LLM Poisoning Case: A Reported Fake Think Tank

What the Hanover Institute fake think tank published

On August 17, 2026, Responsible Statecraft reported that the Hanover Institute for Public Policy is not a real think tank, and that it has put out over 100 reports since it began publishing on August 6. None of the reports carry a byline.

Alice Lee, an analyst at NewsGuard, a company that tracks disinformation, said the site mimics a typical credible American think tank. She pointed to the generic name, the site layout and the red-white-blue color scheme.

The only disclosure is a small disclaimer at the bottom of the page. It says Piro, Inc created the organization on behalf of the Israeli Government Advertising Agency. The outlet adds that Piro has received $900,000 from the Israeli government for its work, subcontracted through Havas Media, a French public relations conglomerate.

Why NewsGuard says AI chatbots reward strong citations

The NewsGuard analyst says large language models (LLMs, the AI systems behind chatbots) favor concrete statistics and data along with strong citations and sources, and that these articles have all of them. That is the analyst's reading of why the format would suit AI chatbots.

Responsible Statecraft describes the format in more detail. The institute's "data reports" come with footnotes and tables of contents, and they argue in a neutral tone. The outlet says this helps them appeal to chatbots like Claude or Gemini.

Many of the reports also follow a formula, according to the outlet. They open with an innocent question that someone might ask a chatbot. The appeal to chatbots is the analyst's and the outlet's reading of the format, not a measurement of any chatbot.

Piro's AI Story Optimization pitch on LLM poisoning and AI chatbots

Piro's own website says the firm writes content engineered for how LLMs evaluate credibility, and it names that service AI Story Optimization. Responsible Statecraft adds that others refer to this practice of influencing artificial intelligence as LLM poisoning. The label comes from those others; Piro's site uses its own product name.

A Piro co-founder described the approach in a LinkedIn post the month before the report. The post says that a question to ChatGPT, Gemini or Perplexity about a product category returns one confident paragraph. It also says the firm spent months reverse-engineering how that paragraph is built.

The outlet also reports on Piro's filing. The agreement Piro submitted to the Department of Justice does not explicitly state that its work for Israel is to influence AI.

GPTZero flagged 11 of 12 Hanover Institute articles

GPTZero flagged 11 of the 12 sampled Hanover Institute articles as AI-written with high confidence, and the twelfth with moderate confidence. Responsible Statecraft chose the 12 articles at random. GPTZero is an AI detection tool that, the outlet says, claims a low false-positive rate.

The sample is 12 of the site's more than 100 reports. So the result covers that sample, not every report on the site.

Drop Site on AI chatbots, citations and training data

Responsible Statecraft writes that Israel also contracted former Trump campaign manager Brad Parscale, under a $46.5 million contract, to create pro-Israel websites engineered to influence chatbots. These are separate sites from the Hanover Institute.

The outlet then cites a Drop Site investigation from the month before. Drop Site found that many chatbots, particularly Microsoft Copilot and Google Gemini, had been successfully trained on data from those websites.

The outlet adds that other chatbots frequently cite the same websites. They do so without flagging them as part of an Israeli influence operation.

How to check the sources AI chatbots cite

You can check any AI answer that names its sources by opening each cited page and looking for an author, a funder and a disclosure. This section is our general advice, not something the reporting found.

In the Hanover Institute case, the reports carry no byline, and the funder appears only in a small disclaimer at the bottom of the page. If you stop at the organization's name, you see neither.

Three questions are enough for a quick check. You can answer each one from the cited page itself:

• Does the page name the person who wrote the piece?
• Does it say who funds the organization?
• Does the footer or an "about" page carry a disclosure?

When the page answers none of them, treat the claim as unverified. For how much traffic AI search sends to sites at all, see AI Referral Traffic Grew 8x: Shopify Q1 2026 Data. For a related problem, where the author shown on a piece is not the one who wrote it, see AI Written Articles Under a Real Writer's Byline.

Three facts on record about this poisoning case

The reporting records three things about this LLM poisoning case. The Hanover Institute published over 100 unsigned reports, Piro's own site describes content engineered for how LLMs evaluate credibility, and GPTZero flagged 11 of 12 sampled articles as AI-written.

Intent is a separate question. The outlet calls the effort a likely attempt to dupe chatbots; that is its inference, and we don't treat it as established.

For readers, the practical step is the one covered above. Open the sources an AI answer names, and check who wrote them and who paid for them.