Politico dropped a scoop that reads like spy fiction: Israel has launched a covert campaign to influence how large language models portray the country. The operation includes a fake think tank, a French PR giant, and a website designed not for human readers — but for the crawlers that feed ChatGPT and Perplexity.
TL;DR: Israel has set up a fake think tank, the Hanover Institute for Public Policy, in an apparent attempt to dupe AI chatbots into giving pro-Israel answers. In just over a week, the institute published at least 100 articles apparently tailor-made to influence tools like ChatGPT and Perplexity, according to Responsible Statecraft.
What Is the Hanover Institute and Who Created It?
The Hanover Institute for Public Policy is a website that presents itself as a research organization, but researchers and journalists have found no evidence it is a genuine think tank. According to Responsible Statecraft, it appears to be a front created as part of an Israeli influence operation, publishing at least 100 articles in a little over a week — a pace no legitimate policy institute could sustain. The campaign was first exposed by Politico, whose reporting was summarized on X by journalist Jacob Wendler.
The operation reportedly runs through a French PR giant together with the newly created “research institute,” according to Calcalist’s ctech. The content floods the internet with pro-Israel material designed to influence the information AI systems retrieve and cite. That is the core trick. Why court human journalists when you can court the machines that answer questions for millions of people?
Arab News notes that governments have long used advertising, lobbying, television, and social media to shape foreign opinion. The Hanover Institute represents the next step: targeting the AI chatbots that increasingly serve as a first stop for information. The site looks authoritative. It cites itself. And chatbots cannot easily tell the difference.
How Does a Fake Think Tank Manipulate AI Chatbots?
It works by exploiting how LLMs are built. Chatbots like ChatGPT and Perplexity answer questions using text scraped from the web, and they tend to trust sources that look credible — academic formatting, institutional branding, citations. A fake think tank manufactures exactly that appearance. Mathrubhumi’s reporting describes the campaign’s toolkit: specialized AI optimization techniques, targeted Q&A content, and custom llms.txt files deployed through covert websites like the Hanover Institute.
The Q&A format matters most. Frontnieuws explains that the content is not primarily intended for human readers at all — it is written so that AI models will ingest it and reproduce it when users ask about Gaza or the IDF. If a page is structured as a question and an authoritative-sounding answer, a retrieval-based chatbot may quote it verbatim.
Think about the scale. Perplexity and similar tools synthesize answers from whatever ranks or crawls well. A coordinated flood of 100+ polished articles creates a corpus the models treat as consensus. No human needs to read a single word. The audience is the algorithm.
What Is llms.txt and Why Does It Matter in This Campaign?
llms.txt is an emerging web standard — a file placed on a website that tells AI models what content to read and how to interpret it. Think of it as a robots.txt for the LLM era. Site owners use it to guide crawlers toward the pages they want chatbots to ingest. In this campaign, the Hanover Institute used custom llms.txt files to steer AI systems directly to its pro-Israel content, as Mathrubhumi reports.
This is what makes the operation clever. Instead of hoping chatbots stumble onto its articles, the institute explicitly instructed the machines where to look. RT’s coverage of the Politico investigation describes a website whose contents were designed so AI models would use them when answering user questions.
The file format also confers an air of technical legitimacy. Models and their crawlers treat llms.txt guidance as a site’s own curated description of itself. A covert operator can therefore hand the model a pre-packaged narrative, complete with suggested framing. It is search-engine optimization reborn for the chatbot age.
How Many Articles Has the Hanover Institute Published So Far?
At least 100 articles in just over a week, according to Responsible Statecraft — a volume that itself betrays the operation’s purpose. Nick Cleveland-Stout, who analyzed the campaign, noted on X that the Hanover Institute put out over 100 reports in roughly a week. GIGAZINE’s write-up confirms the same figure: a fake think tank established to deceive AI chatbots, publishing at least 100 articles seemingly aimed at that goal.
The pace is the tell. Genuine policy institutes publish deep research on a monthly or quarterly cycle. A site churning out dozens of keyword-targeted pieces per day is optimized for ingestion, not scholarship. Responsible Statecraft describes the articles as “tailor-made to influence chatbots” — formatted, structured, and timed for machine consumption.
The number also reveals intent. One hundred articles in a week creates enough surface area for AI crawlers to encounter the narrative repeatedly across queries. Repetition reads like consensus to a language model. That is the entire strategy.
Which AI Tools Were the Campaign’s Main Targets?
The primary targets were ChatGPT and Perplexity, the AI assistants that millions of people now use instead of traditional search engines. According to reporting summarized by Mathrubhumi, the campaign relied on “specialized AI optimization techniques, targeted Q&A content, and custom llms.txt files via covert websites like the Hanover Institute to manipulate ChatGPT and Perplexity responses.” That detail matters. It shows the operation was not built for human readers at all.
The choice of targets is strategic. Chatbots answer questions directly, without showing users ten blue links to evaluate. If a model’s training or retrieval pipeline ingests slanted content, the bias surfaces as a confident, single-sounding answer. Arab News notes that governments have long used advertising, lobbying, television appearances, and social media to shape foreign opinion — but AI answers represent a newer, less scrutinized channel. Why argue with a human when you can pre-argue with the machine?
Perplexity is especially vulnerable because it retrieves live web content and cites sources. A site publishing large volumes of plausible, question-shaped articles can end up quoted as an authority. ChatGPT, meanwhile, absorbs web text during training. Both paths lead to the same destination: a chatbot that presents a curated narrative as neutral fact.
Who Exposed the Operation and What Did Politico Report?
The story was broken by Politico, with reporter Jacob Wendler announcing the scoop on X: “Israel has launched a new effort to sway public opinion in the US by influencing how LLMs paint it in response to chatbot queries. The foreign influence campaign includes a fake think tank set up to feed content to tools like ChatGPT.” Responsible Statecraft followed with its own analysis, confirming the scale of the operation.
According to Calcalist’s CTech, the campaign was run “through a French PR giant and a newly created research institute,” flooding the internet with pro-Israel content designed to influence what AI systems tell users about Gaza and the IDF. The involvement of an established public relations firm distinguishes this from amateur trolling. It was professionalized media work aimed at machines.
Researcher Nick Cleveland-Stout independently documented the operation on X, noting that the Hanover Institute “has put out over 100 reports in the last week.” The combination of Politico’s reporting, follow-up coverage, and open-source research built a paper trail. RT’s Spanish-language coverage added that the Israeli government allegedly created a website whose contents were designed specifically for AI models to use when answering user questions.
Is This the First Known Attempt to Influence LLM Answers?
No — but it appears to be among the most organized and best-documented state-linked attempts. Arab News frames the campaign within a longer history of influence operations, noting that for decades governments have relied on advertising, lobbying, television, and social media to sway foreign audiences. Targeting LLMs is the newest chapter of an old practice. The tools changed. The goal did not.
Earlier documented efforts included coordinated edits to Wikipedia and SEO manipulation aimed at ranking favorable pages higher in Google results. What makes the Hanover case distinct is the deliberate engineering of content for AI ingestion: question-and-answer formats, dense publication volume, and llms.txt files that instruct crawlers on how to read the site. Mathrubhumi’s reporting describes these as “specialized AI optimization techniques” — a phrase that essentially means SEO for chatbots.
The activist and independent press picked up the story quickly. Activist Post called it “teaching AI to love Israel one bot at a time,” while Frontnieuws observed that the content was “not primarily intended for human readers” at all. That inversion — machines as the audience — is what marks this as something structurally new rather than merely another propaganda push.
Can Chatbot Users Detect AI-Optimized Propaganda Content?
Ordinary users face a hard task, because the content was designed to look credible and authoritative. The Hanover Institute presented itself as a legitimate research organization, complete with reports formatted like policy papers. As Responsible Statecraft put it, the institute “published at least 100 articles that appear tailor-made to influence chatbots” — tailored precisely so that neither users nor the models would flag them as unusual.
There are warning signs, though. A volume of over 100 reports published in just over a week is far beyond what a genuine new think tank produces, as GIGAZINE’s coverage highlighted. Other red flags include a newly registered domain, no verifiable staff or institutional history, and articles structured as direct answers to politically charged questions. Users can also ask chatbots for their sources and check whether those sources actually exist and hold up.
The deeper problem is structural. Chatbots present synthesized answers without the visible link ecosystem that let savvy search users skim for bias. When the cited source itself is a fabrication built for the machine, even careful questioning may not surface the manipulation. Media literacy alone cannot fix a poisoned pipeline.
What Does This Mean for the Future of Search and AI Answers?
It means the battlefield of information warfare has shifted from rankings to retrieval and training data. If a website can be purpose-built to feed answers into ChatGPT and Perplexity, then every actor with a budget — governments, corporations, anyone — has an incentive to build one. Politico’s reporting suggests this is already happening, not hypothetically. The Frontnieuws coverage stressed that the content was produced for AI consumption first, humans second or not at all.
For AI companies, the pressure is now on provenance. Models that cite live sources need ways to assess whether a source is an organic publication or a manufactured one. Otherwise, “AI optimization” becomes the new SEO, except users cannot see the equivalent of a search results page to judge it. Arab News notes that this campaign shows how the traditional tools of state influence are being adapted to a landscape where chatbots increasingly mediate what people learn about world events.
For users, the lesson is skepticism toward confident answers on contested topics. Ask for sources. Verify them. Assume that any well-funded interest may already be writing for the machine you are talking to.
Frequently Asked Questions
What is the Hanover Institute for Public Policy?
It is a fake think tank created in Israel as part of an apparent influence operation targeting AI chatbots, according to Politico and Responsible Statecraft. It presented itself as a research organization but had no established track record. Its actual purpose appears to have been publishing content designed for AI models to ingest and repeat.
How fast did the fake think tank publish its content?
Extremely fast. In just over a week of operation, the Hanover Institute published at least 100 articles, according to Responsible Statecraft and researcher Nick Cleveland-Stout. That pace far exceeds what a genuine new policy institute could produce and is itself a tell of the operation’s mechanical nature.
Which AI chatbots did the campaign try to influence?
The main named targets were ChatGPT and Perplexity, according to Mathrubhumi’s summary of the campaign. The operation used targeted Q&A content and custom llms.txt files to steer their responses. The broader goal was shaping how LLMs portray Israel in response to user queries, as Jacob Wendler reported.
How does llms.txt help manipulate AI models?
The llms.txt file is a standard meant to tell AI crawlers how to read and prioritize a site’s content. In this campaign, custom llms.txt files were used to guide chatbots toward the prepared Q&A material, making it more likely the models would ingest and reproduce it. It essentially works like SEO instructions written specifically for machines.
Summary
The Hanover Institute affair shows how influence operations have adapted to the age of chatbots. A few points stand out:
- Israel, via a French PR firm, set up a fake think tank that published over 100 articles in about a week, built to feed ChatGPT and Perplexity rather than human readers.
- The campaign used AI-specific techniques: question-shaped content and custom llms.txt files designed to steer model ingestion.
- Politico broke the story, with independent researchers like Nick Cleveland-Stout documenting the operation’s scale.
- Users cannot rely on chatbot confidence alone — verifying cited sources is now a basic hygiene habit.
- Expect more of this: any actor with a budget can build content for machines, and AI companies will need real provenance controls.
Read the full reporting at Politico via Responsible Statecraft, Arab News, and CTech — and always ask your chatbot where its answers come from.