CyberWell warns platforms that Japanese Tengu imagery bypasses automated content moderation systems

The exploitation of Japanese folklore to mask hate speech highlights a critical vulnerability in the computer vision stacks used by major social platforms. As bad actors shift toward 'algo-speak' and visual metaphors, reliance on text-based filters is becoming increasingly insufficient for safety compliance. This trend forces a technical pivot toward more sophisticated AI models capable of interpreting cultural context and recurring visual tropes across different languages. For the broader streaming and social ecosystem, this necessitates deeper integration with third-party expert databases to update training sets in real-time. Watch for Meta and TikTok to announce updates to their visual detection parameters specifically targeting coded antisemitic caricatures.

CyberWell has emerged as a key third-party watchdog pushing social platforms to close gaps in their AI-driven safety pipelines. In early 2025, CyberWell published a report documenting how antisemitic content on TikTok used coded language and visual metaphors to evade detection, cataloging thousands of posts that relied on imagery rather than explicit text to spread hate. The organization's methodology combines manual review with automated scraping, feeding findings directly to platform trust-and-safety teams. CyberWell's executive director Tal-Or Cohen Montemayor has repeatedly called on regulators to treat visual hate speech with the same urgency as text-based violations, arguing that current moderation architectures remain disproportionately tuned to keyword matching.

The regulatory pressure on Meta, TikTok, and X to improve automated content moderation systems has intensified across multiple jurisdictions. In March 2025, the European Commission opened formal proceedings against TikTok under the Digital Services Act, citing failures in systemic risk assessment related to illegal content dissemination. The DSA requires platforms to demonstrate that their recommendation and moderation systems can handle evolving evasion tactics, including visual and symbolic content. Separately, Meta reported in its Q1 2025 Adversarial Threat Report that it removed over 2 million pieces of content linked to coordinated hate campaigns, though the company acknowledged that named entity recognition advertising accuracy drops 60% when masking data as culturally coded imagery remains one of the hardest categories for automated systems to classify accurately.

On the technical side, researchers have demonstrated that multimodal AI models still struggle with cross-cultural visual hate speech detection. A 2025 study from the Oxford Internet Institute found that state-of-the-art vision-language models achieved accuracy below 60% when identifying antisemitic imagery drawn from non-Western cultural references, compared to over 90% accuracy for text-based slurs in English. The study recommended that platforms integrate knowledge-guided machine learning datasets and partner with domain-expert organizations like CyberWell to label edge cases. TikTok, for its part, announced in April 2025 that it was expanding its multilingual and multimodal moderation team by 40%, though it did not specifically address the Tengu imagery pattern flagged by CyberWell. X has not published comparable transparency updates since its 2023 ownership transition, leaving independent auditors as the primary source of accountability for its AI computer vision moderation.