Israel Hired a PR Firm to Build a Fake Think Tank That Already Has ChatGPT and Perplexity Citing It
The Hanover Institute for Public Policy has a professional website, a stated mission, and over 100 published reports on Israel and Palestine. It also does not exist.
FARA documents filed with the US Justice Department reveal the organization is a front created by Piro, Inc., a boutique ad agency, on behalf of Havas Media Germany, which holds the Israeli Government Advertising Agency (LaPam) account in the United States. Piro received $900,000 from the Israeli government for its work. The specific Hanover Institute campaign was budgeted at $100,000 to “create and disseminate informational materials intended to educate the U.S. public regarding Israel and related issues.”
The site went live on August 6. It has published more than 100 unsigned reports in 10 days.
Engineered for LLMs, Not Humans
The reports are not designed for human readers. They are structured around the exact question formats that retrieval-augmented systems ingest and cite: “Is the IDF the World’s Most Moral Army?” “Is Israel Carrying Out a Deliberate Campaign of Starvation in Gaza?” “What Caused the Displacement of Palestinians in 1948?” Each article includes data tables, footnotes, citations (often to Israeli government sources), and a neutral academic tone — the signal pattern that causes LLMs to assign credibility to a source.
Piro’s own website calls this “AI Story Optimization”: content “engineered for how LLMs evaluate credibility.” The wider industry term is LLM poisoning. The mechanism is simpler than it sounds: flood retrieval corpora with plausible-looking, citation-heavy content before the model re-trains or the RAG index updates, and the model will surface it as an authoritative source.
It is already working. Politico ran neutral test queries through ChatGPT and Perplexity. Both cited Hanover Institute material in their responses. The site’s content on Gaza, anti-Zionism, and antisemitism is appearing in AI-generated answers without any disclosure of the government origin.
The Infrastructure
The site was built by Res, a Seattle startup that markets itself explicitly on helping clients influence what AI systems say about them. After Politico contacted Res for comment, all mentions of the firm disappeared from the Hanover Institute website.
Responsible Statecraft analyzed 12 reports using GPTZero. GPTZero flagged 11 with “high confidence” AI-written; one as “moderate confidence.” None carry bylines. The site’s own disclaimer — buried at the bottom — states: “This material is distributed by Piro, Inc. on behalf of Havas Media Germany GmbH on behalf of the Israel Government Advertising Agency (LaPam).”
Not Isolated
This is not the first Israeli AI influence operation. Brad Parscale’s firm runs a separate $46.5 million contract to produce pro-Israel websites also designed for LLM ingestion. A Drop Site investigation found Microsoft Copilot and Google Gemini had already been trained on content from those sites without flagging its origin as a state-sponsored influence operation.
The Hanover Institute operation is distinctive because its intent is explicit in both the FARA filing and Piro’s public marketing. The company is not pretending this is journalism. It is selling a service that reverse-engineers how AI systems assign credibility, then manufactures sources optimized for that credibility formula.
The Structural Problem
Chatbots and RAG-backed systems can be systematically manipulated before model weights ever change. The attack surface is the retrieval corpus: the broader internet that LLMs draw from when generating responses. State actors with media budgets can exploit this at industrial scale. The Hanover Institute campaign illustrates that countermeasures — provenance metadata, source credibility signals, FARA-disclosure integration into model training — do not yet exist at the speed the threat is moving.
What is novel here is not the propaganda. It is the explicit targeting of AI systems as the primary distribution mechanism, the use of FARA-registered foreign agent infrastructure to do it, and the confirmation that it is already working in production.