A U.S. Justice Department FARA filing shows Havas Media Germany hired Clock Tower X for an Israel-backed U.S. communications campaign, funded with $1.5 million a month for four months plus potential matching funds. The September 2025 statement of work calls for SEO, new websites, and content intended to shape GPT framing; later reporting tied the firm to a broader site network, but the evidence reviewed does not prove that any AI model's training or answers changed.
A Foreign Agents Registration Act filing received by the U.S. Justice Department on September 18, 2025 shows that Havas Media Germany hired Clock Tower X for a U.S. communications campaign connected to Havas's work on behalf of the State of Israel. The filing does not show a contract between Clock Tower X and any AI company.
The attached statement of work says Havas would contribute $1.5 million per month for four consecutive months, a $6 million client-funded pool, alongside any matching funds contributed by Clock Tower X. It describes the campaign as an effort to combat antisemitism in the United States.
What the filing requires
The official filing lists a broad content and distribution program: at least 100 root creative assets and 5,000 variants per month, at least 80% of content tailored to Gen Z audiences, and paid distribution targeting 50 million monthly impressions. It also calls for monthly search-engine optimization using MarketBrew AI and at least 20 pages per month across priority search terms.
One deliverable is the deployment of websites and content to produce "GPT framing results on GPT conversations." That wording documents an attempt to shape the web information environment around chatbot answers. It does not establish that a model provider agreed to alter training data, ranking systems, or product behavior.
Responsible Statecraft reported the filing on September 29, 2025 and highlighted the same audience, search, and chatbot-framing provisions. The outlet identified former Trump campaign manager Brad Parscale as Clock Tower X's leader and reported that the filing named an Israeli Foreign Ministry official as the campaign contact.
Later reporting describes a larger network
Anadolu Agency reported on July 29, 2026, citing an investigation by Drop Site News, that a later or expanded Clock Tower X campaign was valued at $46.5 million and involved hundreds of articles across sites including Allyvia.org, FactSignal.org, and Paxpoint.org. Anadolu also reported that some linked pages appeared in Microsoft Copilot and Perplexity answers and had been archived by Common Crawl.
The reviewed DOJ filing substantiates the initial $6 million client contribution and its AI-search deliverables; it does not itself substantiate the later $46.5 million figure. Treating those amounts as separate claims avoids presenting a reported expansion as though it were contained in the original filing.
What the evidence does not show
A page appearing in Common Crawl or a search-grounded chatbot citation does not prove that a specific model was trained on it, that it affected model parameters, or that it durably changed answers. No reviewed source establishes cooperation by OpenAI, Anthropic, Google, Microsoft, or Perplexity with the campaign.
For ML teams, the documented tactic reinforces the need for provenance checks, source-diversity controls, adversarial evaluation, and clear disclosure in retrieval-augmented systems handling politically contested claims.
Key Points #
- 1The DOJ filing documents a $6 million Havas-funded pool, plus potential matching funds, and explicitly calls for SEO, new websites, and GPT-framing work.
- 2A later $46.5 million figure comes from July 2026 reporting about a broader or expanded network, not from the original filing.
- 3Common Crawl inclusion and chatbot citations do not prove model training, provider cooperation, or durable changes to AI answers.
Scoring Rationale #
The official filing documents an organized attempt to influence the web and search environment used by AI systems, creating a meaningful information-integrity concern for search-grounded applications. Evidence of direct model-training influence or durable changes to model behavior remains unproven, limiting the score.
Sources #
Primary source and supporting public references used for this report.
Practice interview problems based on real data
1,625 SQL & Python problems across 15 industry datasets — the exact type of data you work with.