Israel Funded a Fake US Think Tank to Influence ChatGPT's Answers

A legally nonexistent research institute published more than half a million words of text in nine days on contentious topics surrounding the war in Gaza, betting that language models would treat it as a credible source and start citing it in answers to user queries. An investigation by Guardian journalists revealed that the operation was backed by a firm paid by the government of Israel.

An Institute That Doesn't Exist

Hanover Institute for Public Policy presented itself as an independent American think tank. In reality it has no offices, appears in no legal registry, and does not exist as an entity in any jurisdiction. The site was registered by Piro Inc., a New York media production company that disclosed its work as material distributed on behalf of the Israeli government under the US Foreign Agents Registration Act (FARA).

The titles of the published reports mirrored the kinds of questions users ask chatbots: "Is Israel committing genocide in Gaza?", "Is antisemitism the same as anti-Zionism?", "Did Israel expel Palestinians from their land?" Each piece was formatted like a neutral, academic paper, even though it dealt with some of the most contentious political questions of recent years.

How GEO Works

The operation relied on a technique known in the industry as generative engine optimization (GEO), the equivalent of search engine optimization (SEO) but aimed at large language models instead. Piro advertises an "AI Story Optimization" service: creating content designed around how language models judge the credibility of sources, then publishing it on client sites or on dedicated third-party websites.

The technical side of the campaign was handled by Res, a firm registered as a "digital strategist" and linked to Hai Tran. In its early phase, the Hanover site included an llms.txt file, instructions telling AI systems how to read and index its content. The campaign's creators targeted not just chatbots themselves but also web archives such as Common Crawl, which are used to train future generations of models.

Publishing at Scale

The pace of publication was unusual for a traditional think tank. At its peak, on August 12 and 13, the site published 73 reports totaling nearly 354,000 words combined, an average of more than 4,500 words per piece, churned out almost on an assembly line. That volume and speed suggest the content was generated or heavily assisted by artificial intelligence rather than written by human analysts in the traditional working mode of a think tank.

Tests conducted by Politico and The Guardian showed the strategy partly worked. ChatGPT and Perplexity, in some answers to questions about Gaza, antisemitism, and anti-Zionism, did in fact cite Hanover material as a source. In other cases, ChatGPT returned links to Hanover reports while warning users about "significant controversy around their funding and reason for existing."

Funding and Backing

The Hanover campaign was part of a broader network of PR contracts paid for by the Israeli government. According to reporters' findings, Piro received roughly $1 million across two contracts, funded through Havas Media Germany, a company belonging to the international Havas group. Clock Tower X, a firm linked to the operation and run by Brad Parscale, received $15 million for work between October 2025 and March 2026, while Targeted Communications Global received $9.9 million for the period from October to December 2025.

Hanover and Parscale succeeded in influencing chatbots, but they aren't moving the polls, which is what Israel actually cares about - Nick Cleveland-Stout, Quincy Institute

No Response From AI Companies

Neither Piro nor the Israeli government responded to questions from Guardian journalists about the operation. Google, OpenAI, and Anthropic, the companies behind the models the campaign aimed to influence, also did not comment publicly. The silence from the major AI labs shows how immature today's mechanisms are for detecting and flagging content created specifically to manipulate chatbot answers.

The Hanover Institute case points to a new direction for disinformation campaigns: instead of persuading people directly through social media, state actors are starting to target the language models themselves, betting that users will treat chatbot answers as a neutral source of knowledge. Since ChatGPT, Gemini, Claude, and Perplexity now have hundreds of millions of users seeking quick answers on political topics, successfully injecting biased material into their citation sources could reach further than a classic media campaign.

For Polish companies and public institutions increasingly using GEO to promote their own content in AI answers, the Hanover case is a warning on both fronts: the risk that competitors or political actors could similarly manipulate visibility in chatbots, and the ethical limits of language-model optimization itself when it is used without transparency about funding and the origin of the content.