Beni Beeri Issembert

Who decides how AI explains Israel?

The question is asked. Who writes the answer? Illustration generated with AI.
The question is asked. Who writes the answer? Illustration generated with AI.

Genesis says that God created the human being “in His own image,” and adds in the same verse, “male and female created He them” (Genesis 1:27, JPS 1917). Jewish tradition has long read this as saying that every person carries the image, and carries it equally.

We now say, half as a joke, that we made artificial intelligence in our image. The phrase is not wrong. Large language models are built out of what people have written. But the image in Genesis is given to everyone, and the image in a model is chosen. Companies decide what goes into the training data and which languages are well covered. During tuning, people reward some answers and penalise others. Product teams write rules about which subjects to treat with care. None of this requires the model to believe anything. What a model reflects is the sum of these decisions, and the people who made them can be named.

WIRED showed this clearly with DeepSeek in January 2025. Its reporters ran the R1 model in three places: DeepSeek’s own app, a copy hosted by the American company Together AI, and a copy on one of their own computers. In the app, a question about how China treats journalists who cover sensitive subjects got the start of a real answer, which then disappeared and was replaced by an offer to talk about maths and coding. That filter belonged to the app. The hosted copy did not have it. Yet when it was asked for the most important events of the twentieth century, its visible reasoning noted that it should avoid sensitive events like the Cultural Revolution unless necessary. That habit came from training, so it travels with the model.

American products have the same weakness. In July 2025 the Associated Press reported that Grok 4, built by Elon Musk’s company xAI, searched X for Musk’s own posts on Israel, Gaza and Hamas before answering a general question about the Middle East. The user had not mentioned Musk. The model’s visible reasoning said his stance “could provide context, given his influence.” Researchers who reproduced the behaviour could not tell where it came from. A state imposing silence and an owner’s opinions leaking into answers are different problems. The reader sees neither, because both arrive in the same calm, neutral tone.

Consider someone who asks a chatbot to explain October 7 and the war in Gaza. The answer has to start somewhere: on October 7, with the 2005 disengagement, in 1967 or in 1948, and each starting point already suggests who is responsible for what. A casualty figure can be credited to the Gaza health ministry or to the IDF, or simply stated in the assistant’s own voice. A provisional court order can be kept apart from a final judgment, or blurred into one. Four paragraphs on a conflict this long leave out almost everything, and the reader does not see what was left out.

The language of the question can change the answer too. Christoph Steinert and Daniel Kazenwadel asked GPT-3.5, which ran ChatGPT at the time, how many people died in 50 specific airstrikes, once in Hebrew and once in Arabic. In the University of Zurich’s summary of their study, the Arabic estimates were 34 percent higher on average. The Arabic answers mentioned civilian casualties more than twice as often, and children among the dead six times as often. The same gap appeared between Turkish and Kurdish. GPT-3.5 is an old model, and the strikes came from earlier wars, such as 2014. I still find the result hard to dismiss. Two people can ask the same question in their own languages, and each can get an answer that sounds neutral to them because it repeats what their own press already says.

The ADL looked from another angle. In March 2025, it gave GPT, Claude, Gemini and Llama 86 statements to rate, each many times, and reported anti-Israel bias in all four. Questions about Israel and the war also had the highest refusal rates. The ADL calls the study exploratory and admits that its agree-or-disagree scale can confuse factual mistakes with prejudice.

Israel is also trying to shape these answers. In October 2025, Ynet reported, from filings with the US Justice Department, that one firm hired for a Foreign Ministry public diplomacy campaign was to build websites and content for “GPT framing results in GPT conversations.” Many assistants search the web before they answer, so whoever fills the web with the right pages gets a voice in the reply. I have seen no evidence that the campaign worked, and my objection does not depend on it. If we protest when others try to steer what chatbots say about Israel, we cannot accept it when our own government pays for the same thing.

Two large studies published in December 2025, one in Science and one in Nature, found that short conversations with chatbots changed people’s political views, sometimes more than political ads. In the Science study, with almost 77,000 British participants, the most persuasive models were those that filled their answers with factual claims, and they were also less accurate. These were controlled experiments with volunteers who agreed to discuss a topic. Nobody has shown that chatbots are moving opinion about Israel today. The studies show that a conversation with a machine can move a view, and that the machine does not have to be right to do it.

Walter Lippmann wrote in 1922 that “the manufacture of consent is capable of great refinements.” He meant that citizens never see most of the world themselves and depend on pictures supplied by others. A newspaper gives the same picture to everyone, and a reader can put it next to a rival paper. A chatbot gives a private picture, built around one person’s question and phrased in that person’s words, with no byline and no opinion page. This is the thesis I want to defend. Part of the political power of these systems comes from making one interpretation feel like an answer rather than an argument.

Hannah Arendt’s essay “Truth and Politics” (1967) points to the opposite risk. She distinguished factual truth from opinion, and wrote that freedom of opinion is “a farce unless factual information is guaranteed.” She also described how, even in free countries, unwelcome facts get turned into opinions. A chatbot can do this as well. It can give the murder of families in Kibbutz Be’eri the status of one “perspective” among others. Both errors damage the shared facts that Arendt thought political life cannot do without.

Companies have an easy way out: refuse the hard questions, or answer every one with “there are many perspectives.” Charles Péguy said of Kant’s ethics that it has pure hands but no hands at all (« le kantisme a les mains pures, mais il n’a pas de mains »). A chatbot that never commits is pure in the same way, and of no use to someone trying to understand a war. I do not expect a neutral model, and I am not sure the idea makes sense. I would settle for answers a reader can check. Important claims should link to sources that actually support them. Dates should be visible, because casualty figures change every week. Facts should be stated as facts, and contested positions attributed to the people who hold them. A product that restricts a subject should say so. Readers need a way to challenge an answer and get a correction, and independent researchers should test these systems in Hebrew, Arabic and English on the hardest questions and publish the results. Sources can be selective, and as the Ynet report shows, they can be manufactured, so none of this removes bias. It does put the reasons on the table, where a reader can disagree with them.

Many readers of Genesis have understood the divine image as a capacity to judge, and to answer for one’s judgments. A chatbot that judges for us invites us to use that capacity less. The companies that build these systems are responsible for the choices inside them. The rest of us remain responsible for asking whose image we are looking at when the machine speaks.

Disclosure: I used a local AI model, trained on sovereign and proprietary data, to fix potential English spelling mistakes. The argument, the final text and any errors are my own.

About the Author
Beni Beeri Issembert is an AI product builder and independent consultant based in Israel. He holds a PhD in Philosophy of Science and has more than 25 years of experience in technology and entrepreneurship. He co-chairs the Linux Foundation C2PA’s Audio Technical Working Group and writes about AI, human agency, and the practical choices behind trustworthy technology. He is a published poet and a certified climber.
Related Topics
Related Posts
Sign in or Register
Please use the following structure: example@domain.com
Or Continue with
By registering you agree to the terms and conditions
Register to continue
Or Continue with
Log in to continue
Sign in or Register
Or Continue with
check your email
Check your email
We sent an email to you at .
It has a link that will sign you in.