This is a research prototype. The data and analyses are preliminary and not yet validated — we'd welcome your .
A pro-Israel influence campaign covertly deployed a fake think tank website and used Generative Engine Optimization platforms to manipulate the training data and outputs of commercial AI chatbots like ChatGPT and Perplexity.
An Israeli government-backed communications campaign reportedly operated a website presenting itself as the Hanover Institute for Public Policy, despite reporting finding no corresponding established think tank. The site published material designed for chatbot retrieval, and testing found ChatGPT and Perplexity citing Hanover content in neutral queries.
Domain classification, causal taxonomy, severity scores, and national security assessments were LLM-classified and may contain errors.
Using AI systems to conduct large-scale disinformation campaigns, malicious surveillance, or targeted and sophisticated automated censorship and propaganda, with the aim of manipulating political processes, public opinion, and behavior.
Also classified as
The campaign aims to pollute AI training repositories like Common Crawl and manipulate chatbot outputs to distort public consensus on geopolitical events.
Commercial LLMs lack the robustness to detect and filter out coordinated Generative Engine Optimization (GEO) and data poisoning attempts.
The incident involves a coordinated, state-funded influence campaign designed to manipulate public opinion at scale by gaming AI chatbot outputs on sensitive geopolitical topics.
Human
Due to a decision or action made by humans
Intentional
Due to an expected outcome from pursuing a goal
Post-deployment
Occurring after the AI model has been trained and deployed
The risk is caused by human actors intentionally designing and executing a campaign to manipulate the outputs of deployed AI systems.
Added categories peak at 3: Substantial (Epistemic Harm, Inferred)
This incident represents a notable shift in information warfare, where a state actor covertly funded a campaign to groom commercial AI chatbot outputs using Generative Engine Optimization. By seeding biased narratives into web repositories, the campaign successfully manipulated ChatGPT and Perplexity outputs on sensitive geopolitical issues, establishing a concerning precedent for AI-era disinformation.
The AI systems targeted by the campaign are commercial large language models like ChatGPT, Claude, Gemini, and Perplexity, which function as conversational assistants. These systems augment human capabilities by answering questions and providing information, but require human oversight to verify their outputs.
Limited Risk: The incident involves chatbots and AI-generated content platforms (like ChatGPT, Perplexity, and Res) which are subject to transparency obligations under the EU AI Act to ensure users are aware they are interacting with AI-generated outputs.
The campaign specifically targets commercial chatbots and question-answering engines like ChatGPT and Perplexity to manipulate their conversational outputs on geopolitical topics.
The reports from The Guardian and Politico provide detailed, corroborating evidence of the influence campaign, including specific FARA filings, contract amounts, and technical details of the Generative Engine Optimization (GEO) techniques used. The role of the AI systems (both the Res platform and the target LLMs) is explicitly documented, and the testing of chatbots confirming the campaign's success provides direct empirical support.
The campaign could be viewed purely as a traditional public relations or 'Hasbara' effort, but the explicit technical targeting of LLM training data and search optimization (GEO) makes it a distinct AI-centric information manipulation incident.
Technology
LLM Susceptibility to GEO
Chatbots are easily manipulated by content optimized for retrieval.
Lack of Source Verification
LLMs cite or regurgitate unverified sites like fake thinktanks.
Generative Engine Optimization
Platforms like Res enable active manipulation of chatbot responses.
Data Inputs
Unchecked Training Repositories
Common Crawl ingests biased web content without quality checks.
Manipulated Q&A Formatting
Site content is structured specifically to mimic user queries.
Use of LLM-targeted Files
Use of llms.txt files designed to guide machine readers directly.
Human Factors
User Overreliance on Chatbots
Users trust chatbot outputs as neutral without checking sources.
Deceptive Website Presentation
Professional site design and fake credentials deceive reviewers.
Process and Methods
Inadequate Content Filtering
Chatbot developers lack robust methods to filter coordinated influence.
Lack of Algorithmic Attribution
Difficulty in identifying automated or optimized state-sponsored content.
Management
Commercialization of GEO Platforms
Platforms like Res prioritize profit over the integrity of AI outputs.
AI Platform Risk Blindness
AI developers failed to manage risks of coordinated optimization.
Regulatory Environment
Weak AI Data Governance Laws
No regulations prevent the intentional poisoning of AI training data.
Lack of LLM Grooming Regulation
No clear legal framework addresses the manipulation of chatbot outputs.
AI system involved
Identified by automated extraction; can differ from the alleged developers above.
