Israel Pays Millions to Influence How AI Chatbots Discuss Gaza and Experts Say It's Working

Israel Pays Millions to Influence How AI Chatbots Discuss Gaza and Experts Say It's Working

An investigation alleges Israel targeted ChatGPT, Google Gemini, and Perplexity with LLM poisoning, raising fresh concerns over AI chatbot manipulation and disinformation.

LS
Linsey Smith
Jul 29, 2026
2 min read

The Israeli government has spent millions of dollars on a covert operation designed to manipulate how AI chatbots like ChatGPT, Perplexity, and Google Gemini discuss the war in Gaza, an investigation by Drop Site News reveals, and disinformation experts confirm the campaign is successfully infiltrating both live chatbot responses and underlying training data.

Former Trump campaign manager Brad Parscale, operating through his firm Clock Tower X, signed a $46.5 million contract with the Israeli government to build a network of at least 10 websites promoting pro-Israel narratives. The sites cover topics ranging from alleged links between Gaza journalists and Hamas to the claim that Israel's military actions comply with international law. The intended audience is not primarily humans, most sites average only a few hundred monthly visitors, but rather artificial intelligence systems (1).

Parscale explicitly framed the goal as "LLM poisoning," stating in his original contract that he would deploy "websites and content to deliver GPT framing results on GPT conversations". When Drop Site tested Perplexity on whether the U.S. should enhance military cooperation with Israel, the chatbot answered "Yes" and cited Allyvia.org, one of Parscale's sites, as its top source. Microsoft Copilot similarly pulled from Clock Tower X websites without disclosing their origin as Israeli-funded content.

More alarming, the sites are now being ingested into Common Crawl, the massive data repository that has trained roughly 80% of OpenAI's GPT-3. Drop Site found the 10 sites were crawled 912 times between January and June 2026, with monthly crawl frequency climbing from just 2 in January to 376 in May. Research from Anthropic suggests that as few as 250 malicious documents can create a "backdoor" vulnerability in a large language model regardless of its overall training-data size . Chatbots tested by Drop Site alongside Check First CEO Hervé Letoqueux confirmed that Gemini and Copilot contain Clock Tower X content in their training data.

About the Writer

More from Mindplex

Keep reading

Three more ideas worth your time.

Browse MindBytes

Discussion

Join the discussion

Sign in to share a response with the community.

Type @ to mention someone Type / or use + to add a block Highlight text, then choose Link
Loading editor

Comments cannot be edited after posting because they become part of the reputation record. Give yours a quick review first.