Researchers Show Hybrid Approach Catches Coded Antisemitic Speech Online
Researchers at American University reported that combining automated detection with human expertise improves identification of coded antisemitic language on social media. The work, published in The Conversation on May 18, 2026, documents how euphemisms and in-group terminology allow hateful speech to evade standard moderation filters.
The research team cited FBI reporting on a March 2026 attack outside Detroit, where a man drove a truck into a synagogue after posting messages including "Israel is a cancerous/malignant growth" and "Israel is pure evil." The posts exemplify how coded speech uses indirect language to mask intent while remaining visible to those familiar with the terminology.
Why Keyword Filters Fail
Content moderation systems relying on literal keyword matching miss coded hate speech because the signals are contextual and evolving. A phrase flagged on one platform may shift meaning or vocabulary on another, and hateful messages often blend text with images in ways automated systems struggle to parse together.
Alternative platforms including BitChute, GETTR, Parler, Rumble, and Truth Social apply weaker moderation standards, creating blind spots for researchers and moderators trying to track coordinated activity across the internet.
The Hybrid Workflow
The researchers advocate pairing pattern-discovery algorithms with human annotation and thematic coding. Moderators and data scientists review algorithmic findings to identify subtle indicators of extremist rhetoric, then feed those insights back into detection systems to improve accuracy.
This approach addresses two common problems: algorithms flag too many false positives when applied to in-group language, while humans alone cannot scale to monitor millions of posts. The combination trades speed for precision.
What Matters for Implementation
Success metrics include improved recall on subtle hate categories, fewer false positives on innocuous language, and transparent documentation of how coded terms are classified. Research teams publishing datasets and annotation guidelines allow others to build on the work rather than starting from scratch.
Moderation teams and third-party monitoring services should expect to see published evaluation metrics showing how hybrid systems perform against baseline keyword approaches. Operational playbooks remain limited - this research describes the problem and approach but does not provide vendor-specific benchmarks or step-by-step deployment guides.
The American University team received internal funding through the university's Signature Research Initiative. The article includes a content warning for examples of antisemitic speech.
Learn more about AI Research Courses and Generative AI and LLM Courses to deepen your understanding of pattern detection and human-in-the-loop systems.
Your membership also unlocks:
