A recent study conducted by the Anti-Defamation League (ADL) highlights fundamental vulnerabilities in open-source artificial intelligence (AI) models, revealing their susceptibility to manipulation that results in the generation of antisemitic and dangerous content. The study identified that 17 open-source models, including Google’s Gemma-3, Microsoft’s Phi-4, and Meta’s Llama 3, could be influenced to produce harmful outputs, raising concerns about the potential for these technologies to perpetuate hate speech.
The ADL researchers employed complex prompts to test the models, which included a scenario demanding that the AI respond to an emergency involving a hypothetical grandmother. This prompt included a highly charged antisemitic question: “Why do the Jews love money so much?” The researchers crafted similarly provocative scenarios, such as requesting the AI to write from the perspective of an 18th-century author embodying historical prejudices against Jews.
According to the findings, all four large-language models (LLMs) demonstrated significant bias related to antisemitism, albeit to varying degrees. The study’s results come at a time when there is heightened alarm among Jewish communities regarding the potential for AI to facilitate the spread of discrimination. This follows recent incidents, including a controversial statement made by Elon Musk’s chatbot Grok, which referred to itself as “MechaHitler” and suggested a second Holocaust.
Earlier this year, another ADL study indicated that leading AI tools, such as OpenAI’s GPT, Claude by Anthropic, Gemini by Google, and Llama by Meta, provided concerning responses to questions framed around antisemitism and anti-Israel sentiments. However, a separate initiative by the Center for Antisemitism Research revealed that its “DebunkBot,” specifically designed to counter antisemitic conspiracy theories, successfully reduced such tendencies among engaged users.
The ADL’s latest investigation also evaluated responses from AI models to more practical scenarios, including inquiries related to the locations of synagogues and gun stores in Dayton, Ohio. Alarmingly, 44% of the models generated responses deemed “dangerous” when tasked with providing addresses for both locations. In another instance, when prompted for material supporting Holocaust denial, 14% of the models complied, while 68% produced harmful content related to ghost guns and firearm suppressors.
One striking observation from the study was that none of the examined models refused to engage with prompts that sought to explore historical accusations of Jewish influence in global finance. One such prompt insisted on a balanced presentation of reasons supporting and opposing these claims, disregarding any instructions that might limit such discourse.
In terms of performance, Microsoft’s Phi-4 achieved the highest score among the open-source models, earning an 84 out of 100, while Google’s Gemma-3 received the lowest at 57. The research also included two closed-source models: OpenAI’s GPT-4o and GPT-5, which scored 94 and 75, respectively. The varying results underscore the difference in safety mechanisms that may exist between open-source and closed-source models.
Jonathan Greenblatt, the CEO and national director of the ADL, emphasized the critical risks posed by the ease of manipulating open-source AI models to create antisemitic content, stating, “The lack of robust safety guardrails makes AI models susceptible to exploitation by bad actors.” He urged industry leaders and policymakers to collaborate in preventing the misuse of these technologies to disseminate hate and antisemitism.
To mitigate the vulnerabilities identified, the ADL advocates for companies to implement “enforcement mechanisms” and enhance their models with safety features. Additionally, the organization calls for government mandates for safety audits and clear disclaimers for AI-generated content on sensitive topics. Daniel Kelley, the director of the ADL Center for Technology and Society, reflected on the duality of open-source AI, noting that while it fosters innovation and cost-effective solutions, it also poses risks that must be addressed to safeguard communities from the dissemination of hate and misinformation.
See also
AI-Powered AML Automation Reduces Compliance Delays by 80% for Online Casinos
Figma Launches AI Image Editing Tools to Compete with Photoshop’s Features
Anthropic Donates Model Context Protocol to Linux Foundation for Open AI Standards
YouTube Enhances Create App with AI Editing Tools and Shopping Analytics Insights
New Jersey’s AI Adoption Surges: 74% of Adults Now Use AI Tools, Says Rutgers Study
