The AI Story War: PR Firms, Comment Armies and the Documented Battle Over What Google, Wikipedia and Chatbots Say About Palestine

By the Kimi Travels Team · Updated September 2026

---

A generation ago the fight over Palestine's story was fought in newspapers and TV studios. Today it is fought inside the machines that answer your questions: the search bar, the encyclopedia and the AI chatbot. The intuition many readers share — "if 100 websites say one thing and 3 say another, the AI repeats the 100" — is essentially correct, and that is precisely why organised public-relations operations now treat search results, Wikipedia pages and language-model answers as terrain to be captured. This is not a conspiracy theory; it is a documented campaign, with company names, government contracts, fired employees and platform threat-reports. We documented the media side of this story in our feature on hasbara, Hollywood and the story machine; this article goes one layer deeper, into the machines themselves.

How AI Learns: The Majority-Wins Machine

Start with mechanics, because they explain everything that follows. A language model is trained on an enormous slice of the public web — and when an AI answer cites "sources," it typically retrieves the highest-ranked pages it can find, then summarises them. In both cases, majority signal wins: whatever dominates the index, the encyclopedia and the news archive ends up dominating the answer. Nobody programs bias in deliberately; the machine amplifies whatever the ecosystem feeds it. The consequence is simple and enormous: a coordinated operation that floods search results, encyclopedia pages and comment sections is not just doing PR — it is editing the raw material the machines learn from. Whoever controls the flood controls the echo. And the flood is documented.

STOIC: The Case OpenAI Itself Documented

The single most important receipt in this file comes from the AI industry itself. On May 30, 2024, OpenAI published its threat report "Disrupting Deceptive Uses of AI by Covert Influence Operations" — and among the five campaigns it disrupted was one run by STOIC, a commercial Israeli company, which OpenAI described as a "for-hire threat actor." STOIC had used ChatGPT to generate personas, comments and articles promoting pro-Israel and anti-Hamas messaging across the web, targeting audiences in the United States and Canada — the same network was simultaneously disrupted by Meta, as NBC News, TIME and NPR reported that week. Read that again slowly: the company that builds ChatGPT officially documented an Israeli PR firm using ChatGPT to fabricate the appearance of organic public support. This is the majority-wins machine, caught in the act, by the machine's own makers. No rumour about AI manipulation has better documentation than this one — and it did not come from Palestine's side. It came from OpenAI's transparency report.

The Comment Armies: Act.IL and the Flood by Design

STOIC was not an invention; it was a productisation of infrastructure that already existed in public view. Act.IL — a social networking app documented on Wikipedia itself — was built to mobilise thousands of Israel-supporters to coordinate responses to "anti-Israel content" online; investigative reports by the Forward (May 2018) described it as a government-linked app boosting comments on major news outlets' Facebook posts during the Gaza shootings, while Israel's own strategic-affairs establishment openly celebrated the model (Mondoweiss, July 2017). In January 2025, Haaretz documented a pro-Israel AI-powered bot that had gone rogue while mass-commenting on social media. In March 2024, Al Jazeera reported on a network of fake accounts promoting attacks on UNRWA across multiple platforms. A decade of this activity is on the record — apps, bots, volunteer armies and now AI-generated personas, all designed to manufacture the "100 to 3" majority that the machines will later learn from.

The Contracts: Project Nimbus and the 28 Fired Workers

Now the money, stated precisely. In 2021, Google and Amazon won Project Nimbus — a $1.2 billion contract providing cloud and AI services to the Israeli government and military. When Google employees protested in April 2024, the company fired 28 of them — documented by NBC News, NPR and TIME, alongside the workers' own campaign, "No Tech for Apartheid." Here is the honest, documented version of the popular claim that "Israel pays Google": there is no evidence Google takes censorship bribes — but there is a billion-dollar services contract between the companies that answer the world's questions and the state being questioned, plus fired workers who objected, plus the moderation complaints documented by 7amleh and Human Rights Watch. Contracts are not bribes; they are something subtler and more durable — shared commercial interest between the machine's owner and the state it is asked to judge. Precision here is not a nicety. It is the difference between a conspiracy claim that dies in an hour and a contract record that stays on the books.

The Editing Wars: Wikipedia's Bans and the ADL Ruling

Wikipedia deserves its own chapter, because its pages train both humans and machines. In January and February 2025, the encyclopedia's own arbitration committee banned eight editors from Israel–Palestine articles after years of documented edit-warring (Times of Israel, February 1, 2025) — and the Anti-Defamation League's March 2025 report "Editing for Hate" alleged coordinated manipulation of the topic from one direction. But the counter-document matters just as much: in June 2024, Wikipedia's editor community formally declared the ADL a generally unreliable source on the Israel–Palestine conflict — a ruling covered by CNN — on the grounds of its advocacy record. So what does the record actually establish? That every version of every page on this topic is contested ground; that both directions of organised editing have been caught; and that the encyclopedia's internal courts are among the few places where the flooding can be litigated in public. When an AI tells you "Wikipedia says," it is delivering a truce line, not a verdict.

What the Measurements Show: Bias Studies in Both Directions

Is there hard measurement of machine bias on this conflict? Yes — and honesty requires reading all of it. A 2026 peer-reviewed comparative study in the European Journal for Ecological Psychology (Attia) examined Google Gemini and ChatGPT framing of the Israeli–Palestinian question and documented measurable asymmetries. Digital Action's August 2024 investigation documented AI systems dehumanising Palestinians — including image generators stereotyping Palestinians as figures of war and terror. The ADL's March 2025 study documented anti-Jewish and anti-Israel bias patterns in the same models — from an organisation whose own advocacy agenda (and its Wikipedia defeat) should be weighed alongside it. And the human layer beneath it all is quantified: 7amleh, the Arab Center for the Advancement of Social Media, together with Utrecht University, documented 3,520 cases of digital-rights violations against Palestinians across platforms — the "Platformicide of Palestine" analysis reported by Al Jazeera in August 2026 — on top of HRW's 51-page Meta censorship report. The measurements disagree in places; the mechanism is not disputed by anyone: what dominates the data dominates the machine, and the data is being fought over in the open.

Your Counter-Toolkit: Reading an AI Answer About Palestine

Everything above turns the reader from a target into an auditor, using six rules. One: make the AI show its sources — then click them, because a summary without sources is a rumour with formatting. Two: prefer primary documents over any outlet: UN OCHA situation reports, ICJ and ICC rulings, the CPJ journalist-killings database, 7amleh's violation logs, Human Rights Watch and Amnesty reports — documents outlive narratives. Three: triangulate across ownership: when a Qatari-funded channel, an Israeli paper like Haaretz or +972 Magazine, and a corporate wire all print the same figure, believe it. Four: be most suspicious of unanimity — the flood's signature is the suspiciously uniform "everyone agrees," and STOIC and Act.IL prove such agreement can be manufactured. Five: date-stamp everything, because AI answers mix 1948, 2024 and 2026 into one confident voice. Six: remember that documented is not the same as controlled — the same web that gets flooded also caught STOIC, exposed Act.IL, banned the edit-warriors and defeated the ADL in front of the whole encyclopedia. The machine war is real, but it is winnable, one document at a time — as the international-law record shows, and as the survivors of the original betrayal documented in the key that never rusts would be the first to insist.

Frequently Asked Questions

How does AI decide what is "true" about Palestine?
AI systems do not decide truth — they reproduce majority signals. Language models are trained on web-scale text, and AI search answers summarise the highest-ranked pages, so organised flooding of search results, Wikipedia and comment sections directly shapes what the machines repeat. That vulnerability is not theoretical: OpenAI's own May 2024 report documented the Israeli firm STOIC fabricating public-support content with ChatGPT.

What was the STOIC case?
On May 30, 2024, OpenAI published its threat report "Disrupting Deceptive Uses of AI by Covert Influence Operations," identifying STOIC — a commercial Israeli company — as a "for-hire threat actor" that used ChatGPT to generate personas, comments and articles for pro-Israel, anti-Hamas messaging targeting US and Canadian audiences. Meta disrupted the same network simultaneously. It remains the best-documented case of an AI tool being used to manufacture the appearance of public opinion on this conflict.

Does Israel "pay Google" to hide information?
The documented version is more precise than the rumour. There is no evidence of censorship bribes — but Google and Amazon hold the $1.2 billion Project Nimbus contract providing cloud and AI services to the Israeli government and military, and Google fired 28 employees in April 2024 for protesting it. Combined with the moderation violations documented by 7amleh (3,520 cases) and HRW's Meta report, the documented entanglement is substantial — contracts and moderation records, not bribes, and that is exactly why it survives scrutiny.

Can Wikipedia be trusted on Israel–Palestine?
Treat it as contested ground, not verdict ground. Wikipedia banned eight editors from the topic's articles in early 2025 after documented edit-warring, and in June 2024 its editor community declared the ADL a generally unreliable source on the conflict. The encyclopedia's internal courts expose the manipulation in public — but for decisions that matter, read the page's edit history and cited sources, then go to the primary documents themselves.

How do I verify what an AI tells me about this conflict?
Six steps: demand sources and click them; prefer primary documents (UN OCHA, ICJ, ICC, CPJ, 7amleh, HRW, Amnesty); triangulate across differently-owned outlets; distrust suspicious unanimity, which STOIC and Act.IL prove can be manufactured; date-stamp every claim; and remember that documented is not controlled — the system that gets flooded also catches the flooding, in transparency reports, arbitration bans and court rulings. Documents win; rumours die.