이스라엘이 만든 가짜 미국 싱크탱크, AI 챗봇 여론 조작 시도

Fake US thinktank set up and funded by Israel sought to game AI for propaganda In effort to prime chatbots to make pro-Israel arguments the site published 124 reports, over 560,000 words in nine days, Guardian analysis shows A pro-Israel messaging website badged with the name of a thinktank that does not exist has published more than half a million words in nine days, built on a commercial platform that promises to optimize content so that AI chatbots will cite it. The site gives Israel’s position on subjects including the torture of Palestinian prisoners , Israeli war crimes and whether Israel has deliberately starved Palestinians in Gaza, all presented as neutral research. The New York media production company Piro Inc filed the site with the US justice department this month under the Foreign Agents Registration Act (Fara), a 1938 law that requires anyone working for a foreign government to formally disclose it, as material it distributes for the Israeli government. The site publishes in the name of the Hanover Institute for Public Policy , which apparently does not exist as a legal entity in any jurisdiction, has no physical address, and carries no named staff and no bylines on any of its reports. Its own terms of use are governed by the laws of “the state in which the Institute is established”, which they decline to name. The existence of the site was first reported on 14 August. Since then the institute has added a page headed funding, which states: “An earlier version of this page said the Institute operated without external funding, and that was true when it was written.” The site is one part of a broader effort, financed by tens of millions of dollars from the Israeli government and routed through third parties including the European advertising group Havas Media, to prime chatbots to make Israel’s arguments for it. A contract between Havas and another American firm working on the campaign commits it to the “deployment of websites and content to deliver GPT framing results on GPT conversations”. The Guardian contacted Piro Inc, Havas Media and the Israeli government advertising agency LaPam for comment. Comment was also sought via a contact email address published on the Hanover Institute website. It comes at a point when Israel’s standing in US public opinion has collapsed , and when Israeli officials have begun briefing media outlets that money spent on so-called “Hasbara” efforts at public diplomacy overseas has been wasted. The site According to a Guardian analysis, between 6 and 14 August the Hanover Institute published 124 reports totaling more than 560,000 words, an average of about 4,500 words each. On 12 and 13 August alone it produced 73 reports, amounting to almost 354,000 words in two days. Nothing has been published since the first reports about the Hanover Institute. All but one of the 124 headlines on the website opens with a question resembling one that a user might type into a chatbot, including “ Is Anti-Zionism Antisemitism? ”, “ Did Israel Expel Palestinians From Their Land? ”, “ Is Israel Carrying Out a Genocide in Gaza? ” The reports are presented in a quasi-academic manner and are punctiliously sourced, with citations including the World Bank , UN agencies , the Palestinian Central Bureau of Statistics , Amnesty International and the text of the Genocide Convention . Nick Cleveland-Stout, a research associate at the Quincy Institute who has reported on the site, told the Guardian that his first impression was of something impressive. “It succeeds in masquerading as very academic – the website design, the logos, everything just looks good. But the more you read it, you’re able to realise that a lot of it makes no sense.” For example, the information the Hanover Institute site presents is persistently framed to blunt criticism of Israel. Where human rights organisations have concluded that Israel is committing genocide or apartheid, the reports reliably observe that no court has entered a verdict on the claim – a formulation that appears in 20 of the 124 reports according to the Guardian’s count. Casualty and aid figures are attributed to “a party to the events” , a caveat used in 55 reports and applied to Israel’s own counts as readily as to Palestinian ones. The reader is left with the impression that little about Gaza can be established. One study – a 2022 analysis of antisemitic comments on the Facebook pages of British, French and German media outlets – is cited 454 times across 88 of the 124 reports, including in pages on apparently unrelated topics including land registration in the occupied territories , Aipac’s influence on US media coverage , the Gaza death toll , the 1948 expulsions and the blockade of Gaza . Cleveland-Stout noted the cumulative effect of this inclusion. “Almost every article concludes with saying, well, just talking about this though could contribute to antisemitism,” he said. “And it’ll throw in the study about Facebook comments in Britain.” Building for the machines The Guardian’s analysis found that initially the site carried a kind of file that tells AI systems what a site contains and how to read it (an llms.txt file), but which did not describe the Hanover Institute at all. It was the standard file of Res , an “AI-native content platform that helps B2B [business-to-business] teams get cited by ChatGPT, Perplexity, Claude, and Gemini”, which sells what the industry calls generative engine optimization. The Guardian archived the file on 13 August. It has since been replaced with a bespoke version written for machine readers, and Res is no longer named anywhere on the site. Res’s founder, Hai Tran, registered with the US justice department on 22 July as a subcontractor to Piro on the Israeli account, giving his position as “digital strategist”. Tran told Politico that his role was disclosed in that filing. The Guardian contacted Res for comment. Piro’s own website advertises a service it calls AI Story Optimization, under which it will “author content engineered for how LLMs evaluate credibility”, with the work shipping “on your site or on trusted third-party properties we build”. Cleveland-Stout said the more consequential target was not the possibility of users getting to the site with a live search, but the training data. “There’s kind of two ways to do it,” he said. The first is publishing sites and hoping chatbots cite them. The second, “the more concerning and potentially more effective thing”, is getting material into repositories such as Common Crawl that supply the underlying training data for commercial models. “When it regurgitates narratives the chatbot is trained on, it won’t even cite the sources,” he said. “You won’t be able to fact-check it.” The older network has already made that journey. The seven sites run by Clock Tower X – the firm run by former Trump campaign manager and digital director Brad Parscale – appeared 294 times in Common Crawl’s July index, the repository that supplies training data to commercial AI developers. The Hanover Institute does not appear yet, its reports having gone up after that crawl closed. It is at the start of the same pipeline. Every page of the Hanover site carries a foreign-agent disclosure at its foot. Cleveland-Stout put that to Perplexity, which he said told him that users could visit the website and judge its trustworthiness for themselves. “But how many people are doing that?” he said. The Guardian’s tests of chatbots at the time of reporting showed that scrutiny of the Hanover Institute’s output had already made a mark on chatbot content, however. Asked for new research published by the Hanover Institute, ChatGPT’s webchatbot offered links to four reports , but also warned that the institute was subject to “significant controversy over who funds it and why it exists”, offering a gloss of and links to Quincy Institute’s reporting. The chain Piro is contracted for $1m across two work orders according to Fara filings, but this sum is