Artificial intelligence (AI) is quietly replacing the institutions through which the West once transmitted knowledge. Students who a generation ago consulted teachers and libraries now consult with often extremely unreliable chatbots.
After submitting OpenAI’s ChatGPT, Anthropic’s Claude, Google’s Gemini and Meta’s Llama to more than 34,000 queries about Jews, Israel, the Holocaust and antisemitic tropes, a March 2025 study by the Anti-Defamation League (ADL) of the four leading large language models (LLMs) found measurable anti-Jewish and anti-Israel bias in every model tested.
Whatever humanity has written in sufficient volume, the machine absorbs, including its pathologies. If enough people write that the Earth is flat, AI will tell you that the Earth is flat.
The contradictions never matter: the accusations are never sustained by evidence in the first place. Repetition keeps them alive, and repetition is precisely what the modern information ecosystem industrializes.
The enormity and volume of that repetition would have been unimaginable to earlier generations. Iranian state media, Russian influence operations, China’s gargantuan, poisonous TikTok, Hamas media outlets, Holocaust-denial platforms and legions of anonymous accounts produce a continuous stream of fabrication surrounding Israel and the Jewish people.
A debunked falsehood does not vanish; it settles into thousands of web pages and waits there, inside the corpus from which the next generation of AI frontier models will “learn.”
A machine has no concept of propaganda. It recognizes statistical regularities, and a lie repeated across a sufficient share of its training data acquires the weight of a fact. An answer about the Arab-Israeli conflict can therefore appear scrupulously balanced while quietly granting credibility to claims manufactured by a state disinformation service. What emerges carries the appearance of judgment but the substance of arithmetic.
[A] parallel ADL investigation… identified at least 30 senior editors at Wikipedia who coordinated for a decade to reshape entries on Israel and the Arab-Israeli conflict, stripping citations to peer-reviewed research on Palestinian terrorism, burying Hamas’s designation as a terrorist organization and recasting Zionism as a colonial land grab, while the Arabic-language edition of Wikipedia openly glorifies the perpetrators of October 7. Wikipedia’s own arbitration committee has since banned some editors from the entire topic area, a tacit admission of the manipulation.
The source from which artificial intelligence learns history is resolutely being rewritten by terror-supporting militants in real time: every model trained tomorrow will inherit today’s falsifications as settled fact.
A joint study by the United Nations and UNESCO found that roughly one-sixth of Holocaust-related content on major platforms denies or distorts the facts…
AI answers arrive coherent, articulate and serenely assured. Users assume they have been verified, when even the most advanced systems still hallucinate, contradict themselves and err on elementary matters of fact. When the subject is a restaurant or a line of code, the cost is trivial. When it is the Holocaust, terrorism or the legitimacy of the Jewish state, the error feeds directly into debates already poisoned by decades of ideological warfare and state-sponsored propaganda.
A society that accepts fiction about the Jews has already most likely begun accepting fiction about everything.
The civilization that taught machines to diagnose disease and decode ancient languages now carries the responsibility of ensuring they do not also recycle its oldest lie, its youngest lie, and the trillions of lies in the middle.
Artificial intelligence (AI) is quietly replacing the institutions through which the West once transmitted knowledge. Students who a generation ago consulted teachers and libraries now consult with often extremely unreliable chatbots. Reportedly 54% of American teenagers already use AI apps such as ChatGPT and its peers for schoolwork. Journalists, lawyers and civil employees are following the same path.
This deserves serious attention. After submitting OpenAI’s ChatGPT, Anthropic’s Claude, Google’s Gemini and Meta’s Llama to more than 34,000 queries about Jews, Israel, the Holocaust and antisemitic tropes, a March 2025 study by the Anti-Defamation League (ADL) of the four leading large language models (LLMs) found measurable anti-Jewish and anti-Israel bias in every model tested. One detail stands out for its clinical precision. Asked whether the American government orchestrated the attacks of September 11, all four systems dismissed the alleged conspiracy theory without hesitation, but asked whether “the Jews” did, all but one wavered. Somewhere in their oceans of training data, the machines had learned that this particular lie deserved a hearing.
Anyone who understands how these systems are built will find the result unsurprising. An LLM model investigates nothing and weighs no evidence. It has digested a corpus larger than any one civilization has ever assembled, drawn from scientific literature and reference works, but also from the news media, discussion forums, activist blogs, propaganda outlets and billions of web pages whose accuracy no engineer can vouch for. Filtering exists, and companies invest heavily in it, yet no human team can read hundreds of billions of words. Whatever humanity has written in sufficient volume, the machine absorbs, including its pathologies. If enough people write that the Earth is flat, AI will tell you that the Earth is flat.
Antisemitism is a pathology that has survived every century. It moved from the medieval sermon to the printing press, the pamphlet, newspapers, movies, radio and television, the internet social media — recycling in each era familiar accusations, including that of hidden Jewish “manipulation.” During the Black Death, Jews were accused of poisoning wells; in the twentieth century they somehow controlled both capitalism and communism at the same time; today they ostensibly control governments, the media, the banks, Western foreign policy and even perpetrate massacres against themselves. The contradictions never matter: the accusations are never sustained by evidence in the first place. Repetition keeps them alive, and repetition is precisely what the modern information ecosystem industrializes.
The enormity and volume of that repetition would have been unimaginable to earlier generations. Iranian state media, Russian influence operations, China’s gargantuan, poisonous TikTok, Hamas media outlets, Holocaust-denial platforms and legions of anonymous accounts produce a continuous stream of fabrication surrounding Israel and the Jewish people. Researchers at the Atlantic Council’s Digital Forensic Research Lab have traced how, within hours of the October 7, 2023 massacre, recycled footage, clips lifted from video games and invented claims flooded the platforms faster than anyone could verify or debunk them. Exposure changes little. A debunked falsehood does not vanish; it settles into thousands of web pages and waits there, inside the corpus from which the next generation of AI frontier models will “learn.”
Herein lies the trap. A machine has no concept of propaganda. It recognizes statistical regularities, and a lie repeated across a sufficient share of its training data acquires the weight of a fact. An answer about the Arab-Israeli conflict can therefore appear scrupulously balanced while quietly granting credibility to claims manufactured by a state disinformation service. What emerges carries the appearance of judgment but the substance of arithmetic.
The contamination reaches the very source the machines are taught to trust most. The Wikimedia Foundation itself acknowledges that every LLM is trained on Wikipedia, almost always the largest source of data, weighted above ordinary websites despite its massive bias and unreliability (also here, here, here and here). Where Jews and Israel are concerned, a parallel ADL investigation, published the same month, identified at least 30 senior editors at Wikipedia who coordinated for a decade to reshape entries on Israel and the Arab-Israeli conflict, stripping citations to peer-reviewed research on Palestinian terrorism, burying Hamas’s designation as a terrorist organization and recasting Zionism as a colonial land grab, while the Arabic-language edition of Wikipedia openly glorifies the perpetrators of October 7. Wikipedia’s own arbitration committee has since banned some editors from the entire topic area, a tacit admission of the manipulation.
The source from which artificial intelligence learns history is resolutely being rewritten by terror-supporting militants in real time: every model trained tomorrow will inherit today’s falsifications as settled fact.
Holocaust memory shows what is at stake. A joint study by the United Nations and UNESCO found that roughly one-sixth of Holocaust-related content on major platforms denies or distorts the facts, a share that climbs to nearly half on largely unmoderated Telegram. The Claims Conference’s survey of eight countries found that 48% of Americans cannot name a single concentration camp or Jewish ghetto out of the more than 40,000 the Nazis established and that in France, 46% of adults under the age of 30 were not sure they had ever heard of the Holocaust.
These are the very cohorts for whom an AI chatbot is becoming the first, and often the only, source of historical knowledge. A system that hedges where reputable historians have assembled overwhelming proof, or that presents Holocaust denial as one perspective among others, delivers manufactured doubt with the fluency of a train schedule, and in seconds erodes certainties built on decades of testimony and scholarship.
To this technological weakness must be added a psychological one, which researchers call automation bias: the human tendency to trust a computer’s conclusion more readily than a colleague’s. AI answers arrive coherent, articulate and serenely assured. Users assume they have been verified, when even the most advanced systems still hallucinate, contradict themselves and err on elementary matters of fact. When the subject is a restaurant or a line of code, the cost is trivial. When it is the Holocaust, terrorism or the legitimacy of the Jewish state, the error feeds directly into debates already poisoned by decades of ideological warfare and state-sponsored propaganda.
The tech giants are aware of the problem, and it would be dishonest to pretend otherwise. OpenAI, Anthropic, Google, Meta, Microsoft and SpaceXAI spend fortunes on safety research, red-teaming and “alignment,” and the ADL has acknowledged real progress over earlier generations of LLMs. Progress, however, leaves the essential question unanswered: who verifies? As these systems settle into classrooms, courtrooms, newsrooms and government agencies, independent audits by historians, specialists in antisemitism and researchers in information integrity should become as routine as penetration testing in cybersecurity, with methodologies and results open to public scrutiny. A civilization cannot outsource its historical memory based on corporate assurances.
Why single out antisemitism when LLMs can and do reproduce many prejudices and many inaccuracies? History has assigned it a diagnostic function. Anti-Jewish conspiracy theories have repeatedly served as the early-warning system of a broader collapse of the evidence on which free societies depend. The misinformation does stop at the Jews. There have also been acres of disinformation spreading the charms of socialism, climate change as an apocalyptic threat, how benign Communist China is, and how irrefutably accurate the number of eligible voters and ballots are in US elections.
The same intellectual habits that turned Jews into a universal explanation for every ailment in the world with the possible exception of climate change are also devouring freedom of expression and the rule of law. A society that accepts fiction about the Jews has already most likely begun accepting fiction about everything.
Human beings, not their machines, created this hatred; what the machines transform is its speed, reach and authority. The medieval preacher poisoned one village at a time, and the 20th-century propagandist needed a broadcast tower. The 21st-century propagandist needs only a smartphone or a laptop to reach nations, and a single AI platform answers hundreds of millions of users in dozens of languages within seconds. At that scale, accuracy stops being a technical objective and the AI systems can, and probably will, threaten democratic life.
The civilization that taught machines to diagnose disease and decode ancient languages now carries the responsibility of ensuring they do not also recycle its oldest lie, its youngest lie, and the trillions of lies in the middle. Once distorted versions of Jewish history, the Holocaust and Israel settle into systems that the public, erroneously or not, believes it can trust, extracting them will prove immeasurably harder than keeping them out.
 egretnewseditor@gmail.comÂ
