Israel creates fake think tank in likely attempt to dupe AI chatbots

Who Should Read

A backend developer whose web-search or RAG pipeline feeds external documents to an LLM as evidence

Core Mechanics

• A think-tank-shaped content mill for chatbots
  • The Hanover Institute has published over 100 reports since Aug 6, all on Israel and Palestine.
  • None have bylines, and a small footer disclaimer says Piro, Inc created it for the Israeli Government Advertising Agency.
• Built in the format LLMs favor
  • The reports have footnotes, tables of contents, and a neutral tone, and start with questions people might ask a chatbot.
  • A NewsGuard analyst said LLMs favor concrete statistics and strong citations, which these articles all have.
• Piro sells 'AI Story Optimization'
  • Piro's site says it authors content engineered for how LLMs evaluate credibility.
  • Others call this practice 'LLM poisoning', meaning seeding data to skew chatbot answers.
• The text looks AI-written
  • GPTZero flagged 11 of 12 random articles as AI-written with high confidence, and one with moderate confidence.
  • The reports claim to cite peer-reviewed research but often cite Israeli government sources like the IDF and the Foreign Ministry.

Community Reactions

• Commenters expect this to become routine
  • One said fake personas and organizations will drown out real information, forcing chatbot makers to seek authoritative sources.
  • They added that such authority sites could charge for data access.
• Comparison to a CPU backdoor
  • One commenter argued that skewing LLM data is like a backdoor in a CPU, since LLMs are core components of automations.
  • They gave examples such as an offensive-language filter flagging certain political views unintentionally.
• It reportedly didn't work well
  • One user said Googling 'the Hanover institute' returned an AI summary citing Palestinian activist sites and no link to Hanover itself.
  • Asking for the website produced a URL that wasn't clickable.
• Pushback that the article overreaches
  • One commenter said the site's About page discloses FARA registration 7732 and its funder, so it isn't hiding anything.
  • They called it LLM-friendly SEO and said no-byline publishing is a normal institutional convention.

How to Apply

• Audit your RAG source domain allowlist
  • Today, pull the top 50 domains your search or crawl pipeline retrieves.
  • Manually check unfamiliar think-tank or institute sites and add them to an allow or block list.
• Store trust signals as metadata
  • Add fields per document for byline presence, sponsor or ad-agency disclosure, and registrations like FARA.
  • When showing citations, label sources as 'no byline, funder disclosed'.
• Require multi-source checks on sensitive topics
  • Add a rule that a claim must be confirmed by 2+ independent sources before it appears in an answer.
  • Count repeated citations of the same statistic from one domain as one source.
• Use AI-text detection only as a secondary signal
  • Use GPTZero-style scores to prioritize review, not to auto-block.
  • Combine them with the metadata checks above.

Try It

# Paste into Claude or ChatGPT (simple checker based on the article's case)

You are a source-credibility checker for a RAG pipeline. Assess whether the web document below is likely influence content designed to be cited by chatbots.

Checklist:
1. Are the author or responsible institution named?
2. Are sponsors or ad-agency relationships disclosed?
3. Does it repeatedly cite only parties to the events?
4. Are titles mass-produced as questions people would ask a chatbot?
5. Does it have stats, footnotes, and a table of contents but few independently verifiable sources?

Document:
{{document}}

Output: yes/no per item with a one-line reason, then a risk level (low/medium/high) and a verdict on whether to include it in RAG.

Next step: put the 5 domains your service cites most into {{document}} and compare the results.

Terminology

Related Papers

• You could have invented transformers from memory

  Viewing linear, MLP and attention layers as memory stores: capacity goes ~64 to ~2,700 to over a million

• Pointing AI at archives found a forgotten meteorite, lost rhinos, and more

  Embedding search plus a cheap filter model surfaced unrecorded events in 400 years of archives

• EmbeddingGemma 2: An open, lightweight multimodal embedding model

  One open 740M model embeds text, images, audio and video into a single shared vector space

• RAG-NAROK: Retrieval-Aware Knowledge Corpus Poisoning in RAG with Source-specific Refutation

  Refutation documents targeting retrieved sources took top rank and evaded anomaly detectors up to 88.37%.

• Text, Pixels, or Both? Evaluating Input Representations for Multimodal Document QA

  Page images beat OCR text on accuracy but cost more; a question-only router cut latency 30%.

• Microsoft exec called AI scraping 'the largest theft of labor in human history'

  NYT lawsuit unredacted filings reveal Microsoft and OpenAI internal views on scraping

Related Resources

• https://responsiblestatecraft.org/israel-influence-chatgpt/
• https://hanoverinstitute.com/hanover/about
• https://ads.openai.com/