A Fake Think Tank Wrote 100 Reports in 11 Days. Its Audience Is ChatGPT.

TL;DR

On August 17, a Politico investigation and Responsible Statecraft exposed the Hanover Institute for Public Policy, a think tank that exists only as a website. It appeared on August 6, published more than 100 anonymous "data reports" in 11 days, and carries a disclosure tracing its funding through two intermediaries to the Israeli government. Per Responsible Statecraft's analysis, GPTZero flags 11 of 12 sampled reports as high-confidence AI-written. The intended reader is not you: in Politico's neutral-query tests, ChatGPT and Perplexity were already citing Hanover material. LLM grooming now has a state budget, a FARA paper trail, and working results.

A think tank with no thinkers

hanoverinstitute.com looks like every mid-tier Washington policy shop: sober layout, red-white-and-blue palette, a tidy catalog of research. What it does not have is people. The reports carry no author names, and the headlines are phrased as questions: "Is There a Policy of Starvation in Gaza?", "Is the IDF the World's Most Moral Army?", "Who Is the Aggressor in the Israel-Palestine Conflict?"

Question-shaped headlines are not a house style, they are a targeting decision. They mirror how people type queries into chatbots, which makes the pages unusually likely to get pulled in when a model goes looking for grounding. NewsGuard analyst Alice Lee told Responsible Statecraft the sites appear built to reach Americans arriving through search engines or AI chatbots. The volume backs that up: more than 100 reports since August 6 is a pace no human research staff hits.

The confession is in the llms.txt

Here is the detail that makes this a builder story. The site ships an llms.txt, the emerging standard for serving your content to language models as a clean, ingestible digest. We pulled the file, and the legally required disclosure sits right there in the machine-facing copy: the material is "distributed by Piro, Inc. on behalf of Havas Media Germany GmbH on behalf of the Israel Government Advertising Agency (LaPam)." The FARA confession is addressed directly to the models being groomed.

The paper trail matches. FARA filings with the US Justice Department tie more than a dozen Hanover articles to a $100,000 campaign to "educate the American public about Israel and related issues," per Politico's reporting as relayed by Arab News. Responsible Statecraft puts Piro's total take from the Israeli government at $900,000. Piro co-founder Daniel Rosenberg told Politico the firm was retained "to place accurate and sourced information in the public record." His own LinkedIn had already posed the sharper version: "how do you make sure AI knows how to tell them?"

One touch worth savoring: the research catalog in that llms.txt lists three domains, and one of them is AI bias measurement. The fake think tank publishes reports about AI bias, addressed to AIs, while biasing them.

How LLM grooming actually works

Classic SEO gamed a ranking function. LLM grooming games a citation. The near-term channel is not pretraining, it is retrieval: when ChatGPT, Perplexity, or your own RAG stack answers with web grounding, it fetches sources, weighs credibility signals, and quotes. Fill the index with pages whose headlines match the question distribution and whose bodies are dense with statistics, footnotes, and a studiedly neutral tone, and the citation layer does the rest. Think of it less as buying billboards and more as quietly donating pre-written books to every library in town, shelved under exactly the questions you expect patrons to ask.

The engineering behind it is professional. Politico traced technical markers on hanoverinstitute.com to Res, a Seattle startup whose business is getting clients recommended by AI models, and an identical copy of the Hanover homepage hosted on Res infrastructure went dark after Politico contacted the company. Generative engine optimization is a real industry now. This is what it looks like with a government client.

The $46.5 million umbrella

Hanover is the newest and cheapest piece of a documented, much larger effort. In July, Time revealed that Clock Tower X, the firm of former Trump campaign manager Brad Parscale, holds a $1.5 million-a-month Israeli contract aimed in part at shaping how ChatGPT, Claude, and Gemini characterize Israel and the war. Drop Site News then published the FARA agreement behind the $46.5 million campaign, and the contract language does not hedge: the deliverable includes "websites and content to deliver GPT framing results on GPT conversations." Network sites like paxpoint.org and factsignal.org were already live, and Parscale's team told Axios in April it was "seeing success" getting AI systems to pick up its material.

Eleven days from launch to citation

Politico ran neutral prompts about Gaza and antisemitism through ChatGPT and Perplexity and watched both cite Hanover Institute material, repeatedly. That is the whole return on investment: a $100,000 content campaign, eleven days old, surfacing inside the answer boxes of mainstream chat products. Nobody hacked anything. The models did exactly what grounded search is designed to do. The well they drew from had been salted.

What this means if you build with AI

First, if your product does grounded answers, RAG over web content, or agentic research, your retrieval layer is now an adversarial surface with state-level budgets behind it. The credibility heuristics most stacks lean on (institutional design, citation density, neutral register, a well-formed llms.txt) are precisely the features this operation manufactured on purpose.

Second, the detection story should worry you more than the operation. This one unraveled because FARA compelled a disclosure line and because the content flunked an AI-text detector. Both are trailing indicators, and the next operator, foreign or domestic, will not file paperwork. There is no crawl-time signal that separates a fake think tank from a thinly staffed real one.

Third, this is a genre now, not an incident. Time already sells ads that only agents can see; grooming is the unpaid version of the same insight, that the highest-leverage reader on the internet is no longer human. Source allowlists, provenance weighting, and adversarial red-teaming of your retrieval pipeline stop being nice-to-haves the moment governments are on the other side of the query.

Key Takeaways

• Politico and Responsible Statecraft exposed the Hanover Institute for Public Policy, a fake think tank that published 100+ anonymous "data reports" in the 11 days after its August 6 launch.
• The site's own llms.txt carries the FARA disclosure: distributed by Piro, Inc. on behalf of Havas Media Germany GmbH on behalf of the Israel Government Advertising Agency (LaPam).
• FARA filings tie more than a dozen Hanover articles to a $100,000 campaign; Responsible Statecraft's GPTZero sampling flagged 11 of 12 reports as high-confidence AI-written.
• It works: Politico's neutral tests found ChatGPT and Perplexity citing Hanover material on Gaza and antisemitism queries within days of launch.
• It sits under a documented $46.5 million campaign involving Brad Parscale's Clock Tower X, whose FARA contract promises "GPT framing results on GPT conversations."
• For builders: retrieval is now an adversarial surface, and the credibility signals RAG stacks trust (citations, neutral tone, institutional polish) are exactly what gets manufactured.

Sources: Responsible Statecraft, Politico (reporter's thread), Arab News, Drop Site News via Quincy Institute, Time, Hanover Institute llms.txt (primary artifact)