The ADL tested 25,000 conversations.
Here's what to do about it.
In January, the ADL published the first standardized evaluation of how the major AI models handle antisemitic content. More than 25,000 conversations. Thirty-seven categories. Six models: ChatGPT, Claude, DeepSeek, Gemini, Grok, Llama.
The finding, in their words: all six could benefit from improvement. Performance varied across the board. And the models were better at refuting the old tropes, Jews controlling the media, Holocaust denial, than at handling anti-Zionist and extremist content.
Ten days ago they published something sharper. They ran the same kind of test in Persian. Their Senior Director's words:
Same question. Different language. Different answer.
We are not here to repeat that work. The ADL has a research institute and we do not, and their evidence is better than anything we could produce. If you want the numbers, go and read them. We will link to them every time.
We are here for the part nobody is doing.
Because knowing the models are broken does not help the person who just got a bad answer and wants to do something about it. Reports go to AI companies, policymakers, and educators. Nobody is handing anything to the person standing there with their phone.
That is what we build. The tools on this site are live now and they are free. Anyone can use them. No expertise, no login, no waiting for an institution to act.
And we are building the platform to do this at scale.
See it for yourself
Copy a prompt, paste it into any AI, and watch what happens. Two minutes.
Get the promptsADL AI Index, launched 28 January 2026. "Six Leading AI Models Show Varied Ability to Detect and Counter Antisemitism and Extremism," ADL, 28 January 2026. "AI Chatbots Consistently Fail to Identify and Reject Antisemitism in Persian," ADL, 8 July 2026. Quote from Daniel Kelley, Senior Director, ADL Center for Technology and Society.