Did an Israeli campaign try to influence ChatGPT? The contract, OpenAI’s non-response and the evidence

Fact-check · Updated 28 October 2026 · The November 2025 non-response concerns one specific inquiry.

A 2025 US foreign-agent disclosure brought a new form of political communication into the spotlight: producing web pages that might shape how artificial-intelligence assistants discuss Israel. The documents are real. But some headlines went further than the evidence by suggesting Israel had bought ChatGPT or struck an agreement to retrain its underlying model. The distinction matters for understanding both public diplomacy and independent reporting.

What the US documents actually show

In September 2025, Clock Tower X, associated with political consultant Brad Parscale, registered under the US Foreign Agents Registration Act (FARA). The filing describes work through Havas Media Germany in support of an Israeli-government communications campaign in the United States. The accompanying scope of work set out digital publishing, advertising, audience research and search-engine optimization. It framed the client’s stated purpose as combating antisemitism, while outside observers questioned the effects of government-funded political messaging on information access.

A campaign aimed at search and AI answers

The published draft plan included 100 original creative assets per month and 5,000 adaptations, with at least 80 percent aimed at Generation Z. It targeted a minimum of 50 million monthly advertising impressions. Under its search-related work, the contractor proposed publishing pages and improving their visibility, including a line about achieving “GPT framing results” in AI conversations. This is evidence of an intention to influence the information environment. It is not evidence that the contractor could change an OpenAI model’s internal weights or control ChatGPT responses.

OpenAI offices in San Francisco, archive photograph from June 2025
OpenAI offices in San Francisco, photograph taken June 2025; contextual photograph, not evidence of a contract. Photo: Coolcaesar, CC BY 4.0. Wikimedia Commons(https://commons.wikimedia.org/wiki/File:1515_Third_Street.jpg) · CC BY 4.0(https://creativecommons.org/licenses/by/4.0/)

Why the OpenAI non-response became a story

The Business & Human Rights Resource Centre asked multiple technology companies to comment on allegations that AI and online platforms could be targeted by the initiative. On 14 November 2025 it recorded that OpenAI had not responded to its request. That is a narrow and verifiable fact about a particular inquiry. It does not, by itself, mean that OpenAI approved the campaign, participated in it, or admitted any allegation. In the same case the organization listed several other companies that had not supplied comments.

Is “training ChatGPT” the right description?

A public website can potentially affect future information ecosystems: search rankings, citations, retrieved passages or material available to future data collection. Those processes are not the same as direct model training. Commercial AI systems have independent training pipelines and safeguards, and there is no public proof in these filings that a specific ChatGPT model incorporated the campaign’s materials or subsequently changed its answers. The phrase “Israel signed a deal with OpenAI” is therefore misleading. The disclosed commercial arrangement was with a communications contractor, not with the developer of ChatGPT.

United States Department of Justice building, Washington DC, 2026
US Department of Justice headquarters in Washington, DC. FARA declarations are held by the Department. Photo: APK, CC BY 4.0. Wikimedia Commons(https://commons.wikimedia.org/wiki/File:Robert_F._Kennedy_Department_of_Justice_Building.jpg) · CC BY 4.0(https://creativecommons.org/licenses/by/4.0/)

A separate documented incident in 2024

One earlier case must not be confused with the Clock Tower contract. In May 2024, OpenAI reported that it had blocked accounts associated with an Israel-origin influence operation called Zero Zeno, linked to the political consultancy STOIC. The accounts used AI to draft comments and articles for third-party platforms. OpenAI reported disrupting that use of its tools; it did not describe cooperating with the campaign. This earlier incident shows that influence operations can misuse AI products, but it does not establish that the later Clock Tower effort succeeded.

How readers can assess claims fairly

The strongest method is to read the FARA filing, separate contractual objectives from demonstrated outcomes, identify who was actually a contracting party, and check the dates of company responses. A government may legitimately explain its position, while covert or misleading presentation warrants scrutiny. The same evidentiary standards should apply regardless of whether a campaign favors or criticizes Israel. A chatbot answer is not proof of who influenced it: sources, retrieval systems and independent replication matter far more than a viral headline.

Original documents and further reading

All articles(https://israel.re/en/)

Coolcaesar · Wikimedia Commons · CC BY 4.0 · resized