ADL and the AI Labs: What Greenblatt Claimed About Fixing Antisemitism in Models, and What Is Verified

Posts spreading on October 13, 2026 say that ADL CEO Jonathan Greenblatt, speaking from Israel, announced the Anti-Defamation League has worked with leading AI developers to eliminate antisemitic output from their models. The names attached are OpenAI, Anthropic, Google, Meta and X, and the wording is that the ADL helped them tackle the issue "right in the source code." Critics replied that the ADL decides what counts as extremism, pointing to groups such as Turning Point USA appearing in its Glossary of Extremism.

We could not find a transcript, ADL release or lab statement for the October remark. What we can verify is a body of ADL testing and a public statement from March 2026 that the group is in regular contact with AI companies. This post lays out what is claimed, what is documented, and what is missing, so you can judge how far the headline goes. For broader context on bias and values in models, see our coverage of AI models' political compass study and what bias in AI means.

Claimed vs. verified at a glance

What the ADL has published about AI

A clipboard with a shield and checks, representing ADL AI antisemitism testing of chatbots

March 2025: four chatbots, 34,400 responses

The ADL press release of March 25, 2025 described what it called the most comprehensive evaluation to date of anti-Jewish and anti-Israel bias in GPT, Claude, Gemini and Llama. Each model was queried 8,600 times for 34,400 responses. The ADL reported measurable bias in all four, with Meta's Llama the lowest scoring on bias and reliability, and said GPT and Claude showed significant anti-Israel bias on questions about the Israel-Hamas war. It also reported that models refused Israel-related questions more often than other topics. The research was done in partnership with Builders for Tomorrow, a venture philanthropy and research organization the release describes as focused on combating anti-Jewish and anti-West ideologies. That funding and framing context is worth knowing when weighing the method.

December 2025: 17 open-source models

A second ADL study tested 17 open-source models, including Google's Gemma-3, Microsoft's Phi-4 and Meta's Llama 3, plus GPT-4o and GPT-5 as closed-source comparisons. It built a guardrail score from refusal rates, rates of evading safety rules, and rates of harmful content. Phi-4 scored best at 84 out of 100 and Gemma-3 lowest at 57. The ADL reported dangerous responses in 44 percent of cases when asked for addresses of synagogues and nearby gun stores in Dayton, Ohio, and harmful content for 68 percent of prompts about ghost guns and firearm suppressors.

January 2026: the AI Index

The Algemeiner's report on the ADL AI Index says the index ranked six models on detecting antisemitism, anti-Zionism and extremism: Claude 80, ChatGPT 57, DeepSeek 50, Gemini 49, Llama 31 and Grok 21. Testing ran between August and October 2025, with more than 25,000 chats across 37 sub-categories, scored by human and AI evaluators. The ADL's own words, per that report: every model "showed gaps," and the index was meant as "a roadmap for AI companies to improve their detection capabilities." We could not open the ADL's own page for the index, so the scores here are as reported by the Algemeiner.

What Greenblatt has said about working with labs

In his March 16, 2026 State of Hate speech, Greenblatt said the ADL's Center on Antisemitism Research is working with its Center for Technology and Society "to understand how the major AI platforms are building their products," and that "We are in regular contact with the AI companies -- and I think we are making a difference."

His example was OpenAI's video tool: "last year, we showed OpenAI how its new video generation AI tool Sora made it easy to create antisemitic videos," and nine months later, he said, Sora 2 "refuses to generate videos around several of the concepts we specifically discussed with them."

That is the strongest documented description of the relationship, and it is a testing-and-feedback relationship, not a joint engineering one. It is also the ADL's own account. OpenAI has not, in sources we found, confirmed that its changes came from ADL input. Labs routinely say they work with outside experts, and a refusal change in a product is rarely attributable to one party.

Reading the "source code" claim

A padlock with a thin pick, standing for how outside testers probe model guardrails rather than edit code

The viral wording says the ADL guided labs in tackling the issue within the source code. Technically, that is unlikely to be literal. A language model's behavior comes from its training data, post-training such as reinforcement learning from human feedback, the system prompt, and layers of classifiers and moderation around the model. Outside groups can contribute test sets, definitions and red-team findings that labs use for any of those. Few outside parties edit the code or weights.

Understanding that matters for judging what could change. If the ADL's contribution is evaluation data and guidance, the lab still decides what to adopt, and the effect shows up as different refusals and different classifications, not as a visible patch. That is consistent with the Sora example. It also means the claim is hard to check from the outside: the only way to see an effect is to run the same prompts before and after, which the ADL and independent researchers can do, and labs rarely publish.

Why critics object

The pushback in the posts has two parts. First, the ADL's definitions are contested. The group maintains a Glossary of Extremism that includes entries critics consider political, and Turning Point USA is the example raised. Second, critics argue that a single advocacy organization influencing AI safeguards could cause lawful speech to be suppressed across several large platforms at once.

Those are arguments about governance, and we are not ruling on them here. A few neutral points help readers evaluate them:

• Definitions drive measurement. Whether a model "fails" a test depends on who writes the prompts and the answer key. The ADL's reports state methodology and prompts, which allows others to challenge them.
• One tester is not a standard. The AI industry uses many outside evaluators, from government institutes to academic labs to advocacy groups on various sides. A lab relying on any single group's definitions concentrates judgment in one place. Our post on 406 scientists urging the airgapping of cyber-capable models shows a different model, with calls for an international body and published containment failures.
• Consistency matters. If labs accept input on one kind of harm, critics reasonably ask how other groups' concerns are handled. Publishing the process, not only the outcome, answers that.

What this means if you build with models

If you ship a product on top of a model, you inherit its refusal behavior and its blind spots. A few practical steps follow from this story:

1. Run your own evals on sensitive categories that matter to your users, and keep the prompts and results. Do not rely on one group's index, including this one.
2. Layer your own moderation for user-generated content, with your own policy and appeal path, rather than assuming the base model's refusals match your context.
3. Log model version changes. Refusal behavior shifts between versions. Sora 2's changes, as Greenblatt describes them, are an example of behavior changing without an announcement you would read in a changelog.
4. Expect outside pressure on safeguards from many directions. Advocacy groups, governments and users will each want different things. Having a documented policy is better than reacting one complaint at a time. For agent products, a runtime control like AgentBeam, the agent security platform from the explainx.ai team, can enforce what your agents may and may not do regardless of what the base model decides.

For how value choices show up in model behavior more broadly, see our write-up on Bill Gates arguing for federal AI law over self-regulation and the Qwen censorship audit, which looked at a different kind of guardrail question.

How to check the claim yourself

1. Look for the primary record. An ADL transcript, press release or video from the event would settle exactly what Greenblatt said. As of October 14, we did not find one.
2. Ask the labs. Statements from OpenAI, Anthropic, Google, Meta or X describing an ADL relationship would confirm or narrow it. We found none on this specific announcement.
3. Compare model behavior over time. The ADL and independent researchers publish prompts. Re-running them on current model versions shows whether refusals changed.
4. Separate the three claims. Testing (documented), ongoing contact (documented), and structural changes to models (not documented) are different strengths of claim.

What we do not know

• The exact words and context of the October remark, beyond summaries on X.
• Whether X and xAI are engaged with the ADL on this work. The ADL's January index ranked Grok last, and the Algemeiner notes the chatbot's earlier antisemitic outputs, so a cooperative relationship would be notable and needs a source.
• Whether any lab changed a model because of ADL findings, other than the Sora example in the ADL's own account.
• How the ADL's definitions are applied to political content outside antisemitism, which is the center of the critics' concern.

Bottom line

The documented story is narrower than the viral one. The ADL tests major AI models, publishes the results, and says it is in regular contact with AI companies. Greenblatt cites one concrete behavior change, in Sora. No formal partnership, joint announcement or source code change is documented, and the critics' objections concern who sets definitions of harm for systems used by hundreds of millions of people. We will update this post if the ADL, a lab or a transcript adds detail.

Related reading

• AI models' political compass: left-libertarian study
• What is bias in AI: types, examples, mitigation
• Bill Gates on Meet the Press: federal AI law beats self-regulation
• Qwen censorship audit
• 406 scientists urge airgapping cyber-capable AI models
• OpenAI fires three safety researchers

Details reflect public sources as of October 14, 2026 and may change if the ADL or AI labs publish more.