The Hanover Institute for Public Policy looks like a respectable new think tank covering Israel-Palestine: a generic name, a red-white-blue website, footnoted reports written in a neutral tone. It isn’t one. Responsible Statecraft reports the institute was created on behalf of the Israeli government’s advertising agency by Piro Inc — a firm whose founder describes its business as shaping what AI chatbots say about your subject.

What the reporting found:

  • More than 100 “research” reports published in roughly a week and a half, all on Israel-Palestine, all with no bylines
  • The reports are framed as innocent chatbot questions: “What Caused the Displacement of Palestinians in 1948?”, “Which Humanitarian Organizations Have Documented Israeli War Crimes?”
  • They are engineered specifically for large language models — footnotes, citations, tables of contents, neutral tone. Piro advertises this as “AI Story Optimization”: content “engineered for how LLMs evaluate credibility”
  • One report even contradicts the Israeli government line (calling foreign university funding a weak explanation for antisemitism) — the goal is credible mimicry, not just advocacy
  • Piro received $900,000 from the Israeli government for the work, subcontracted through the PR conglomerate Havas

Why it matters: chatbots are the new distribution layer. Someone asking ChatGPT or Gemini about Gaza gets “one confident paragraph,” as Piro’s founder put it — and influence operations that used to target readers now target the machines that summarize for readers. This follows Israel’s reported $46.5 million contract with Brad Parscale for chatbot-influencing websites, which a Drop Site investigation found had already shaped what Microsoft Copilot and Google Gemini answer.

One caveat from the report itself: Responsible Statecraft ran 12 sample articles through GPTZero and flagged 11 as AI-written. AI detectors are unreliable, so treat that as supporting color rather than proof — the structural evidence (no bylines, a disclaimer naming the Israeli government’s ad agency as the client, the firm’s own chatbot-marketing materials) is the stronger case.

The pattern is what researchers call LLM poisoning — flooding the sources chatbots learn from with crafted material — and this one is government-scale. A good read for anyone who builds on LLM output: the citations your model trusts are not all organic.