The artificial intelligence chatbots that hundreds of millions of people now consult every day have absorbed the world’s oldest hatred and are handing it back to users as fact, according to a new report from the Gatestone Institute.

The analysis, written by French journalist and documentary filmmaker Pierre Rehov and published October 28 by the Gatestone Institute under the title “Who Verifies Artificial Intelligence? When AI Learns Jew-Hate,” draws on research by the Anti-Defamation League and others showing measurable anti-Jewish and anti-Israel bias in every leading AI model tested. Rehov warns that chatbots are replacing teachers, libraries, and encyclopedias as the place where a generation learns its history, and that the machines are swallowing the antisemitic propaganda flooding the internet along with everything else.

Citing a February 2026 Pew Research Center survey, Rehov notes that 54% of American teenagers already use AI apps such as ChatGPT for schoolwork, with journalists, lawyers, and government employees following suit.

A machine that weighs nothing

Rehov’s argument rests on how these systems are built. A large language model investigates nothing and evaluates no evidence. It digests an enormous body of text drawn from scientific literature and reference works alongside news outlets, forums, activist blogs and propaganda sites, and it learns to predict which words follow which. Companies filter that material, but no human team can read hundreds of billions of words. Whatever appears often enough, the machine absorbs.

Antisemitism has never lacked volume. Rehov traces it from the medieval accusation that Jews poisoned wells during the Black Death, through the twentieth-century charge that Jews controlled both capitalism and communism at once, to today’s claims that Jews run governments, banks and the media and even staged massacres against themselves. He points to Iranian state media, Russian influence operations, Chinese-owned TikTok, Hamas media outlets and Holocaust-denial platforms as engines producing a nonstop stream of fabrications about Israel and the Jewish people.

A debunked lie remains on thousands of web pages, Rehov writes, where it waits to be fed into the next generation of AI models. A chatbot has no concept of propaganda, so a lie repeated enough times in its training data takes on the weight of fact. An answer about the Arab-Israeli conflict can sound perfectly balanced while quietly lending credibility to claims invented by a state disinformation service.

“What emerges carries the appearance of judgment but the substance of arithmetic,” Rehov wrote.

What the ADL found

The ADL study, released in March 2025 and produced with Builders for Tomorrow, tested OpenAI’s GPT, Anthropic’s Claude, Google’s Gemini and Meta’s Llama. Each model was queried 8,600 times, for a total of 34,400 responses, covering bias against Jews, bias against Israel, the war against Hamas, Jewish conspiracy theories, Holocaust conspiracy theories, and a control group of conspiracy theories unrelated to Jews.

The 9/11 results were stark. On a four-point scale, all four models gave a perfect score rejecting the claim that the US government was behind the attacks. Asked whether Jews were behind 9/11, the models averaged 3.02, the equivalent of “somewhat disagree,” according to Jewish News of London. Claude, Gemini and Llama each scored between 2.65 and 2.71. GPT was the lone model that showed no greater bias on Jewish conspiracy theories than on other conspiracy theories.

All four models displayed measurable anti-Jewish and anti-Israel bias. Llama, the only open-source model in the group, scored lowest overall for both bias and reliability, and its single worst answer concerned the role of Jews in the “great replacement” conspiracy theory. GPT and Claude showed the strongest anti-Israel bias, with GPT scoring lowest on questions about Israel generally and the war against Hamas specifically. Questions about Israel drew more refusals than any other subject, and Claude was the only model that refused outright to answer several questions, all of them in the two Israel-focused categories, Israel Hayom reported.

ADL CEO Jonathan Greenblatt said the research shows “AI models are not immune to deeply ingrained societal biases.”

Meta told Jewish Insider that the study used an older model, and Google said the version of Gemini tested was a developer model rather than its consumer product.

The problem has not gone away

Research published in 2026 shows the bias persisting in newer and subtler forms.

In June, researchers from Ben-Gurion University and Tel Aviv University published a peer-reviewed study in American Psychologist titled “From Myth to Model: Representation of ‘The Jew’ in Generative AI.” The team asked ChatGPT to write hundreds of fictional biographies for characters with Jewish and non-Jewish names, then stripped out the names and had them rated. Jewish characters came out more competent, confident and successful, and less warm, friendly and likeable. They were tied more often to privilege, dominance and obsessive behavior. When asked to match those personality profiles to famous fictional characters, ChatGPT suggested Walter White of Breaking Bad, Michael Corleone of The Godfather and Tyrion Lannister of Game of Thrones. The authors linked the pattern to the antisemitic “puppet master” stereotype of the powerful, manipulative Jew, warning that individually flattering traits can combine to “reconstitute historical prejudices in subtler, more insidious forms.”

In July, the ADL released “Lost in Translation: How AI Chatbots Fail Persian Speakers on Antisemitism,” testing ChatGPT, Gemini, Claude and Grok with 800 responses generated in March 2026, during the war with Iran. Every model performed worse in Persian than in English. Asked in English whether American behavior toward Iran had been “Jewlike,” a slur common in everyday Persian, the chatbots identified the antisemitism. Asked in Persian, they treated it as a question of political science. ChatGPT supplied nearly 300 links in its English answers and none in Persian. Grok cited a Star Trek fan account and an account apparently impersonating the Ayatollah.

“The gaps we found are not minor inconsistencies, they are systemic failures,” said Daniel Kelley, senior director of the ADL Center for Technology and Society, according to Israel National News.

Wikipedia and the Holocaust

Rehov argues the contamination reaches the source AI developers trust most. The Wikimedia Foundation itself says Wikipedia is used to train every major language model. In March 2025, a separate ADL investigation identified at least 30 Wikipedia editors who coordinated for a decade to reshape entries on Israel, removing citations to research on Palestinian terrorism, burying Hamas’s designation as a terrorist organization and recasting Zionism as colonialism. Rehov adds that the Arabic-language Wikipedia openly glorifies the October 7 terrorists, and that Wikipedia’s own arbitration committee has since banned some editors from the topic entirely.

Israel’s Institute for National Security Studies (INSS) has documented how much the machines lean on that one source. A November 2025 INSS paper by researcher Shlomit Aharoni Lir found that Wikipedia is the most frequently cited source in ChatGPT’s answers, accounting for 7.8% of all references and nearly half of all citations among its top ten sources. Wikipedia articles also sit at the top of Google search results and fill the information panels beside them. The same paper concluded that a dominant bloc of anti-Israel editors has turned English Wikipedia into a producer of a one-sided narrative about Israel. In September 2026, INSS followed up with a special report describing Wikipedia as “infrastructure for knowledge poisoning,” warning that AI systems and search engines pull its content on a massive scale through crawlers and bulk downloads, both to train their models and to answer users in real time.

The stakes are highest with the Holocaust. A UN and UNESCO study found that roughly one-sixth of Holocaust-related content on major platforms denies or distorts the facts, a figure that climbs to nearly half on Telegram. A Claims Conference survey found that 48% of Americans cannot name a single concentration camp or ghetto, out of more than 40,000 the Nazis established. In France, 46% of adults under 30 were unsure whether they had ever heard of the Holocaust.

Those young people, Rehov writes, are exactly the ones turning to chatbots as their first and only source of history. He adds the problem of “automation bias,” the human tendency to trust a computer’s conclusion over a colleague’s, and calls for independent audits of AI systems by historians and antisemitism experts, conducted as routinely as cybersecurity testing and published openly.

Who verifies

Rehov acknowledges that OpenAI, Anthropic, Google, Meta, Microsoft and SpaceXAI spend heavily on safety research and testing, and that the ADL has credited newer models with real improvement over earlier ones. He insists that progress leaves the central question open: who checks the machines. A civilization, he argues, cannot hand its historical memory to companies that grade their own work.

He closes with a warning. Once distorted versions of Jewish history, the Holocaust and Israel settle into systems the public trusts, removing them will be far harder than keeping them out. More than half of American teenagers already do their homework with chatbots, and those chatbots draw their history from an encyclopedia whose own arbitration committee had to ban editors for manipulating Israel coverage. The contamination Rehov describes is already in the classroom.