Contested Neutrality: How Israel Buys the Neutrality of Artificial Intelligence

Within days, Israelis will head to polls, while we in Amman, Ramallah and Gaza follow an outcome we will not vote on, yet whose price we will pay. Whoever occupies the Israeli prime minister’s office will determine the fate of the West Bank, the future of the war on Gaza, and the shape of relations between Israel and Jordan in the years ahead. Even Naftali Bennett, who is presented as the alternative to Netanyahu, has called for the annexation of more than 60% of the West Bank and refuses to rely on Arab parties to form any government.[1] As for Palestinians inside Israel, the Central Elections Committee voted in September to disqualify their two main lists,[2] before the Supreme Court reinstated them in the electoral race on 2 October.[3] We are thus facing an election in which our fates will be decided, while Palestinians inside Israel must contest even their right to take part in it.

Against this backdrop, AI models have come to play a significant role as a source from which voters draw their information. According to a survey conducted by the Israel Internet Association on a representative sample of 1,005 Israelis, a quarter of Israelis are considering using AI tools to reach their electoral decision.[4] When an Arab user asks ChatGPT, Claude or Gemini about these elections, they receive a calm, balanced answer that presents “different points of view” with apparent fairness. It may seem that our problem with these tools is their bias toward Israel, yet this argument alone no longer offers anything new and has become part of a mutual accusation: the US-based Anti-Defamation League (ADL) argues precisely the opposite, asserting in an extensive report that large language models (LLMs) are biased against Israel.[5] Each side, Arab Palestinian and Israeli alike, now sees these tools as biased against it. The more productive question, therefore, is not against whom the machine is biased, but who is working to shape what it says, and how.

Turning to the production of narrative as a front of war, Netanyahu himself, in his address to the United Nations General Assembly in September 2026, described the media and social media as the “eighth front” that Israel is fighting alongside its military fronts.[6] Although he did not explicitly mention AI models, they appear to be the latest extension of this front. In Israel, artificial intelligence is not regarded as a mere consumer tool, which is consistent with Netanyahu’s launch of a national AI program whose declared aim is to establish Israel as a global superpower in the field.[7] Nor has this thinking been confined to laboratories and the military: narrative itself has become a front.

In September 2025, the Israeli government signed a contract with Brad Parscale, Donald Trump’s former campaign manager, which explicitly provides for the creation of websites and content designed to achieve framing outcomes within GPT conversations.[8] The contract was subsequently doubled, reaching US$4.5 million per month. Parscale’s firm set up ten websites dressed in the guise of research platforms and fact-checking outlets, which deny the genocide in Gaza and cast doubt on the killing of the child Hind Rajab. Their average traffic does not exceed a few hundred visitors per month, because their target audience is not human in the first place, but chatbots.[9] In the current election campaign, Israeli parties are likewise investing in influencing the answers these bots provide.[10]

Herein lies the point that the debate over “bias” overlooks: these campaigns succeed not despite the machine’s neutrality, but because of it. Chatbots are designed to present “both sides” of every issue. When a user asks, “Is Israel committing genocide?”, the chatbot looks for two sides to present on an equal footing. This is where the Israeli websites perform their function: written in a measured tone and replete with sources, they are selected by the machine to represent “the other side” in its answer. This is the dynamic described by Hervé Letoqueux, former head of operations at VIGINUM, the French agency responsible for monitoring foreign digital interference.[11]

In this sense, those seeking to exert influence do not need to make the machine biased; it is enough to secure a seat for their narrative as “the other side.” Once the question of genocide is turned into a debate between two equivalent sides, genocide is transformed from a fact into a point of view. Nor does this require much effort: a study by Anthropic found that around 250 poisoned documents are sufficient to teach a language model undesirable behavior, regardless of its size.[12]

Accordingly, the ADL report is not merely a complaint, but part of an organized Israeli effort: bodies that monitor these tools and lobby the companies, others that publish content from which the machine learns, parties seeking visibility in its answers, and companies that test their systems in Israel specifically.

As for us, on the side of the Arab Palestinian narrative, studies are fewer and more scattered. The most notable is a study published this year in Third World Quarterly, which demonstrated that ChatGPT is biased against Palestinians in both Arabic and English, to the extent that some of its answers ignore the right of return and deny the Nakba.[13] Yet such studies remain confined to universities; they are translated neither into pressure on the companies nor into content that reaches the machine, and there is no effort by institutions, research centers or government bodies to influence the models’ answers.

The problem, then, runs deeper than an absence of pressure. Strictly speaking, the issue is not that the models favor one party over another in their answers; the machine builds its answers from what is available, and what is available is unequal. A study of these tools’ answers to Israeli voters found that 84% of them drew on Israeli news websites,[14] whereas Arabic-language websites account for no more than 0.6% of all websites on the internet.[15]

And when a Palestinian inside Israel, whose parties were nearly disqualified, asks about these elections in their own language, they are asking a tool that was not designed for them. Anthropic’s neutrality tests measure the balance between left- and right-wing positions, and the people surveyed to develop its standards came from the United States, the United Kingdom and India,[16] with nothing on Arabic or on the conflict.

What this means for us is the recognition that narrative is no longer decided in newspapers and on screens alone, and that the knowledge infrastructure feeding AI models has become an arena of contestation with its own tools and budgets.

What may be required at the Arab level is threefold. First, Arab research centers should undertake periodic, systematic audits of what these tools say in Arabic about our causes, publish the findings, and put them before the companies. Second, we should demand source transparency from these companies: if Claude, ChatGPT and Gemini flagged that Parscale’s websites are funded by the Israeli government while other tools omitted this, then labeling government-funded content is technically feasible and should become the rule rather than the exception. Third, we should build Arabic content that is rigorous, documented, well organized and accessible, because these very qualities are what make content likely to be picked up in the answers of AI models.

So, the narrative does not end with the announcement of the Knesset election results. And our absence from that narrative—as Arabs and Palestinians—remains, for now, a choice, not a fate.

[1] Political Studies Unit, “The 2026 Knesset Elections: Contours of the Political and Party Map in Israel” [in Arabic], Situation Assessment, Arab Center for Research and Policy Studies, 5/8/2026, accessed on 22/10/2026, at: https://www.dohainstitute.org/ar/PoliticalStudies/Pages/2026-knesset-elections-israeli-political-and-party-map.aspx

[2] “Arab Parties Denounce Ban from Israel’s October Election,” Al Jazeera, 24/9/2026, accessed on 22/10/2026, at: https://www.aljazeera.com/news/2026/9/24/israel-moves-to-bar-both-main-arab-lists-from-october-election

[3] “Israel’s Supreme Court Overturns Election Panel Ban on Arab Parties,” Al Jazeera, 2/10/2026, accessed on 21/10/2026, at: https://www.aljazeera.com/news/2026/10/2/israels-supreme-court-overturns-election-panel-ban-on-arab-parties

[4] “Quarter of Israelis Consider Using AI for Election Votes, Survey Shows,” The Jerusalem Post, 31/8/2026, accessed on 20/10/2026, at: https://www.jpost.com/israel-election-2026/article-907116

[5] “Anti-Jewish and Anti-Israel Bias Found in Leading AI Models, New ADL Report Finds,” ADL, 25/3/2025, accessed on 1/10/2026, at: https://www.adl.org/resources/press-release/anti-jewish-and-anti-israel-bias-found-leading-ai-models-new-adl-report

[6] “Prime Minister Netanyahu’s 2026 UN General Assembly Speech,” Jewish Virtual Library, 24/9/2026, accessed on 30/9/2026, at: https://jewishvirtuallibrary.org/prime-minister-netanyahus-2026-un-general-assembly-speech

[7] “Netanyahu Launches National AI Program: ‘Israel Will Be a Global Power’,” Israel National News, 10/8/2026, accessed on 18/10/2026, at: https://www.israelnationalnews.com/news/431502

[8] “Exhibit AB to Registration Statement, Registration No. 7649,” FARA eFile, U.S. Department of Justice, 18/9/2025, accessed on 22/10/2026, at: https://efile.fara.gov/docs/7649-Exhibit-AB-20250918-1.pdf

[9] Nick Cleveland-Stout, “Israel Is Paying Millions to Train AI Chatbots How to Talk About Gaza. It’s Working,” Drop Site News, 28/7/2026, accessed on 22/10/2026, at: https://www.dropsitenews.com/p/israel-brad-parscale-ai-chatbots-gaza

[10] “Poll-rigging Bots and AI Robocalls: The Dark Tech Running Israel’s Election Campaigns,” Haaretz, 11/8/2026, accessed on 22/10/2026, at: https://www.haaretz.com/israel-news/security-aviation/2026-08-11/ty-article-magazine/israel-elections-2026-the-dark-tech-running-israels-election-campaigns/0000019f-eb5f-d101-a5bf-eb5f63800000

[11] Nick Cleveland-Stout, “How Is ‘Israel’ Poisoning AI Models with Its Narrative of the Gaza War?” [Arabic translation], Noon Post, 30/7/2026, accessed on 22/10/2026, at: https://www.noonpost.com/380607/

[12] “A Small Number of Samples Can Poison LLMs of Any Size,” Anthropic, 9/10/2025, accessed on 20/10/2026, at: https://www.anthropic.com/research/small-samples-poison

[13] G. King, D. Abbas & D. Jegić, “The Political Economy of AI: A Content Analysis of ChatGPT Outputs and Anti-Palestinian Media Bias,” Third World Quarterly (2026), pp. 1–18, accessed on 22/10/2026, at: https://laur.lau.edu.lb:8443/xmlui/handle/10725/17706

[14] “We Ran a Massive AI Election Simulation: Here Is What the Chatbots Recommended to Our AI Personas (Based on 7,051 Citations),” Chatoptic, 28/5/2026, accessed on 22/10/2026, at: https://www.chatoptic.com/blog/israel-2026-elections-ai-visibility-simulation

[15] “From Roots to Relevance,” Strategy& / Abu Dhabi Arabic Language Centre, accessed on 22/10/2026, at: https://www.strategyand.pwc.com/m1/en/industries/culture-and-entertainment/from-roots-to-relevance/from-roots-to-relevance.pdf

[16] “Preparing for the 2026 US Midterms,” Anthropic, 14/10/2026, accessed on 22/10/2026, at: https://www.anthropic.com/news/2026-us-midterms