Israel Funds Campaign to Shape Sources Behind AI Answers

Israel has funded a multimillion-dollar campaign that targets Google News visibility and the web sources behind AI-generated answers. The operation does not require direct access to ChatGPT, Claude, Gemini, or their underlying models. Instead, it seeks to publish, rank, and distribute enough favorable material that AI systems encounter those narratives when searching the web.

That distinction is the heart of the story. Israel is not paying an AI company to secretly rewrite a chatbot. It is funding an outside communications network that tries to influence the information environment feeding search engines, answer engines, and retrieval systems.

The campaign is associated with Clock Tower X, a company controlled by former Trump campaign manager Brad Parscale. Public filings describe work involving websites, social content, search optimization, and material intended to affect framing in GPT conversations. The strategy challenges a reassuring assumption about generative AI: that an answer becomes trustworthy when it includes current web sources.

The Contract Targets Both People and Machines

The campaign treats web publishing as infrastructure for influencing human audiences and automated answer systems at the same time.

Clock Tower X registered under the Foreign Agents Registration Act in September 2025. The filing identifies Havas Media Germany as an intermediary working on behalf of the State of Israel.

A federal contract filing says Clock Tower would provide strategic communications, planning, and media services for a nationwide campaign in the United States. The stated public purpose was combating antisemitism.

The original agreement described a campaign valued at six million dollars. Later filings and reporting placed the expanded Clock Tower contract at nine million dollars, with payments structured at roughly one and a half million dollars per month.

The work was not limited to conventional advertising. Contract language described a “Search and Language Operation,” alongside the deployment of websites and content intended to deliver framing results in GPT conversations.

That wording matters because it identifies AI answers as an explicit campaign objective. It does not merely describe using AI to produce advertisements faster.

The planned campaign also set large distribution goals. According to reporting based on the filing, at least 80 percent of the material was intended for Generation Z audiences. Target platforms included TikTok, Instagram, YouTube, podcasts, and other digital outlets.

The campaign sought at least 50 million impressions each month. Those impressions would help messages reach people directly, but distribution also creates links, citations, reposts, and engagement signals that search systems can detect.

Clock Tower reportedly planned to use Market Brew, an optimization platform that models how search engines might rank content. Search engine optimization, or SEO, is the practice of making pages easier for search systems to discover and rank.

There is nothing unusual about an organization using SEO. Companies, political campaigns, newsrooms, governments, and advocacy groups all compete for visibility.

The unusual part is the declared objective of shaping how AI systems frame conversations. The campaign appears designed to connect traditional influence techniques with the source-selection processes used by modern answer engines.

That creates a wider surface than one chatbot. A page that ranks in ordinary search can enter Google News discovery, appear in an AI-generated summary, or become a citation in a chatbot response. It can also be copied by other publishers, multiplying its apparent presence.

An early contract investigation reported that the operation would create and deploy content across multiple platforms while developing websites intended to affect GPT framing. Later coverage found that the effort continued beyond the initial agreement.

Clock Tower and Parscale have disputed characterizations of the project as a covert foreign influence operation. The company’s FARA registration, however, formally discloses work connected to a foreign principal.

The distinction is legally and editorially important. FARA registration does not establish that every claim published by a registrant is false. It tells readers that the work was performed within a paid relationship involving a foreign government.

Google News users rarely see that context before opening a story. AI users are even less likely to encounter it when a system extracts one claim from a page and compresses it into an answer.

Why Google News Is Part of the AI Supply Chain

Google News is no longer only a destination for readers because its indexed stories can also become evidence for machines.

Generative AI products rely on several information paths. Their base models learn statistical patterns from large collections of data assembled before training. Many products then use live search, licensed databases, or retrieval systems to answer questions about recent events.

Retrieval-augmented generation, usually called RAG, is a process that gives a model selected documents before it writes an answer. The model uses those documents as temporary evidence without permanently retraining on them.

This process makes a chatbot more current. It also transfers considerable power to the systems selecting the documents.

If a search layer retrieves reliable reporting, official records, and diverse analysis, the model has a stronger foundation. If it retrieves coordinated advocacy pages that resemble independent sources, the final answer can inherit their framing.

Google News matters within this environment because it organizes a large part of the public news web. Its clusters, indexing decisions, and search visibility affect what readers find. They also influence which pages other search and monitoring tools encounter.

That does not mean Google News directly trains every major AI model. It also does not mean inclusion guarantees that a chatbot will cite a page.

The connection is more indirect and more difficult to audit. A story gains visibility through search, gets linked by another publisher, appears in a dataset, and later becomes available to an AI retrieval tool. Each stage can amplify the previous one.

AI companies use different retrieval arrangements, ranking systems, and safety controls. OpenAI, Anthropic, Google, Microsoft, and other providers do not publish complete lists of every source considered for every answer.

Users therefore see only the final selection. Even when citations appear, they rarely reveal why one source outranked another or whether multiple cited pages trace back to the same campaign.

This is where content networks gain leverage. Ten websites can present the same argument with different branding, headlines, and domain names. A search system might interpret them as independent confirmation unless it detects their shared ownership or coordinated production.

Google News has policies intended to evaluate transparency, originality, deceptive practices, and publisher accountability. Enforcement across the open web remains difficult because relationships can run through contractors, advertising agencies, shell companies, or loosely branded projects.

A March 2026 website investigation identified a network of sites publishing favorable material about Israel. Parscale confirmed efforts to affect search and chatbot results, while rejecting the label of a foreign influence campaign.

That acknowledgment removes one major ambiguity. The effort to influence machine-readable sources is not merely an interpretation imposed by critics. It corresponds with both the contract language and the operator’s description of the strategy.

What remains uncertain is the campaign’s actual effect. Finding a campaign-backed page in an AI answer proves that the page entered a retrieval path. It does not prove that the campaign changed the model’s overall position, permanently altered training data, or controlled future responses.

Results can vary with the user’s wording, location, account settings, search provider, model version, and date. A source appearing in one answer can disappear in the next.

Still, successful placement matters. Many users ask one question, read one summary, and move on. A campaign does not need permanent control if it can influence the small set of sources seen at the moment of decision.

For knowledge workers, this makes source lineage essential. A personal or organizational AI knowledge base is only as dependable as its underlying documents and provenance records.

The Real Contest Is Source Selection, Not Model Training

The campaign’s central bet is that controlling discoverable evidence can be more practical than changing an AI model itself.

Public discussion often describes the operation as an attempt to “train ChatGPT.” That phrase is understandable, but technically imprecise.

Clock Tower cannot independently add material to OpenAI’s training dataset. It cannot force Anthropic to retrain Claude on a campaign website. It cannot instruct Google to treat a specific article as authoritative.

What it can do is manufacture a large supply of searchable material. That material can target phrases people use, answer predictable questions, earn links, and compete for placement in search results.

This is closer to information placement than direct model training. The campaign seeks to make preferred claims available at the moments when a machine gathers evidence.

The approach reflects a change in how online persuasion works. Traditional campaigns targeted voters through television spots, mailers, search advertisements, and social feeds. Those channels remain important, but AI introduces an intermediary that summarizes material before a person sees it.

That intermediary can erase visible signs of persuasion. A reader opening an advocacy website can inspect its branding, tone, authors, and disclosure. A chatbot might reduce the same page to one neutral-sounding sentence.

The source may still be linked, but the persuasive context is weaker. Many users do not open citations, especially when the answer sounds confident and includes several references.

The Clock Tower strategy therefore pits source transparency against source saturation. Transparency asks users and platforms to examine who produced a claim. Saturation aims to ensure that preferred claims are repeatedly available across the search environment.

This model does not belong to one country or political position. Any government, corporation, lobbying group, activist network, or commercial operator can attempt it.

The barriers are relatively low. A coordinated group can generate articles, translate them, build specialized websites, and optimize pages much faster with generative AI. Automation also makes it easier to revise the network when political priorities change.

Reporting indicates that the Clock Tower sites changed their emphasis as events changed. Early material reportedly presented Israel as peaceful and technologically productive. Later content addressed the wars involving Gaza and Iran, alongside criticism of journalists and international institutions.

That responsiveness reveals the campaign’s operational advantage. A static propaganda archive becomes stale. A continuously managed source network can target the exact questions rising in search traffic.

Google News and chatbot systems face related problems here, but their jobs are not identical. A news aggregator must identify publishers, cluster similar coverage, and rank timely stories. A chatbot must select sources, reconcile claims, and compose an answer.

Both can fail when apparent diversity conceals common coordination. Five domains do not represent five perspectives if one contractor directs all five.

AI answers add another complication. Models often blend sources at the sentence level, making it difficult to identify which page contributed each phrase. Citation interfaces can suggest stronger traceability than the system actually provides.

Campaign operators do not need every claim accepted. They can focus on ambiguous questions where framing carries more weight than a single verifiable fact.

Consider a user asking whether a military action was justified, whether media coverage is biased, or whether a country remains a reliable ally. These questions require selecting events, legal interpretations, and moral premises. Source choice can determine the boundaries of the answer before generation begins.

A campaign-backed article might emphasize attacks on Israeli civilians, regional security threats, or disputed reporting practices. A Palestinian rights organization might emphasize civilian deaths, displacement, occupation, or restrictions on aid.

A dependable answer should identify those perspectives and separate verifiable facts from contested interpretation. A source-optimization campaign benefits when the system instead treats one perspective as neutral background.

This is why the issue cannot be solved by banning one set of domains. That response would invite arbitrary censorship and create another avenue for political pressure.

Platforms need stronger provenance signals. These can include ownership disclosures, FARA status, shared analytics identifiers, identical publishing patterns, common hosting infrastructure, and repeated language across nominally separate outlets.

They also need diversity measures based on institutional independence, not domain count. Three pages funded through the same contract should not carry the same evidentiary weight as three unrelated organizations.

Evidence of Influence Is Real, but Control Is Not Proven

Campaign pages appearing in AI responses demonstrate access to the pipeline, not command over the final answer.

Recent reports have treated the appearance of Clock Tower-linked material in chatbot results as evidence that the strategy is working. That conclusion needs careful boundaries.

If an AI system cites one of the network’s pages, the campaign has achieved discoverability. It has placed material inside the system’s pool of usable evidence.

That is a meaningful result. The contract explicitly described efforts to influence search and GPT framing, so retrieval aligns with the stated objective.

It is not the same as proving broad persuasion. Researchers would need repeated, controlled tests across models, prompts, locations, languages, and dates to establish a systematic shift.

Those tests would also need a baseline. Without archived responses from before the campaign, analysts cannot confidently attribute a change to Clock Tower rather than current events, model updates, or other newly published information.

Independent audits face another problem: AI answers are unstable. The same question can produce different sources during separate sessions. Search indexes also change as pages gain or lose visibility.

A proper study would record the full response, citations, model version, retrieval time, user location, and prompt wording. It would then classify cited sources by ownership, funding, and editorial independence.

No publicly available evidence reviewed for this article establishes that Israel gained durable control over the general behavior of ChatGPT, Claude, or Gemini.

The stronger verified claim is narrower. Israel funded a disclosed communications contractor whose plans included websites, search optimization, large-scale distribution, and attempts to affect GPT framing.

An April investigation reported that Israel had paid Clock Tower nine million dollars and renewed its work. It also described concern among pro-Israel advocates that AI platforms were drawing heavily from material they considered hostile to Israel.

That concern helps explain the timing. Public support for Israel had weakened during the Gaza war, particularly among younger Americans.

A 2025 international opinion survey found predominantly unfavorable views of Israel in many surveyed countries. Views of Israeli Prime Minister Benjamin Netanyahu were also broadly negative.

An influence campaign can increase message visibility, but it cannot automatically reverse those attitudes. Public opinion responds to policies, personal experience, trusted relationships, images from conflict zones, and reporting from many institutions.

This creates the campaign’s central weakness. It treats a legitimacy problem partly as an information-distribution problem.

More favorable content can fill search gaps or challenge false claims. It can also backfire if audiences discover that apparently independent sites belong to a government-funded network.

Disclosure is therefore not a minor compliance issue. It changes how a reasonable reader evaluates the material.

FARA provides one disclosure mechanism for activities in the United States. Yet a reader must already know to search the federal database, connect a domain to a contractor, and interpret the filing.

AI interfaces could surface that context automatically. A citation label might state that a source was produced under a registered foreign-principal relationship. It could also identify advocacy funding, government ownership, or common control among cited domains.

Such labels would not determine truth. Government documents can be accurate, while independent sites can publish errors. The label would give users information needed to judge incentives and independence.

Platforms must also avoid partisan enforcement. Influence campaigns are not uniquely Israeli, and scrutiny should not depend on whether a platform agrees with the message.

Governments including Russia, Iran, China, and the United States have used digital communications to shape foreign audiences. Political organizations across ideological lines optimize material for search and social distribution.

The defensible standard is behavioral and transparent. Platforms should examine coordinated production, concealed sponsorship, misrepresented independence, and manipulation of ranking systems regardless of the sponsor.

Google News faces particular pressure because publishers want clear reasons for inclusion, demotion, or removal. Overly secretive enforcement can hide mistakes. Overly detailed rules can teach operators how to evade detection.

AI providers face the same tradeoff. They need enough transparency for users and researchers to audit answers, but not a complete blueprint for gaming retrieval.

The campaign also creates a challenge for journalists. Republishing a provocative claim about manipulated AI can amplify the campaign’s visibility, even when the reporting is critical.

Newsrooms should link to primary filings, identify contractual relationships, and distinguish evidence of attempted influence from proof of successful persuasion. Headlines that say a government has “trained” a chatbot risk overstating what the evidence shows.

Readers can apply the same discipline. Open the cited page, identify its publisher, check its date, and look for an author and funding disclosure. Then compare it with an official document and reporting from an unrelated organization.

For longer projects, preserving source context is more dependable than trusting a generated summary alone. A searchable knowledge base can retain documents, dates, and surrounding evidence for later review.

What AI Companies and Publishers Must Do Next

The next phase will show whether platforms can expose coordinated source networks before those networks become normal parts of AI retrieval.

The first signal to watch is citation disclosure. Google, OpenAI, Anthropic, Microsoft, and other providers should clarify when a cited publisher works for a foreign government or coordinated advocacy network.

A visible disclosure system would strengthen the conclusion that provenance has become a core AI safety issue. Continued silence would leave users responsible for investigations that platforms can perform more efficiently.

The second signal is independent measurement. Researchers should test whether campaign-linked domains appear more often for questions about Israel, Gaza, Iran, antisemitism, journalism, and United States foreign policy.

Those studies must compare models and record results over time. A few screenshots can establish that retrieval happened, but they cannot measure frequency or influence.

Consistent overrepresentation across models would strengthen claims that source optimization is affecting AI answers. Rare or short-lived appearances would suggest that the campaign achieved placement without gaining sustained influence.

The third signal is enforcement across the publishing network. Google News and other discovery services can examine whether related sites disclose ownership, produce original reporting, and follow policies against deceptive practices.

Any enforcement should include clear explanations and an appeals process. Removing sites solely because they express a political position would create serious free-expression and fairness concerns.

A more defensible response would focus on undisclosed coordination, false representations of independence, automated duplication, or manipulation that violates existing platform rules.

Publishers should prepare for higher provenance standards even if platforms act slowly. News organizations can publish ownership information, correction histories, author credentials, and machine-readable disclosures.

AI companies can then use those signals when ranking evidence. They should also reduce the weight given to clusters of pages that repeat substantially identical claims under different brands.

Answer interfaces need improvement too. Citations should connect individual claims to specific passages, not merely display several links below an entire response.

A system should distinguish facts supported by primary documents from analysis supplied by an advocacy source. When authorities disagree, the answer should name that disagreement instead of blending positions into false consensus.

Users also need an indication of source concentration. If most evidence comes from organizations sharing funding or ownership, the interface should say so.

These changes will not eliminate influence operations. Search rankings have always attracted manipulation, and political actors will continue adapting to new distribution systems.

The goal is to make coordination visible before generated prose strips away its context.

This case also offers a warning about AI browsing products. A chatbot with live search feels more accountable because it can show citations. Yet citations only improve trust when the retrieval system identifies credible, independent, and relevant evidence.

Otherwise, the interface can launder repetition into authority. A confident answer supported by several coordinated pages may look better sourced than a cautious answer based on one primary record.

That is the broader significance of the Israeli campaign. It makes explicit what many marketers, propagandists, and reputation firms have already recognized: the web is becoming a promptable database.

Publishing for that database differs from publishing for a human reader. Pages can be structured around anticipated questions, concise claims, repeated entities, and language that an AI system can easily extract.

The audience might never visit the page. The campaign succeeds if its framing survives inside an answer.

Google News remains an important battleground because it sits between publishers, search behavior, and automated discovery. However, responsibility does not rest with Google alone.

Every AI provider that retrieves from the open web makes ranking decisions. Every publisher that syndicates material contributes to the evidence environment. Every user who accepts a summary without inspecting sources reinforces the incentive to target machines rather than readers.

The verified facts do not show that Israel controls major chatbots. They show something more limited, but still consequential: a government-funded contractor openly built content and search operations intended to influence what those systems retrieve and say.

The immediate question is not whether one campaign can rewrite an entire model. It is whether AI companies can recognize coordinated persuasion before their products present it as independent evidence.

Readers should watch the citations. Then they should ask who funded the sources, how many are genuinely independent, and whether Google News or an AI summary preserved the context needed to judge them.