Loading page…
The site’s foreign-agent filings and machine-readable design make its purpose unusually visible. Its effect on chatbot behavior, or on future model-training data, remains unproven.
Listen to this story
A disclosed influence operation used a synthetic think-tank site to seed Israel-related material into the information channels AI systems may retrieve. Hanover published 124 question-shaped reports in nine days, while Piro Inc. described related work to US authorities under FARA and promoted AI Story Optimization services. ChatGPT surfaced four reports in tests, but there is no evidence yet of changed model behavior.
Hanover published 124 reports totaling more than 560,000 words between August 6 and 14, 2026.
Piro’s FARA filings list two Hanover-related work orders worth $1 million for LaPam, Israel’s government advertising agency.
Pangram found three sampled reports entirely AI-written except for bibliographies; the sample does not establish authorship across all 124.
A website presented as the Hanover Institute for Public Policy published 124 reports totaling more than 560,000 words in nine days this month, using a think-tank format to address questions about Israel, Palestine and antisemitism. The site was filed with the US Justice Department as material distributed for the Israeli government, while its operator sells services intended to shape how large language models evaluate and explain information.
Between August 6 and 14, Hanover posted an average of roughly 4,500 words per report. All but one headline was phrased as a question resembling a chatbot query, including questions about genocide in Gaza, anti-Zionism and Palestinian displacement. The reports used a quasi-academic presentation and cited organizations such as UN agencies, the World Bank and Amnesty International.
The material presents Israel’s positions on allegations including torture of Palestinian prisoners, war crimes and starvation in Gaza as neutral research. Its framing frequently emphasizes that courts have not reached verdicts on accusations made by human-rights organizations, and repeatedly qualifies casualty and aid figures as coming from parties to the events.
The institute has no named staff, bylines or physical address, and the Guardian found no apparent legal entity behind it. A Pangram analysis of three Hanover articles concluded they were entirely AI-written apart from human-written bibliographies; that sample does not establish how every report was produced.
Piro Inc., a New York media-production and advertising firm, filed the site under the Foreign Agents Registration Act. Its filing identifies Piro as providing strategic communications and media-relations services to LaPam, Israel’s government advertising agency, on behalf of the State of Israel. The paperwork says the work may include publicly distributed, source-supported informational materials intended to educate the US public about Israel and related issues.
Hanover published the reports between August 6 and 14, 2026.
Foreign-agent filings list Piro as contracted for $1 million across two work orders.
The listed $1 million covers Piro’s two work orders. The project sits in a wider campaign routed through Havas Media: Justice Department filings record $15 million in payments to Clock Tower X and $9.89 million to Targeted Communications Global between October 2025 and March 2026. Those payments show a broader spending network, not that all of its work was devoted to Hanover.
Hanover initially carried an llms.txt file associated with Res, a platform marketed to help businesses get cited by ChatGPT, Perplexity, Claude and Gemini. It later replaced that file with a bespoke version for machine readers. Such files are designed to make a website easier for large language models to scrape and interpret.
Piro calls its related commercial offering AI Story Optimization. The company says it engineers content around how language models assess credibility, and maps sources including Reddit, YouTube transcripts, reviews, publisher pages and forums to determine which surfaces shape AI-generated answers. That makes Hanover’s publication pattern more than a conventional public-relations site: it is a disclosed attempt to strengthen a political narrative in the online material AI systems may retrieve or learn from.
There is no demonstrated evidence yet that Hanover has shifted model behavior. At the time of the Guardian’s reporting, the site was absent from Common Crawl’s July index, a web repository used as training data by commercial AI developers; its reports were published after that crawl had closed. Seven sites operated by Clock Tower X appeared 294 times in that same index.
The campaign may instead seek near-term retrieval and citation. In tests described by the Guardian, ChatGPT linked to four Hanover reports when asked for the institute’s new research, but also warned users that the organization faced significant controversy over its funding and purpose. Whether future crawls absorb the material, and whether that changes uncited model answers, remains the central unresolved question.
Editorial analysis
This is a useful early example of generative-engine optimization moving from brand marketing into state-backed political communications. The revealing detail is not simply the 124-report publishing burst; it is the combination of FARA disclosure, content engineered around likely user questions, and files intended to help machines read the site. The next meaningful evidence will be whether Hanover material enters a later Common Crawl index or begins appearing in chatbot citations and answers without clear sourcing. Until then, the campaign demonstrates an attempted route into AI information systems, not a measured change in what those systems believe.
Loading discussion...
Be the first to share a perspective or an experience.
Reader comments
Newest comments first. Replies stay oldest first.