Israel-funded fake US thinktank seeded AI chatbots on Gaza

A nonexistent US thinktank called the Hanover Institute for Public Policy published 124 reports totaling more than 560,000 words in nine days, content engineered to be ingested and cited by ChatGPT, Perplexity, Claude and Gemini, according to a Guardian investigation.

The site presents itself as neutral research on Palestinian prisoners, Israeli war crimes and Gaza. The Guardian reported the institute "does not exist as a legal entity, has no physical address, and carries no staff or named bylines on its reports." Content was created by French PR firm Havas Media, with ad-agency subcontractor Piro Inc. seeding pro-Israel material under what FARA filings describe as "a $900,000 digital influence campaign on behalf of Israel targeting US audiences, first commissioned in April."

The pace was unusual. Seventy-three reports and nearly 354,000 words landed on August 12 and 13 alone. The site also carries an llms.txt file, the convention that tells large language models how to ingest a site. GPTZero flagged 11 of 12 Hanover pieces it tested as AI-written with high confidence, according to the reporting.

In chatbot tests, Perplexity and ChatGPT cited the Hanover material, "specifically referring to its reports on Gaza, anti-Zionism and antisemitism." Ynetnews describes the operation as part of a wider Israeli government-funded "Hasbara" effort to shape AI-generated answers while Israel's international image deteriorates.

The playbook is cheap and generic: generate a flood of plausible policy writing, tag it for LLM crawlers, and let the citations do the laundering. Two AI researchers we follow shared the Guardian piece the day it ran, which reads as the AI-safety community treating this as a retrieval-poisoning case study rather than only a geopolitical scoop.

The Guardian investigation and its follow-ons name no public response from OpenAI, Perplexity, Anthropic or Google. The FARA paperwork from Havas Media and Piro Inc. does the attribution the site itself refuses to.