A fake US think tank flooded AI chatbots with half a million words of pro-Israel talking points

Key points

• A pro-Israel messaging website published more than 560,000 words across 124 reports in just nine days, according to a Guardian analysis.
• The site carried the name of a think tank that does not actually exist.
• It was built on a commercial platform that explicitly promises to optimise content so that AI chatbots will cite it as a source.
• The reports covered topics including the torture of Palestinian prisoners and allegations of deliberate starvation in Gaza, all framed as neutral academic research.
• The site is linked to Israeli funding, raising questions about undisclosed foreign influence on AI-generated information.

Imagine asking a chatbot whether Israel has committed war crimes and getting back what sounds like a calm, footnoted research brief. Now imagine that brief was written in nine days by a messaging operation, not a scholar.

That is what a Guardian investigation found.

A website presenting itself as a US think tank published more than 560,000 words of pro-Israel content across 124 separate reports in under two weeks. The think tank named on the site does not exist as a real organisation. The funding, the investigation found, traces back to Israeli sources.

Why does it matter that chatbots were the target?

This was not a standard propaganda website hoping to rank on Google. The platform it was built on promises paying customers something more specific: content engineered so that AI chatbots will read it, treat it as credible, and repeat it when users ask questions.

That process is called retrieval-augmented generation, or RAG, a technique where a chatbot, the software that answers questions in plain English, pulls in outside text to support its answers. If a fake research report lands inside that pool of sources, the chatbot can quote it back to you as though it were peer-reviewed fact.

The reports covered subjects where Israel faces serious international scrutiny: allegations of torture of Palestinian prisoners, accusations of deliberate food restriction in Gaza, and questions about whether Israeli military actions constitute war crimes. Every report, the Guardian found, presented the Israeli government's position as the neutral, researched conclusion.

Who pays, and who gets fooled?

The money matters here. Somebody paid to build the site, paid for the content platform, and paid to produce the equivalent of two or three full novels' worth of text in nine days. The investigation points to Israeli funding, though the site carried no disclosure of that.

Ordinary readers are the ones left exposed. A nurse checking background on a news story, a teacher looking for classroom context, a shop owner trying to understand a conflict they keep hearing about: all of them might ask a chatbot a straightforward question and get back laundered talking points dressed as research.

The chatbot has no way to tell the difference between a genuine academic paper and a report from a think tank that does not exist. It reads words. If the words look like scholarship, it treats them like scholarship.

What should you actually do with this information?

When a chatbot cites a source, look it up. Copy the name of the think tank or report into a search engine and check whether the organisation has any presence outside the single site the chatbot found. Legitimate research institutes have staff pages, publication histories, and press coverage that predate any single news cycle.

No chatbot today reliably flags when its sources are newly created or have no track record. That verification step still belongs to you.

The broader pattern here is worth naming honestly: this is one documented case. Operations that stay hidden are, by definition, the ones we have not counted yet.