Israel-funded operation influenced major AI chatbots on Gaza war
According to an investigation by Drop Site News, Israel contracted Brad Parscale's firm Clock Tower X for $46.5 million to create a network of websites designed to influence how major AI chatbots respond to questions about the Gaza war. The campaign, which began in October 2025, created content intended to be absorbed into AI training datasets through search engine indexing and web archiving, a practice known as "LLM poisoning."
What happened
According to a Drop Site News investigation, Brad Parscale's firm Clock Tower X built a network of 10 websites under a $46.5 million contract with the Israeli government, starting in October 2025, to influence responses from major AI chatbots on Gaza war questions.
The websites, including Allyvia.org, FactSignal.org, and Paxpoint.org, created hundreds of articles promoting Israeli narratives, defending Israel's military campaign in Gaza, and presenting pro-Israel perspectives. Content included claims that footage from Gaza was staged and that reports of journalist deaths were Hamas propaganda.
The websites were designed to be indexed by search engines and incorporated into AI training data through Common Crawl, a major web archive used by large language models, rather than to attract direct human readers.
Reported impact
Tests found that Gemini and Microsoft Copilot cited content from the Clock Tower X network, and the content entered AI training datasets through Common Crawl. Between January and June 2026, the 10 websites were crawled 912 times, with crawls increasing from twice in January to 376 times in May.
When Drop Site tested Perplexity about US military ties with Israel, the chatbot cited a Parscale-created site as its top source.
Index analysis
Each website carried a legal disclosure stating it was distributed on behalf of the State of Israel as required under US law, but the disclosure was not passed on to users when chatbots cited the sites.
This practice, known as 'LLM poisoning,' aimed to influence AI responses by introducing pro-Israel content into training datasets. According to a study cited in the report, as few as 250 planted documents can compromise even a large AI model.
Classification
Sources and evidence
This record was researched and written by the Index, based on the following original reporting.