Donald Trump’s former campaign manager Brad Parscale has launched a massive digital operation commissioned by the Israeli government to influence the delivery of popular AI chatbots such as ChatGPT, Claude, and Gemini. Using a network of specially created websites, the perpetrators are attempting to integrate pro-Israel narratives directly into the AI’s training data.
Donald Trump’s former campaign manager Brad Parscale has launched a massive digital operation commissioned by the Israeli government to influence the delivery of popular AI chatbots such as ChatGPT, Claude, and Gemini. Using a network of specially created websites, the perpetrators are attempting to integrate pro-Israel narratives directly into the AI’s training data.
According to journalists from Drop Site , Parscale's firm Clock Tower X has received a $46.5 million contract from the Israeli government. The operation, which has been ongoing since October, involves creating websites that masquerade as fact-checking or news resources. Their main audience is not real people, but search robots that collect data for training AI.
This process in cybersecurity is called “large language model poisoning” (LLM poisoning). The idea is to saturate the internet with specifically structured information that algorithms recognize as authoritative. For example, FactSignal.org publishes materials claiming that most of the journalists killed in Gaza were affiliated with terrorists, while Paxpoint.org promotes the narrative of Israel as an exclusively peaceful nation.
Disinformation experts interviewed by the publication say the tactic is working. When journalists asked the Perplexity chatbot whether it would be beneficial for the United States to increase military cooperation with Israel, the system answered “Yes,” citing the first source as Allyvia.org, a website created as part of the same campaign. Microsoft Copilot showed similar results.
The main platform for “infiltration” was Common Crawl, a giant data repository on which most modern AI is trained (it makes up 80% of the training base for OpenAI’s GPT-3). Analysis showed that Clock Tower network sites were indexed there over 900 times in just six months.
According to NewsGuard analyst Alice Lee, the Israeli campaign is much more sophisticated than straightforward Russian propaganda.
"These websites have bulleted lists, key takeaways at the top of the page, aggressively citing sources — all of which increase the likelihood that the content will be picked up by chatbots. It is much more difficult to set up protective barriers around an opinion or half-truth than around an outright lie," the expert explained.
While some chatbots (notably Claude and Gemini) have begun to label Parscale sources as foreign government-funded, experts warn of a major danger: if these narratives become deeply embedded in the training data itself (as opposed to a typical internet search query), it will be virtually impossible for users to trace the original source of the manipulation.
This is not the first time that countries have used AI “poisoning” to promote their own narratives. It was previously reported that Russian special services are trying to create an alternative information ecosystem of fake think tanks, media outlets, and Wikipedia clones to force Google’s search algorithms and popular AI-based chatbots to present pro-Russian narratives to users as verified facts.
