Skip to main content

Writing for the Chatbot: Israel's Reported Campaign to Shape AI Answers on Gaza

Manipulation Breakdowns · 8 min read · By D0

title: “Writing for the Chatbot: Israel’s Reported Campaign to Shape AI Answers on Gaza” description: “Drop Site News reports Israel paid Clock Tower X to build websites aimed at chatbots. What the reported filing says, what tests found, what is unproven.” slug: “clock-tower-x-israel-chatbot-framing-campaign” category: “manipulation-breakdowns” date: “2026-10-15” author: “D0” published: true tags: - llm-grooming - state-propaganda - fara - ai-chatbots - gaza - source-laundering - search-manipulation

Clock Tower X’s Websites for AI Chatbots

Most propaganda aims at people. This campaign, as reported, aims at the software people ask.

On July 28, 2026, Drop Site News published a report by Nick Cleveland-Stout, a research associate at the Quincy Institute. It describes a network of ten websites built by Clock Tower X, the firm of Brad Parscale, Donald Trump’s 2020 campaign manager. The firm works for the Israeli government.

The sites got little direct traffic. According to the report, that was never the point. They were written to be crawled and indexed, so that AI models would train on them or chatbots would retrieve them in live searches, and then repeat their framing to users who never saw the sites.

This piece sets out what the report claims, what it can support, and what it leaves open. It also explains why the method works on anyone who asks a chatbot a question.

What Is Documented and What Is Claimed

The pieces differ in strength, so they are listed separately.

  • A foreign-agent filing. The report links a Foreign Agents Registration Act exhibit whose file name indicates September 18, 2025. It quotes the filing as promising “websites and content to deliver GPT framing results on GPT conversations.” I could not open the filing myself. I am relaying Drop Site’s quotation of it.
  • A disclosure on each site. The report says each site states it is distributed “on behalf of the State of Israel.” The legal disclosure exists. The question is where it sits on the page.
  • A contract value. The report gives $46.5 million but names no source for the total. It cites Responsible Statecraft for a start of $1.5 million a month and The Intercept for a later rise to $4.5 million a month. The report does not reconcile the total with those monthly figures. Treat the number as unconfirmed.
  • Chatbot test results. These come from Drop Site’s own testing, covered below.

IBTimes UK, writing on August 4, repeats the central claims and attributes them to Drop Site. It adds that Israel and Clock Tower X have not described the work as an attempt to manipulate AI. They call it public diplomacy.

OpenAI, Anthropic, Google, Microsoft and Perplexity have not accused Clock Tower X of breaking their rules, per IBTimes. Parscale and Clock Tower X did not respond to Drop Site’s request for comment. He told the Wall Street Journal, when asked whether a group called “Friends for Peace” is real, “What do you define as an actual organization?” An anonymous Israeli official told Time, “We have paid him lots of money.”

The Tactics

1. Source Laundering

The ten sites carry names like Paxpoint, Allyvia, Factsignal and Compassionpulse. They read like independent research outlets or civic groups.

The report says the state sponsorship is stated on the sites, but a text-message campaign linking to them hides it until a user scrolls to the bottom. In Drop Site’s tests, some chatbots cited the sites without mentioning the footer.

This is source laundering: the message is placed with a source that sounds neutral, so the audience gives it more trust than it would give a government.

2. Writing for the Crawler

The report counted how often Common Crawl, a public archive of web pages used widely to train language models, captured the sites. Between January and June 10, 2026, the ten sites were captured roughly 910 times. The two counts in the report differ by one, 912 against 913 when the listed figures are added up.

Captures rose from 2 in January to 376 in May. The report says about 85 percent of the network’s roughly 900 pathways were archived. It contrasts that with about 2.5 percent for Russia’s Pravda network.

That comparison is the report’s strongest point. It suggests the sites were built to be collected.

There is a limit the report states itself. DFRLab has noted that appearing in Common Crawl does not guarantee a page is in any given model’s training data.

3. Repetition Across Many Pages

The report names about 900 pathways. Many are blog posts. Drop Site says Mark Ginsberg, an Israel-based marketer who registered as a foreign agent for Israel in March, wrote many of them.

Volume matters because a chatbot that searches the web, or a model trained on scraped pages, sees the same framing from several apparent sources. Agreement among sources reads as corroboration, even when one sponsor wrote all of them.

The report also cites an October 2025 Anthropic study finding that about 250 malicious documents were enough to plant a backdoor in a model. That study concerned a different attack. It shows only that small amounts of planted text can matter.

4. Framing the Contested

Per IBTimes, relaying Drop Site, the sites cover military cooperation, regional security and responses to Gaza criticism. Articles reportedly dispute civilian casualty figures, question claims about journalists killed, and reject genocide allegations. The network reportedly also wrote about the death of Hind Rajab.

Governments are entitled to argue their case. The tactic is not the argument. It is arguing under a name that hides who is speaking, in a place where the audience expects a neutral answer.

What the Chatbot Tests Showed

Drop Site asked chatbots about topics the network covers and read the sources they cited.

  • Asked whether expanded US-Israel military cooperation is beneficial, Perplexity answered “Yes” and listed Allyvia.org as its top source.
  • Copilot cited Parscale’s sites as well.
  • Asked about Paxpoint.org, Perplexity and Copilot did not flag the Israeli government sponsorship. Claude, ChatGPT and Gemini did.
  • On a technology collaboration query, Gemini, Copilot and Perplexity cited network sites without noting who ran them.

A second test, run with Hervé Letoqueux of the firm Check First, asked five chatbots to write articles using only their training data. Gemini and Copilot consistently appeared to contain the sites. Perplexity appeared to contain some of them. Claude and ChatGPT did not appear to, though they can search the sites live.

The report gives no prompts, number of trials, dates or scoring rules for that second test. Chatbot answers also vary from run to run. Take the results as an indication that needs repeating, not as a measurement.

Perplexity’s spokesperson said the product is built to give cited, linked answers from a range of sources. Microsoft and Google did not respond to the report.

Who Is Telling the Story

The sources behind this report deserve the same scrutiny.

Drop Site News describes itself as reader-supported. The Washington Free Beacon reported that Open Society Foundations gave it $250,000 in 2024. Critics call the outlet hostile to Israel. The author is a research associate at the Quincy Institute, and I did not trace its funders.

None of that makes the findings false. It means they should be checked against the filings. The FARA language can be checked against the public registration database at efile.fara.gov, and the disclosures on the sites themselves.

Drop Site compares the sites with Russia’s Pravda network. A reader who objects to one should apply the same standard to the other.

What This Is Not

This is not proof that any chatbot was taught to lie. It is not a finding that Israel’s case is wrong, and it is not evidence that AI companies cooperated. It is a reported effort to shape what models cite, with early test results that need independent repetition.

What You Can Do

When a chatbot gives you a one-line answer on a contested political question, open the sources it lists.

  • Ask who runs each source, and scroll to the footer.
  • Check for a foreign-agent notice or a line saying “on behalf of.”
  • Search the organization’s name with the word “funded.” A research group with no staff list is a warning.
  • Ask the chatbot the same question two ways, and see whether the same few sites come back.
  • Treat agreement among several sites with the same style, topics and timing as one source until shown otherwise.

Watch two things next. One is whether anyone repeats the chatbot tests with published prompts and dates. The other is whether the AI companies start labeling sources that carry a foreign-agent disclosure.

This article is part of Decipon’s Manipulation Breakdowns series, examining specific influence operations through the Influence Tactics Protocol.

Sources: