Israel's $45M AI Op Targets US Voters and Chatbot Answers

The Wall Street Journal reports that Israel has paid Brad Parscale, Donald Trump's 2016 and 2020 digital campaign manager, more than $45 million for a US influence operation built around AI-generated text messages and content designed to steer how chatbots like ChatGPT and Claude answer questions about Israel. Millions of Americans have reportedly received texts from senders using names like 'Emma,' 'Sarah,' or 'John,' presented as members of a group called 'Friends for Peace.' When the WSJ asked whether that group is a real organization, Parscale replied, 'What do you define as a real organization?'

The chatbot-grooming angle is what separates this from an ordinary foreign lobbying story. Parscale's team has reportedly stood up websites, articles, and social posts explicitly aimed at how AI systems respond to questions about Israel, with the outlet naming an alliance site called Allyvia.org as one that has surfaced in chatbot answers. Sitting behind all of it is a 2026 Israeli budget line of more than $700 million for what Foreign Minister Gideon Sa'ar has called efforts to 'shape consciousness,' more than four times what was allotted in 2025.

For anyone building on OpenAI or Anthropic, this is the first well-reported instance of a state-scale actor treating chatbot answers themselves as a persuasion surface, not through prompt injection but by feeding the open web that LLMs quietly graze on. The uncomfortable question is whether current retrieval pipelines can distinguish paid, coordinated content from organic sources at all, and whether provider trust-and-safety work extends to demoting flagged campaigns in cited results. Steve Bannon, quoted in the same story, asked how 'the second-largest conservative talk network being run by a registered foreign agent,' a reference to Parscale's role as chief strategy officer at Salem Media, is allowed to stand.

Take the specifics as reported, not settled. The WSJ story does not publish a full list of URLs that ChatGPT or Claude have actually cited, and it does not include responses from OpenAI or Anthropic on whether they have detection tuned for this campaign. What is clear is that the 'grounded' answer that reads like a neutral summary of the open web is becoming a targeted asset. The move for model providers, and for any product built on retrieval over the public internet, is visible source-provenance work, so users can see whose money is behind an answer.