CyberWell released new guidance for social media companies earlier this month on improving artificial intelligence content moderation to better target antisemitism, while protecting free expression.
The guidance was submitted as a public comment this month in response to a call for input from Meta’s Oversight Board on the use of large language models (LLMs) in content moderation, addressing both the potential efficiencies of automated review and the persistent challenges of interpreting nuanced content at scale.
“AI gives platforms an unprecedented opportunity to improve the enforcement of their own rules at scale, moving beyond slur 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 draw the line between anti-Jewish scapegoating – a classic form of antisemitism– and misinformation and disinformation campaigns that receive different moderation treatment,” Cohen Montemayor added.
In its submission, Cyberwell detailed several technical challenges that automated systems frequently fail to catch.
AI safety filters miss posts with hidden antisemitism
The report noted that online bad actors use modern AI tools to create tricky new versions of hate speech, meaning computer programs have to understand hidden meanings instead of just searching for banned words.
For example, automated safety filters often miss posts that combine harmless-looking images with hidden insults, such as an AI-generated video showing Adolf Hitler dancing in front of the Auschwitz concentration camp alongside text celebrating the mass murder of Jewish people.
Posts like these also frequently slip past automated checks by using innocent symbols in place of words, such as using a juice box emoji because it sounds similar to the word Jews when spoken.
The research revealed that internet users often try to hide hateful content by pretending they are just jokes, memes, or educational material.
For instance, automated filters often struggle to catch posts that mock the Holocaust by wrapping denialist ideas in absurd scenarios, such as a video calculating how difficult it would be to bake six million pizzas in five years.
Similarly, users frequently attach misleading labels, such as comedy or humor, or educational disclaimers to posts containing explicitly anti-Jewish stereotypes and hate speech, tricking basic computer filters into thinking the content is harmless.
The report also documented non-English hurdles, pointing to Arabic-language content that employs subtle letter modifications and punctuation dots between words to evade automated moderation while amplifying violent messaging and praising attackers.
Content moderation struggles to review hate spread in comment sections
Furthermore, harmful content frequently hides in comment sections where automated filters struggle because they evaluate text in isolation rather than looking at the original post.
For example, under a news post reporting on Israeli embassy workers murdered in Washington, DC, users bypassed filters and praised Hitler by referring to him as “the half moustache guy,” while other commenters used the phrase “promised 3,000 years ago” to mock and attempt to delegitimize the Jewish historical connection to the Land of Israel.
While social media platforms might not consider that specific phrase a policy violation when used in standalone posts, its appearance in comments responding to reports of antisemitic violence constitutes a violation used to mock victims and justify violence against them.
The core problem enabling this is that automated moderation tools fail to analyze how comments relate to the original post, missing how harmless-sounding phrases become toxic when used to target victims.
The report also noted that harmful content slips past automated tools by hiding inside account usernames and display handles and by circulating within audio clips that utilize non-English lyrics or varying lengths to escape detection.
The submission underscored severe inconsistencies in enforcement for high-reach accounts, demonstrating that identical policy-violating posts shared by influencers with large followings remain active even after repeated appeals, whereas ordinary user reposts are instantly removed.
CyberWell's five core recommendations for tech companies
To resolve these vulnerabilities, CyberWell issued five core recommendations for tech companies deploying LLMs for safety enforcement.
Platforms should collaborate with independent subject-matter experts to refine model training and evaluation using real-world annotated datasets.
Companies must ensure consistent moderation across platform surfaces and multi-modal formats, including comments, usernames, images, GIFs, videos, and audio.
Platforms need to improve multilingual and cultural accuracy by testing models against slang, dialects, irony, and coded language.
Tech firms should establish clear escalation pathways for human review during rapidly evolving global events.
Companies must increase public and user transparency through regular reporting on model performance, error rates, enforcement outcomes, and content-specific removal explanations.
“The future of effective trust and safety will depend on combining advanced technology with reliable data and real-world expertise,” Cohen Montemayor said.
“Our working relationship 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 turning its expertise in how antisemitism and online hate evolve into practical datasets, evaluation tools, and technological solutions that can scale, helping 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.