TLDR: A new report by the Anti-Defamation League (ADL) reveals that generative AI video applications frequently fail to block antisemitic, extremist, or hateful prompts. The study, conducted by the ADL’s Center on Technology and Society (CTS), found that these programs generated problematic responses at least 40% of the time. OpenAI’s Sora 2 performed best among the tested models, refusing 60% of prompts, while others like Google’s Veo 3 and Hedra’s Character-3 showed significantly higher failure rates. The ADL emphasizes the urgent need for AI developers to implement stronger safeguards, including enhanced moderation teams and continuous testing against evolving hateful rhetoric.
The Anti-Defamation League (ADL) has released new research from its Center on Technology and Society (CTS) highlighting a significant vulnerability in generative artificial intelligence (AI) video applications. The report, published on Friday, October 24, 2025, indicates that these AI programs exhibit high failure rates when confronted with antisemitic, extremist, or otherwise hateful text prompts, generating problematic content at least 40 percent of the time.
The ADL’s investigation involved feeding a series of 50 antisemitic prompts into several prominent AI video generation tools, including OpenAI’s Sora 2, Google’s Veo 3, and Hedra’s Character-3 model. The objective was to assess their content moderation capabilities against racist and hateful material.
Among the tested applications, OpenAI’s Sora 2 demonstrated the most effective content moderation, successfully refusing to generate content for 60% of the problematic prompts. In contrast, Google’s Veo 3 refused only ten out of 50 prompts, and Hedra’s Character-3 model refused a mere two. An earlier version, Sora 1, refused none of the prompts.
Jonathan Greenblatt, CEO of the ADL, underscored the gravity of the situation in a statement on X (formerly Twitter), noting that “throughout history, bad actors have exploited new technologies to create antisemitic, extremist and hateful content – that’s where we find ourselves today as AI video generation becomes more sophisticated and accessible.”
This research builds upon previous ADL findings from March, which examined AI chatbot applications like OpenAI’s ChatGPT, Anthropic’s Claude, Google’s Gemini, and Meta’s Llama. That earlier report also revealed varying levels of anti-Israel and anti-Jewish biases in AI chatbot responses.
Experts point to the inherent challenge with large language models (LLMs) that power these generative AI tools: their responses often deviate from programmed rules, enabling users to bypass safeguards and create dangerous content. Despite developers implementing safety features, consistent application remains an unresolved issue. The ADL noted that some prompts initially shared with OpenAI after early Sora testing were subsequently blocked in the updated version of the tool.
To combat this growing threat, the ADL has proposed three critical policy changes for AI developers:
1. Heavy Funding for Moderation Teams: Investing significantly in human moderation teams to review and manage content.
2. Aggressive Prompt Testing: Continuously and rigorously testing prompts that involve hateful stereotypes.
3. Dynamic Keyword Updates: Regularly updating keywords and vernacular used in moderation in response to the evolving nature of extremist rhetoric.
The report emphasizes that many terms used in hateful content are obscure and unknown to engineers, necessitating collaboration with civil society groups and experts to provide timely analysis and updated language.
Further review by The Algemeiner into Sora 2’s antisemitic content revealed how users exploit the platform’s features. For instance, by searching terms like “rabbi” and “Jews,” users could find videos depicting antisemitic tropes, such as a man in a kippah immersed in money. One user, “acm156741,” created a series of videos featuring MMA star Jake Paul, first identifying him as Jewish and then associating him with antisemitic stereotypes to circumvent safety measures.
Also Read:
- Elon Musk Intensifies AI Bias Debate with Sharp Criticism of OpenAI and Rivals
- Instagram Enhances Stories with Advanced Generative AI Editing Capabilities
The West Point’s Combating Terrorism Center had previously warned in January 2024 about terrorist groups leveraging AI tools for propaganda, specifically noting the potential for ‘jailbreaking’ models to remove ethical safeguards. The ADL’s latest research confirms these concerns, highlighting the urgent need for robust and adaptable safety protocols in the rapidly advancing field of generative AI.
