CyberWell Calls for Stronger AI Moderation of Antisemitism on Social Platforms

CyberWell has issued new recommendations for social media companies seeking to improve artificial intelligence-based content moderation to better identify antisemitism while protecting free expression.

The guidance was submitted as a public comment this month in response to Meta’s Oversight Board request for input on the use of large language models in content moderation. It addressed both the potential efficiencies of automated review and the persistent difficulty of interpreting nuanced material at scale.

“AI gives platforms an unprecedented opportunity to enforce their own rules at scale, moving beyond static word lists toward context-based decisions that can help reduce human error. However, expert input is crucial to ensuring these systems can effectively address specific forms of online hate,” CyberWell founder and CEO Tal-Or Cohen Montemayor said.

“Antisemitism constantly evolves through coded language, symbols, cultural references, and forms of ‘algospeak’ designed to evade automated detection. AI moderation systems must draw on contextual, linguistic, and subject-matter expertise to recognize the antisemitic message, not just the words, while protecting legitimate expression. Social media platforms have also struggled to distinguish anti-Jewish scapegoating—a classic form of antisemitism—from misinformation and disinformation campaigns that receive different moderation treatment,” Cohen Montemayor added.

In its submission, CyberWell outlined several technical challenges that automated systems frequently fail to identify.

AI safety filters overlook posts with hidden antisemitism

The report noted that online bad actors are using modern AI tools to produce new and difficult-to-detect versions of hate speech. Moderation systems must therefore understand implied meanings rather than merely search for prohibited words.

For example, automated safety filters often miss posts that combine apparently innocuous imagery with hidden insults. One example was an AI-generated video showing Adolf Hitler dancing in front of the Auschwitz concentration camp alongside text celebrating the mass murder of Jewish people.

Such posts can also evade automated checks by replacing words with seemingly innocent symbols. One cited example used a juice box emoji because its pronunciation resembles the word “Jews.”

The research found that internet users often disguise hateful material as jokes, memes, or educational content.

For instance, automated filters may struggle to identify posts that mock the Holocaust by presenting denialist ideas through absurd scenarios, such as calculating how difficult it would be to bake six million pizzas in five years.

Users also frequently attach misleading labels such as “comedy,” “humor,” or “educational” to posts containing explicit anti-Jewish stereotypes and hate speech, prompting basic automated filters to treat the material as harmless.

The report also documented challenges involving non-English content, including Arabic-language posts that use subtle character modifications or punctuation dots between words to evade moderation while amplifying violent messaging and praising attackers.

Context-aware review needed for hate spread in comment sections

Harmful content also frequently appears in comment sections, where automated filters struggle because they assess each comment in isolation rather than considering the post being discussed.

For example, beneath a news post about Israeli embassy workers murdered in Washington, D.C., users bypassed filters by referring to Hitler as “the half moustache guy.” Other commenters used the phrase “promised 3,000 years ago” to mock and delegitimize the Jewish historical connection to the Land of Israel.

Although social media platforms may not consider that phrase a policy violation when used in a standalone post, its use in comments responding to reports of antisemitic violence can constitute an abuse that mocks victims and justifies violence against them.

A central problem is that automated moderation tools do not consistently analyze how a comment relates to the original post. As a result, they miss cases in which phrases that appear harmless in isolation become threatening when directed at victims.

The report also noted that harmful content can evade automated tools by appearing in account usernames and display handles or circulating in audio clips featuring non-English lyrics and varying lengths that make detection more difficult.

The submission highlighted significant inconsistencies in enforcement for high-reach accounts. It said identical policy-violating posts shared by influential users with large followings can remain active despite repeated appeals, while reposts by ordinary users are removed immediately.

CyberWell’s five core recommendations for technology companies

To address these vulnerabilities, CyberWell issued five core recommendations for technology companies deploying large language models for safety enforcement.

Platforms should work with independent subject-matter experts to refine model training and evaluation using real-world, annotated datasets.

Companies should ensure consistent moderation across platform surfaces and multimedia formats, including comments, usernames, images, GIFs, videos, and audio.

Platforms need to improve multilingual and cultural accuracy by testing their models against slang, dialects, irony, and coded language.

Technology firms should establish clear escalation pathways for human review during rapidly evolving global events.

Companies should also strengthen public and user transparency through regular reporting on model performance, error rates, enforcement outcomes, and content-specific explanations for removals.

“The future of effective trust and safety will depend on combining advanced technology with reliable data and real-world expertise,” Cohen Montemayor said.

“Our work with Meta and other major social media platforms demonstrates what that can look like in practice. The potential to scale effective moderation through AI makes it even more important to deepen relationships with trusted experts and partners who understand technology and digital policy,” she explained.

“CyberWell is translating its expertise on the evolution of antisemitism and online hate into practical datasets, evaluation tools, and technological solutions that can scale. These tools can help platforms optimize their moderation systems to identify policy violations, including nuanced and coded forms of hate, and enforce their rules more accurately and consistently,” Cohen Montemayor said.

Source link