams can improve identification of coded hate speech, specifically coded antisemitism. The piece documents examples of online antecedent rhetoric tied to real-world violence and cites FBI reporting that the man who drove a truck into a synagogue outside Detroit in March 2026 posted messages saying "Israel is a cancerous/malignant growth" and "Israel is pure evil." The authors include an academic team from journalism, mathematics and computer science disciplines and disclose internal funding from American University's Signature Research Initiative.

The Conversation frames coded antisemitism as language that relies on euphemisms and in-group signals to evade literal-keyword moderation. Industry-pattern observations: content moderators and automated classifiers focused on surface lexical cues often miss such patterns because the signals are contextual, rely on evolving vocabularies, and can be multimodal across text and images. For practitioners, combining pattern-discovery algorithms with human annotation and thematic coding is a common approach to surface low-precision but high-impact indicators of coordinated or extremist rhetoric.

Industry context: The article highlights that alternative platforms, named in the piece as BitChute, GETTR, Parler, Rumble and Truth Social, sometimes apply weaker moderation, creating monitoring blind spots reported by researchers. This amplifies the operational challenge of detecting coded speech at scale and across platform boundaries. The Conversation situates the research within prior reporting that violent actors often leave an online footprint containing dehumanizing or conspiratorial language prior to attacks, increasing the relevance of cross-platform and longitudinal signal-gathering.

Observed patterns in similar projects: measures of success will include improvements to recall on subtle hate categories, demonstrable reductions in false positives for innocuous in-group language, and transparent annotation schemas that document evolving codes. Observers should watch for published datasets, annotation guidelines, and evaluation metrics from research teams, plus any adoption of hybrid pipelines by moderation teams or third-party monitoring services.

The Conversation article is descriptive about methods and examples; it reports the authors' position that hybrid human-plus-AI workflows can improve detection but does not provide a prescriptive operational playbook or vendor-specific performance benchmarks. The piece includes a content warning that it contains examples of antisemitic hate speech.