A new report from the nonprofit CyberWell, published last month, documents how generative AI tools were used to produce and distribute antisemitic content at a scale that stretched the limits of platform enforcement - accumulating 30 million views and more than 2.8 million engagements across TikTok, Instagram, YouTube, Facebook, and X in the 13-month period from January 2025 to February 2026.

CyberWell, an independent tech-based nonprofit that monitors antisemitic content online using the International Holocaust Remembrance Alliance's working definition of antisemitism, assembled a dataset of 307 verified posts. The report, released on May 18, 2026, identifies which platforms removed content, which did not, and how AI tools including OpenAI's Sora, Google's Veo, X's Grok, and the audio generator Suno were used in the production process. The findings carry direct implications for brand safety and platform governance - areas that marketing professionals track closely given the continued expansion of programmatic advertising across social media inventories.

Scale and timing of the data

The dataset is the most widely viewed CyberWell has ever compiled. According to CyberWell, it recorded precisely 30,345,631 views, 2,664,979 likes, 146,126 shares and retweets, and 43,658 comments. Those numbers come from just 307 posts.

What makes the timeline particularly notable is the concentration of activity. According to CyberWell's report, 98.4 percent of the content identified was posted from June 2025 onward. The period from January to May 2025 yielded only a small number of isolated examples. The sharp inflection point aligned with Israel's 12-day war with Iran in June 2025, when CyberWell identified a rapid spike in the cross-platform dissemination of AI-generated antisemitic content. Shortly after that conflict, trends including the "promised 3,000 years ago" framing and Holocaust-related mockery formats became more widespread on major platforms.

This pattern - where geopolitical events trigger surges in AI-produced hate content - had been observed before, though at a smaller scale. According to CyberWell, AI-generated antisemitic content existed prior to October 7, 2023, but the aftermath of the Hamas attacks gave it greater traction and a more visible presence on social platforms. That earlier period prompted CyberWell to begin monitoring it systematically.

Platform distribution

According to CyberWell, video-based platforms dominate the dataset. TikTok, Instagram, and YouTube together account for 79.2 percent of all posts identified during the research period. TikTok holds the largest single share at 35.8 percent of posts. That concentration on short-form video reflects both the production capabilities of modern AI tools and the algorithmic amplification dynamics of video-first platforms.

Instagram's numbers tell a different story about engagement versus volume. The platform accounts for just 24.8 percent of posts in the dataset but generates 64.9 percent of total engagement. According to CyberWell, Instagram was responsible for 62.43 percent of total views and 92 percent of overall engagement within Meta's portfolio. That disparity - a platform producing roughly a quarter of the posts but nearly two-thirds of the interactions - points to the role algorithmic recommendation plays in amplifying content once it gains initial traction.

X sits at the opposite end of the volume spectrum, representing only 8.1 percent of the dataset. The platform's text-based architecture limited the volume of AI-generated multimedia captured in this analysis. However, X's figures in this dataset represent a smaller slice than CyberWell's 2025 annual report, where X accounted for 37 percent of monitored antisemitic content.

Enforcement gaps and removal rates

Platform removal rates reveal significant divergence. According to CyberWell, 63.5 percent of all 307 posts in the dataset were eventually removed for policy violations. That figure is higher than the 52.53 percent removal rate documented in CyberWell's 2025 annual report across all antisemitic content - but the distinction matters less than the delays involved.

TikTok recorded the highest removal rate in the dataset at 88.2 percent. Meta removed 67 percent of AI-generated antisemitic posts on Instagram and Facebook. YouTube's removal rate was 28.1 percent. X recorded the lowest at 20 percent.

According to CyberWell, TikTok and Meta's comparatively higher removal rates likely reflect their more explicit policy frameworks addressing AI-generated content. Both platforms integrate AI-generated content considerations directly into their existing hate speech and community standards frameworks. TikTok's Integrity and Authenticity policy explicitly states that any content breaking community guidelines - "including those on impersonation, misinformation, and hate speech, even if it's AI-generated" - is prohibited. Meta similarly maintains that content violating its Community Standards will be removed regardless of whether it was created with AI tools.

YouTube and X operate under different frameworks. According to CyberWell, neither platform maintains policy provisions as explicitly focused on AI-generated content as TikTok's. YouTube's community guidelines address harmful stereotypes and denial of violent events through broader provisions rather than dedicated AI-content rules. X introduced a new provision on March 3, 2026, addressing AI-generated content related to armed conflict - creators who post such images without disclosure face a 90-day suspension from Creator Revenue Sharing, with permanent suspension for repeat violations.

The YouTube CEO's defence of AI moderation in December 2025 came amid criticism from creators about inconsistent automated enforcement - a context that matters here, because the same automated systems responsible for monetization enforcement also handle content removal. The AJC-CyberWell partnership article on PPC Land noted that a snapshot of CyberWell's open database from May 2026 showed 13,534 examples of logged antisemitic content across platforms, with entries as recent as May 11, 2026.

The removal rate figures mask the central timing problem. According to CyberWell, even content that platforms eventually removed frequently remained online long enough to accumulate hundreds of thousands or millions of views before enforcement occurred. High eventual removal rates do not necessarily indicate effective enforcement when content that glorifies or incites violence circulates at scale during the delay.

Violent content spreads fastest

CyberWell's engagement data reveals a specific pattern. According to the report, content that glorifies, justifies, or calls for violence against Jews accounts for 33.22 percent of posts in the dataset, yet generates 33.66 percent of total views and 41.41 percent of overall engagement - a disproportionate share of likes, shares, and retweets relative to post volume.

Notably, according to CyberWell's research, violent antisemitic content is more than twice as likely to appear in AI-generated content than in user-generated antisemitic content on social media. That comparison draws on CyberWell's separately compiled 2025 State of Online Antisemitism dataset. The implication is that generative AI tools are being used selectively for the most extreme material.

According to CyberWell, 87.3 percent of posts that explicitly glorified, justified, or called for violence towards Jews were eventually removed across all platforms - a higher rate than the average for the dataset. Meta hosted the most AI-generated content consistent with IHRA Example 1, which covers calls for or justification of the killing or harming of Jews. Despite the higher eventual removal rate for violent material, engagement data shows these posts often circulated widely before platform action.

The AI tools involved

According to CyberWell's report, AI models including Sora (OpenAI), Veo (Google/DeepMind), Grok (X), and Suno significantly lowered the technical barrier to producing multimedia antisemitic content. These tools allowed users to generate synthetic video, audio, and realistic imagery from text prompts and then distribute the results at speed across multiple platforms simultaneously.

Sora, in particular, featured in a substantial share of the content CyberWell identified before OpenAI discontinued the tool in March 2026. According to CyberWell, the discontinuation reflected ongoing challenges in preventing abuse of advanced generative AI systems following widespread misuse to spread problematic content. Sora was used to produce fabricated Disney-Pixar-style trailers mocking the Holocaust - formats designed to appear child-friendly while embedding antisemitic narratives. These videos circulated alongside content promoting violence and content that normalized the grooming or sexualization of minors on the same platform feeds.

Veo3, Google's video generation model released in 2025, was identified by CyberWell in a specific trend starting in July 2025. Users leveraged Veo3 to transform the "promised 3,000 years ago" textual trope - which had previously circulated in comment sections - into video content. The transformation from text to video gave the narrative new reach, and the content subsequently reappeared in comment sections mocking victims of violent antisemitic attacks. The progression illustrates how coded antisemitic narratives migrate from text-based spaces into multimedia formats once users adapt them for AI generation.

Grok, X's AI tool, appeared in several specific examples documented by CyberWell, including a post that used Grok's image-generation capabilities to reinforce the blood libel narrative in the context of the 2025 New York City mayoral election. In that case, a user created an AI-generated video depicting distressing fabricated imagery and paired it with a caption invoking a historical antisemitic accusation.

Suno, an audio generation tool, was used to produce "Boom, Boom, Tel Aviv", an AI-generated song that circulated widely across all major platforms during Israel's conflict with Iran in June 2025. The song's lyrics celebrated and justified violence against Jews. According to CyberWell, this single piece of content and its variants account for a significant share of the dataset's engagement figures, demonstrating how a single user-generated AI audio clip can be replicated and amplified across platforms by multiple accounts.

Youth-oriented formats

A recurring pattern in CyberWell's data involves packaging antisemitic content in formats specifically designed to reach younger audiences. According to CyberWell, these included fabricated Disney-Pixar-style movie trailers, gaming-related audio clips tied to platforms such as Roblox, Minecraft, Clash of Clans, and Clash Royale, and short-form content styled to match youth viral trends on TikTok.

One example from the dataset - an AI-generated audio clip tied to a conspiracy theory about 9/11 - was associated with nearly 60,000 TikTok videos and generated over 50 million views across the first 50 posts alone, according to CyberWell. The audio was embedded repeatedly alongside gaming-related content, distributing the conspiracy theory within youth-oriented spaces where it would not appear out of place.

According to CyberWell, this type of content appeared alongside posts promoting radicalization, violence, and - in some cases - the sexualization or grooming of minors. The co-occurrence on the same platform feeds is notable: AI-generated antisemitic content was not isolated in niche channels but appeared within broader content streams accessible to young users.

Coded language and evasion

According to CyberWell's analysis, the AI-generated antisemitic content that remained online after enforcement actions was typically implicit rather than explicit. Posts relied on coded language, humor, disclaimers such as "#satire" and "for educational purposes", mockery, or stereotypical depictions to obscure intent and present harmful material as neutral or comedic.

Three primary narrative categories dominated the dataset. First, depictions of Jews as greedy or money-obsessed appeared in 33.2 percent of posts. This narrative used AI visuals to reinforce longstanding stereotypes through fabricated imagery - characters depicted stealing coins, animations tied to Jewish cultural music, and visual representations of classic antisemitic caricatures given AI-generated movement and audio.

Second, Holocaust-related hate speech appeared in 21.5 percent of posts, most frequently on Meta's platforms. This category included both explicit denial and content that mocked victims through coded references - for example, using "juice box" emojis or the phrase "6 million pizzas" to refer to Holocaust victims in ways that evaded automated detection tied to explicit slurs.

Third, event-driven violent rhetoric against Jews appeared in 21.2 percent of posts, concentrated on TikTok. These posts responded to real-world events - the October 7 aftermath, the Iran conflict, the December 2025 Bondi Beach Hanukkah shooting in Sydney, Australia - with AI-generated content that fabricated events, celebrated violence, or denied the occurrence of antisemitic attacks.

According to CyberWell, IHRA Example 2 - stereotypical and conspiratorial depictions of Jews - accounted for 75.6 percent of the entire dataset. IHRA Example 1, covering calls for or justification of violence, accounted for 33.2 percent. Ten of the IHRA definition's eleven examples appeared somewhere in the 307 posts.

Governance and the policy gap

CyberWell's report assessed not only platform enforcement but also the regulatory frameworks that govern how AI companies and governments address AI-generated hate speech. The findings point to a structural gap: AI governance frameworks and social media platform regulation continue to operate as separate domains, with limited coordination on how generative AI systems interact with platform recommendation and content moderation systems.

According to CyberWell, major AI companies including OpenAI, Google, and Suno publish safety and usage policies that restrict hate speech and incitement - but these policies rely on broad terms such as "hatred" and "hate speech" without specifying how harms manifest in practice, particularly through implicit or coded forms. Transparency reporting from AI companies on how these guardrails are implemented and updated remains limited.

The EU's AI Act, adopted in 2024, established a risk-based framework classifying AI systems by their potential impact on safety and fundamental rights. Australia adopted a National AI Plan in 2025 with guidance for AI developers and a Voluntary AI Safety Standard. The United States has maintained a more conservative regulatory approach focused on protecting innovation and free expression. According to CyberWell, none of these national frameworks adequately addresses the specific intersection between AI model capabilities and social media platform amplification of hateful content.

This regulatory gap is directly relevant to digital advertising. When platforms carry AI-generated content that violates their own policies - particularly violent or hateful content with high engagement - the inventory surrounding that content affects brand safety measurements. PPC Land's coverage of DSA enforcement actions in October 2025 documented preliminary Commission findings that TikTok and Meta were in breach of transparency obligations. The European Commission's €120 million fine against X in December 2025 for DSA violations added pressure on platforms to improve content governance mechanisms.

"Artificial intelligence has fundamentally changed the scale and speed at which antisemitism can be produced and distributed online," said CyberWell CEO and founder Tal-Or Cohen Montemayor. "Generative AI now allows bad actors to industrialize hate, producing high-impact content that can reach millions, with enforcement often coming only after it has already been widely amplified."

Cohen Montemayor added: "Platforms must move beyond disclosure and invest in systems that identify harmful narratives at scale, including those embedded in audio, visuals and coded formats that evade traditional detection."

The report concludes with recommendations directed at three groups. For platforms, CyberWell calls for explicit policies that apply hate speech and violence rules equally to AI-generated content, expanded detection across multimedia formats including audio and video, algorithmic safeguards to limit early-stage amplification of violent material, and improved cross-account tracking using embedded watermarks. For AI companies, the report calls for clearer definitions of protected categories, expanded red teaming focused on implicit and coded hate content, and greater public transparency around safety guardrails. For policymakers, CyberWell recommends mandatory disclosure requirements for AI companies, closer coordination between AI governance and platform regulation, measures to address youth exposure to AI-generated harmful content, and standardized methods for identifying AI-generated media.

"Generative AI is a powerful technology, but it is also being weaponized at scale. By strengthening automated detection, investing in competent and transparent human moderation, auditing training data and partnering with specialized external stakeholders, platforms and AI developers can address the complex and fast-evolving forms of online hate through sustained collaboration between technology companies, policymakers and expert partners," Cohen Montemayor said.

Timeline

  • October 7, 2023 - Hamas attacks on Israel prompt CyberWell to begin monitoring AI-generated antisemitic content as a regular part of its research methodology; content glorifying the attacks, including AI-generated images depicting Hamas paragliders, circulates on platforms.
  • Early 2025 - Isolated examples appear when users respond to the release of John F. Kennedy Assassination Records with AI-generated songs attributing Jewish responsibility, such as "Kennedy Killers".
  • January 2025 - CyberWell's monitoring period for this dataset begins; the dataset will eventually span through February 2026.
  • May 13, 2026 - AJC and CyberWell join forces to track antisemitism across social platforms, combining survey data from American Jewish experience with CyberWell's real-time platform detection system.
  • June 2025 - Israel's 12-day war with Iran triggers a sharp spike in AI-generated antisemitic content; CyberWell identifies the AI-generated song "Boom, Boom, Tel Aviv" (created with Suno) and flags it for removal; 98.4 percent of all dataset content is posted from this month onward.
  • July 2025 - Users begin leveraging Veo3 to transform the text-based "promised 3,000 years ago" antisemitic trope into video content; the trend escalates from comment sections to multimedia formats.
  • October 24, 2025 - European Commission preliminarily finds TikTok and Meta in breach of DSA transparency obligations.
  • December 2025 - European Commission issues a EUR 120 million fine against X for DSA violations, the first non-compliance decision under the DSA.
  • December 2025 - The Bondi Beach Hanukkah shooting in Sydney, Australia prompts a new wave of AI-generated content using the "promised 3,000 years ago" trope to dehumanize victims; CyberWell documents this as Phase 3 of the trend's evolution.
  • December 14, 2025 - YouTube's CEO defends AI moderation practices amid creator criticism about inconsistent automated enforcement.
  • February 2026 - CyberWell's monitoring period for this dataset ends.
  • March 2026 - OpenAI discontinues Sora as part of a broader business decision, following widespread misuse to spread problematic content.
  • March 3, 2026 - X introduces a new provision requiring disclosure for AI-generated images of armed conflict, with suspension from Creator Revenue Sharing for 90 days for non-compliant posts.
  • May 18, 2026 - CyberWell publishes the AI-Generated Antisemitism report documenting 307 posts, 30 million views, and 2.8 million engagements across five major platforms between January 2025 and February 2026.

Summary

Who: CyberWell, an independent tech-based nonprofit founded by Tal-Or Cohen Montemayor and based in Tel Aviv, published the report. The subjects of the research are the major social media platforms - TikTok, Instagram (Meta), YouTube, Facebook (Meta), and X - along with the AI model providers whose tools were used to generate the content: OpenAI (Sora), Google/DeepMind (Veo), X (Grok), and Suno.

What: CyberWell compiled a dataset of 307 verified AI-generated antisemitic posts spanning 13 months and analyzed their distribution, engagement, and removal rates across platforms. The dataset recorded over 30 million views and 2.8 million user interactions. Violent antisemitic content - content calling for or justifying harm to Jews - accounted for 33.2 percent of posts but generated 41.4 percent of total engagement, and is more than twice as likely to appear in AI-generated content than in user-generated antisemitic content. Platform removal rates ranged from 88.2 percent on TikTok to 20 percent on X.

When: The data covers January 2025 through February 2026, with 98.4 percent of the content posted from June 2025 onward. The report was published on May 18, 2026.

Where: The content appeared on TikTok, Instagram, YouTube, Facebook, and X. The report was released by CyberWell, headquartered in Tel Aviv.

Why: CyberWell's report addresses a gap in existing research on AI-generated content: while studies have examined AI safety guardrails and regulatory frameworks, less attention has been paid to how AI-generated antisemitic content actually manifests on social media and how platforms respond. The findings matter beyond the immediate subject because platform enforcement gaps, algorithmic amplification of violent content, and the interaction between AI tools and social media recommendation systems have direct implications for the safety of advertising environments, the exposure of young users to radicalizing content, and the effectiveness of national and EU regulatory frameworks governing AI and online content.