CyberWell sends social media companies recommendations to improve antisemitism moderation
The guidance was submitted as a public comment this month in response to a call for input from Meta's Oversight Board regarding the use of large language models in content moderation.
CyberWell recently provided guidelines to digital platforms to enhance their artificial intelligence moderation of online content, especially concerning antisemitic material. The recommendations were submitted to Meta's Oversight Board in response to the company's call for input on the use of large language models in content moderation.
According to CyberWell's CEO, Tal-Or Cohen Montemayor, AI offers unprecedented opportunities to enforce rules at scale while moving beyond simple word lists to context-based decisions. However, he stressed the importance of expert input to ensure the systems can effectively identify antisemitic content without stifling free expression.
The guidance highlighted several challenges faced by automated moderation systems. For instance, AI safety filters often fail to detect antisemitic content hidden in seemingly innocuous contexts, such as combining harmless-looking images with coded insults or using innocent symbols to represent harmful language. Moreover, users frequently attempt to disguise hate speech as jokes, memes, or educational material by attaching misleading labels like comedy or humor.
This tactic can deceive basic filters, especially when posts are accompanied by disclaimers that make them appear harmless.
Content moderation also struggles with identifying antisemitic messages in comment sections, where automated tools evaluate text in isolation rather than in the context of the original post. For example, commenters might praise historical antisemitic figures or mock victims by using phrases that seem harmless in isolation but are toxic when used to target Jewish communities.
Similarly, antisemitic content can evade detection by hiding inside account usernames, display handles, or audio clips featuring non-English lyrics or varying lengths.
To address these issues, CyberWell proposed five core recommendations for social media firms using large language models for safety enforcement. These include collaborating with subject-matter experts to refine model training and evaluation using real-world datasets, ensuring consistent moderation across all platform surfaces and formats, improving multilingual and cultural accuracy, establishing clear escalation pathways for human review during rapid events, and increasing transparency through regular reporting on model performance and enforcement outcomes.
Written by urgent.news from Jerusalem Post's reporting — not their text. Machine-written — may contain errors; check the original before relying on it.
