Israel Spent $100K to Sway ChatGPT Answers on Gaza

For AI teams, the Gaza influence operation isn't a distant policy issue—it's an architecture problem. If a $100,000 PR campaign can seed unattributed articles designed to be cited by retrieval systems, then every RAG-enabled chatbot is vulnerable to coordinated content poisoning. This report should force a re-examination of how AI models select and cite web sources when answering charged questions.

According to a Politico report published on Friday, August 14, 2026 and carried by syndicated sources on August 16, Israel has spent tens of millions of dollars in recent years to improve its standing with US audiences, including a $100,000 campaign specifically designed to shape how large language models such as ChatGPT answer questions about Gaza and the Israel-Hamas war. The campaign was reportedly routed through French public relations firm Havas Media, its subcontractor Piro Inc, and the Israel Government Advertising Agency, known as LaPam. At the center of the operation is the Hanover Institute for Public Policy, a website that publishes anonymous reports framed as questions on anti-Semitism and geopolitical issues, including "Is Torture of Palestinian Prisoners a War Crime?" and "Who Is the Aggressor in the Israel-Palestine Conflict?". Politico reported that FARA documents filed with the US Justice Department linked more than a dozen Hanover Institute articles to a PR campaign whose stated purpose was "the creation and dissemination of factual, source-supported informational materials" intended to "educate" the US public about Israel. The Hanover Institute itself describes its work as "documentary, not declarative" and insists its research is purely academic.

“

The risk is not limited to Israel; if one state can attempt this at a reported cost of $100,000, others are likely already experimenting with similar methods.

The backdrop to the operation is a sharp decline in US public support for Israel. A June 2026 Pew Research Center poll found that 60 percent of US adults now view Israel somewhat or very unfavorably, up from 53 percent last year and 42 percent in 2022. Israeli Prime Minister Benjamin Netanyahu faces even worse numbers, with 65 percent of US adults expressing little or no confidence in his policies. The shift began after the October 7, 2023 Hamas attacks and Israel's subsequent military response, and has deepened more recently as the US-Israeli war on Iran raised concerns about costs to American taxpayers. Against that opinion environment, the reported $100,000 effort represents a comparatively cheap but potentially high-leverage attempt to contest the information space where many Americans now encounter news and analysis: through AI assistants.

The mechanics of the operation point to a new and underappreciated vulnerability in modern information infrastructure. Rather than directly hacking or bribing AI developers, the campaign appears to have targeted the public web content that LLMs increasingly rely on through retrieval-augmented generation and browsing capabilities. By publishing unattributed, question-framed reports and seeding them across the web, the Hanover Institute content was reportedly designed to be picked up and cited by large language models. Politico noted that technical markers linked the website to the campaign. If an AI assistant ingests such material and presents it as a factual answer, the narrative can bypass traditional editorial gatekeepers and reach users with the veneer of machine neutrality. This is a form of coordinated content poisoning whose audience is not solely human readers but also the models that digest the open web.

For the defense and geopolitical community, the episode is a clear case study in low-cost narrative warfare layered on top of kinetic conflict. The campaign combines several elements of modern influence operations: an anonymized front organization, foreign agents registration filings, commercial PR intermediaries, and a target set that includes both public opinion and AI model behavior. It demonstrates that state actors now view commercial language models as contested terrain, not merely as neutral tools. The use of FARA-registered contractors also underscores how influence campaigns can operate in a gray zone, technically disclosed in regulatory filings yet obscure enough that the downstream effect on AI outputs may escape scrutiny for months or years.