The investigation into the Hanover Institute revealed far more than a fake think tank that published 124 reports and more than 560,000 words in nine days. Following the money trail down to the records filed under the Foreign Agents Registration Act (FARA) with the United States Department of Justice reveals a deeper architecture: contractors linked to campaigns funded by the State of Israel to study audiences, construct narrative frameworks, produce thousands of content variants, measure shifts in perception, optimize the positioning of narratives and, in one contract, explicitly seek specific framing outcomes in conversations with GPT systems.
The question is no longer simply who manufactures propaganda. It is who is attempting to manufacture the documentary landscape that artificial intelligence will encounter when we ask it what happened.
By Claudia Aranda and LumusAI
The institute that put on a tie
Hanover did something considerably more effective than shouting lies: it put on a tie, added footnotes, and disguised itself as knowledge.
The Hanover Institute for Public Policy appeared on the Internet with all the visual attributes of a respectable think tank —a policy research center or ideas laboratory—: lengthy reports, statistics, tables, academic citations, and references to the World Bank, the United Nations, Amnesty International, and legal documents. Moreover, it dealt precisely with the issues over which the State of Israel faces a political, legal, and historical dispute of the highest order: genocide, apartheid, occupation, hunger in Gaza, the displacement of Palestinians, war crimes, antisemitism, and the political power of organizations such as AIPAC.
The investigation published by Jason Wilson in The Guardian on August 26, 2026, found, however, that behind this academic façade there was no verifiable legal entity, physical address, identified researchers, or authors signing the studies. Between August 6 and 14, the supposed institute published 124 reports totaling more than 560,000 words. On August 12 and 13 alone, it produced 73 documents, almost 354,000 words in 48 hours. Of the 124 titles, 123 began with questions phrased in a manner extraordinarily similar to what a person might type into Google or ask an artificial intelligence directly.
Spanish Revolution subsequently reconstructed the investigation for a Spanish-speaking audience in the report presented by Patricia Salvador on September 3: “Hanover Institute: how Israel is hacking ChatGPT to deceive you.” The verb “hacking” functions here as a journalistic metaphor. There is no evidence of a computer intrusion into ChatGPT or of any covert alteration of its parameters. What has been documented is different and, precisely for that reason, potentially far more significant: the deliberate intervention in the information ecosystem from which generative systems retrieve, rank, or potentially learn content.
Nor has it been established that all 124 reports were written entirely by artificial intelligence. Their volume and speed are compatible with an automated or heavily assisted process, but that inference does not amount to forensic proof.
And here an uncomfortable paradox emerges: the fact that automation cannot be proven does not make the case reassuring. If a machine was not producing material at such speed, then there must have been an extraordinarily organized human or hybrid infrastructure capable of generating half a million words in nine days on some of the most politically sensitive issues on the planet.
The dilemma is not between an alarming explanation and a reassuring one.
It is between the industrial automation of the narrative and the human industrialization of the narrative.
The factory of doubt
The most sophisticated aspect of Hanover was not outright falsehood. It was the architecture of doubt.
The Guardian found a much subtler pattern. When human rights organizations characterized Israeli actions as apartheid or genocide, certain reports repeatedly introduced the caveat that no final judicial ruling had yet been issued. In numerous documents, casualty figures or humanitarian aid data were accompanied by reservations concerning the party that had supplied them. And a 2022 academic study of antisemitic comments posted on the Facebook pages of British, French, and German media outlets was cited 454 times across 88 reports, appearing even in texts addressing subjects as different as land registration in the occupied territories, the Palestinian expulsions of 1948, the blockade of Gaza, or the death toll.
This is a form of disinformation considerably more sophisticated than simply inventing a fact.
A fact can be true and nevertheless form part of a deceptive construction if it is selected, ranked, decontextualized, or systematically associated with other elements in order to produce a particular interpretation.
This is framing: the interpretive structure that determines what appears in the foreground, what is relegated to the background, where doubt is introduced, what causal relationship is suggested, and what association is repeated until it eventually appears natural.
There is no need to write “Israel did not do this” if every appearance of “this” can instead be surrounded by enough caveats to leave the reader believing that nothing can truly be known.
Nor is it necessary to assert that every criticism of Israel constitutes antisemitism. It may be enough to introduce antisemitism again and again into discussions whose relationship to that particular study is, to put it generously, remote.
Crude propaganda demands allegiance.
Propaganda disguised as knowledge offers caution.
That is precisely its advantage.
Following the money
The Spanish Revolution investigation provides an entry point into the case. To understand its true depth, however, it is necessary to leave it behind momentarily and descend from journalistic reporting to the primary sources.
Piro, Inc. is registered under FARA as registration number 7732. FARA requires certain individuals and organizations acting in the United States on behalf of foreign interests to disclose that relationship and certain activities associated with it.
Exhibit A filed by Piro identifies the Israel Government Advertising Agency, LaPam, as the government agency represented. The names of officials Eran Shayovitch and Guy Getz also appear. The document was signed by Daniel Rosenberg under penalty of perjury on June 2, 2026. Document: U.S. Department of Justice, FARA Registration No. 7732, Exhibit A, pages 1–3 of the PDF.
Exhibit B is even clearer. It identifies the foreign principal as “Havas Media Germany GmbH on behalf of the Israel Government Advertising Agency (LaPam).” The contract is dated April 30, 2026, and Piro states that it will provide “strategic communications and media relations services in connection with LaPam’s engagement on behalf of the State of Israel.” Document: FARA No. 7732, Exhibit B, page 5 of the PDF.
Then comes a declaration that eliminates much of the room for euphemism. The form asks whether the activities will include “political activities” as defined under FARA. Piro answers yes and adds: “Activities may include communications activities intended to influence the U.S. public.” Document: FARA No. 7732, Exhibit B, page 6 of the PDF.
That is not The Guardian’s characterization.
It is not Spanish Revolution’s interpretation.
It is not an adjective of mine.
It is what the company itself declared to the United States Department of Justice.
Piro is kind enough to explain the method
Piro currently markets a service called AI Story Optimization.
Its own corporate advertising is extraordinarily useful for understanding what that means.
“The engineering gets you cited. The story makes you chosen,” the company states.
It then explains the procedure.
Piro says it maps “every surface the models read”: Reddit conversations, YouTube transcripts, reviews, media pages, forums, and comparison sites. It then ranks those sources according to how much they actually influence what AI engines say.
The next stage involves creating content “engineered for how LLMs evaluate credibility.” LLM stands for Large Language Model: the technological family to which systems such as ChatGPT, Claude, and Gemini belong.
Then comes a particularly revealing phrase. Piro says that this content can be deployed on the client’s own website or on “trusted third-party properties we build.”
Finally, the circuit is closed: “We measure, revise, and re-earn it,” referring to presence in the answers and citations generated by generative engines.
This is no longer simply advertising.
It is engineering the information environment.
For decades, SEO —Search Engine Optimization— sought to make a webpage appear among the top results on Google.
GEO —Generative Engine Optimization— pursues something different: increasing the probability that a generative system will find, use, represent, or cite particular material when constructing its answer.
The difference is political.
A traditional search engine provides links that readers can still compare. An artificial intelligence system can select the sources, synthesize them, and return a single, orderly, polished answer written in an apparently neutral voice.
Anyone who succeeds in inserting themselves into the sources the machine considers relevant is no longer competing merely for a click.
They are competing to become part of the answer.
And then Clock Tower X appears
Hanover is at the center of this story. Clock Tower X is a parallel branch and should not be mistaken for evidence that Clock Tower created Hanover.
The important connection is different: both branches appear in contracting arrangements linked to Havas and the State of Israel, and Clock Tower’s FARA file demonstrates that influencing the framing produced by GPT systems is not an interpretation journalists added afterward.
It is written into a contract.
Clock Tower X LLC appears under FARA registration number 7649. Its Exhibit A identifies Havas Media Network as the foreign principal and states in the corresponding section: “Supervised: State of Israel.” The document was signed by Bradley Parscale under penalty of perjury on September 18, 2025. Document: FARA Registration No. 7649, Exhibit A, pages 1–3 of the PDF.
In Exhibit B, Clock Tower states that it will provide “strategic communications, planning, and media services” to support Havas’s engagement by the State of Israel to develop and execute in the United States a nationwide campaign contractually presented as intended to combat antisemitism. Document: FARA No. 7649, Exhibit B, page 5 of the PDF.
The attached contract is marked “CONFIDENTIAL.” Its definition of confidential information includes “strategic plans, campaign concepts” and “narrative frameworks,” along with audience research, polling, engagement analysis, distribution tactics, and performance metrics. Document: Services Agreement, internal page 4; page 13 of the FARA PDF.
These are not categories added by a journalist to dramatize the case.
They are written into the contract.
The operations manual
The Initial Statement of Work —the contractual annex specifying exactly what the contractor is required to do— allows us to observe the machinery in operation.
The combined budget begins with $1.5 million per month for four months provided by the client, in addition to any matching contribution by Clock Tower. The base contribution contemplated in the document alone totals $6 million. Document: Initial Statement of Work, internal page 1; page 22 of the FARA PDF.
The campaign was to conduct cultural, demographic, and sentiment research; segment audiences, specifically including Generation Z; identify influential voices, communities, and distribution nodes; construct a “narrative framework and messaging blueprint”; and update messaging and framing monthly on the basis of data and audience testing.
Then the factory appears.
The contract requires the production of at least 100 core creative assets each month and 5,000 monthly variants. At least 80% were to be tailored to Generation Z audiences through TikTok, Instagram, YouTube, podcasts, and other media. Paid campaigns were to generate at least 50 million impressions per month. An impression means that a piece of content was displayed; it does not necessarily correspond to a unique person, because the same user can generate multiple impressions. Document: Initial Statement of Work, internal page 2; page 23 of the FARA PDF.
Then comes the section entitled “Search and Language Operations.”
The contract establishes monthly SEO campaigns using MarketBrew —an AI-assisted search-engine analysis and optimization platform— to improve “the visibility and ranking of relevant narratives.” It also requires a minimum of twenty pages per month targeting priority keywords and categories, along with monthly analysis of changes in the search ecosystem.
And then comes a sentence that changes the entire scale of the case:
“Deployment of websites and content to deliver GPT framing results on GPT conversations.”
Document: Clock Tower LLC & Havas Media Germany GmbH, Initial Statement of Work, internal page 2; page 23 of the FARA PDF.
It is necessary to read exactly what it says.
It does not say using GPT to produce advertising.
It does not say studying GPT.
It does not say buying advertisements around GPT.
It says deploying websites and content to achieve framing outcomes in GPT conversations.
That does not demonstrate that Clock Tower controlled ChatGPT. It does not demonstrate that OpenAI collaborated with the campaign. It does not demonstrate any modification of the internal parameters of any model. Nor does it establish that all of the contractual objectives were achieved.
It demonstrates something different and sufficiently serious: intervening through websites and content in the framing of GPT conversations was explicitly among the objectives of a campaign linked to the State of Israel.
We are no longer speculating about whether someone might want to do this.
Someone wrote down that they wanted to do it.
And budgeted the operation.
The clause that explains why “disinformation” does not simply mean lying
The same contract contains a clause that might initially appear contradictory.
It establishes that campaign materials must not deliberately include content that is “false, defamatory, or unlawful.” Document: Services Agreement, internal page 5; page 14 of the FARA PDF.
That does not contradict the problem.
It explains it.
The idea that disinformation exists only when someone invents a false fact is far too impoverished to explain the contemporary information ecosystem.
A representation of the world can be profoundly deceptive even if every brick used to construct it is technically true. Evidence can be systematically omitted, other evidence disproportionately amplified, exceptions transformed into trends, caveats introduced in only one direction, associations repeated until they become normalized, and apparently independent sources multiplied even though they ultimately depend on the same source of funding.
Reality does not have to be fabricated.
It can also be edited.
And edited on an industrial scale.
The trail does not end with the contract
Clock Tower’s FARA file does not end with that 2025 document either. The official Department of Justice registry shows a Supplemental Statement filed on May 18, 2026, new Short-Forms —abbreviated forms through which certain individuals associated with the registrant disclose their activities— and a succession of Informational Materials, meaning informational materials registered pursuant to FARA.
What is particularly significant is how the official index itself identifies them:
“State of Israel via Havas Media Germany GmbH.”
The file records materials under that identification during 2026, in addition to the supplemental statement filed in May.
This broadens the picture.
We are not looking merely at a contractual project drafted in August 2025. There is a subsequent administrative trail of registered activity during 2026.
And then, in August 2026, Hanover appears.
The second audience is no longer human
Here the nature of the case changes.
Piro explains it with almost pedagogical candor. For its product, there are two simultaneous audiences: “the engine that recommends you, and the person who believes it.”
That may be one of the most important sentences in this entire investigation.
Traditional propaganda needed to persuade people.
The new infrastructure can attempt to persuade first the intermediary that will select information for those people.
It is an upstream operation.
Instead of merely contesting the answer, it intervenes in the river before the answer is constructed.
Hanover therefore assumes an importance that goes far beyond its 124 reports. The supposed institute did not need to attract millions of human readers. It merely needed to exist as a machine-readable source, with an academic appearance, abundant citations, and a structure built around questions extraordinarily similar to those users pose to generative engines.
The old propagandist needed an audience.
The new one can try to secure intermediaries.
Common Crawl and the second door
There is also a second potential route.
Common Crawl maintains an enormous open archive of the web used in research and as raw material for various datasets. Its July 2026 crawl collected 2.14 billion pages between July 7 and 25, equivalent to approximately 364 tebibytes of uncompressed content.
The Guardian found pages belonging to sites associated with Clock Tower within that index. Hanover could not have appeared in that July collection because its reports were published later, in August.
Here the methodological boundary must be absolute.
Appearing in Common Crawl does not mean that a page was used to train ChatGPT.
Between a crawled webpage and an artificial intelligence model lie processes of selection, filtering, cleaning, deduplication, and dataset construction that vary from one developer to another.
There is therefore no evidence to claim that Hanover “contaminated ChatGPT’s training.”
But that does not eliminate the risk.
It means that this particular door cannot be assumed to have been crossed.
The other one already exists: producing content intended to increase the likelihood that generative systems will find, represent, or cite it.
And there is a decisive difference between these two routes. When a system consults a webpage in real time and displays its source, we can still follow the citation, determine who published the content, discover who financed it, and reject it.
When an association has been statistically incorporated during a training process, that traceability may disappear for the user.
Propaganda then ceases to present itself as a source.
It can become a probability.
From source laundering to epistemic capture
From this point onward, the prospective analysis begins. And it is essential to distinguish scrupulously between what has already been documented and what constitutes a possible scenario.
The most conservative scenario already exists: political and commercial actors study which sources generative engines use, produce content specifically designed to appear readable and credible to those systems, place it where machines search, and subsequently measure whether it is cited. Piro describes precisely this circuit as a commercial service.
The immediate risk is citation capture.
If certain states, corporations, lobbies, or billionaires possess the resources to produce vastly greater quantities of content, they can attempt to artificially increase their presence within the documentary universe that these systems consult.
Economic inequality can thus become epistemic inequality.
The second scenario is the manufacture of artificial corroboration.
Suppose a single actor finances twenty apparently independent websites. Each publishes a slightly different version of the same argument. Other domains then cite them, aggregators reproduce them, and social networks link to them.
A machine might encounter twenty URLs.
But behind them there would be only one voice.
This is what we can call provenance laundering: the political or financial origin of a message becomes progressively less visible as it passes through domains, summaries, aggregators, search engines, and generative engines.
For journalism, this means revisiting an old rule.
It is no longer enough to ask how many sources confirm something.
We must ask how many actual owners, funders, and centers of production exist behind those sources.
Twenty URLs can be one voice wearing twenty disguises.
The severe scenario: colonizing the corpora
The next scenario is more serious and has not yet been demonstrated in Hanover’s case.
If producing millions of pages becomes extraordinarily cheap, an operator could design content specifically intended to maximize its survival within web crawls and repositories that might later feed training datasets.
Large language models do not need to memorize a textual slogan in order to be affected by an information distribution. They learn statistical regularities.
If enormous quantities of content systematically associated a particular form of political criticism with antisemitism, or repeatedly subjected Palestinian sources to a standard of suspicion different from the one applied to Israeli sources, the concern would not be finding a secret sentence hidden inside the model.
It would be a statistical gradient.
We have no evidence that Hanover succeeded in doing this.
We do have evidence that the technical infrastructure that would make such an attempt possible exists.
And we have documentary evidence that another contractor within the ecosystem examined included among its tasks the deployment of websites and content to obtain specific framing outcomes in GPT conversations.
The audacious scenario: training the environment
There is an even more sophisticated scenario. An operation could systematically query different models, identify which questions produce unfavorable answers, identify the sources being used, produce new content specifically designed for those queries, place it on domains that the engines regard as credible, and then run the questions again.
Question. Answer. Measurement. Production. Publication. New question. New measurement.
Piro already commercially describes much of that circuit: mapping the surfaces models read, identifying relevant queries, producing content designed for their credibility criteria, deploying it across digital properties, and continuously running the measurement-and-revision cycle.
Automated and applied politically on a large scale, this mechanism could become an adversarial feedback loop.
It would not be hacking in the traditional computer-security sense.
It would be training the environment.
The extreme scenario: manufacturing consensus
The most troubling frontier would be epistemic capture.
States, corporations, and great fortunes could systematically compete to colonize the corpus available to artificial intelligence systems until what a machine perceives as “consensus” reflects, at least in part, who had the greatest economic capacity to produce and distribute sources.
This scenario has not been demonstrated.
But its technological components already exist separately: industrial-scale content generation, networks of websites, programmatic advertising, GEO, automated monitoring of responses, construction of third-party digital properties, massive Internet crawls, and artificial intelligence systems that synthesize enormous volumes of information for millions of people.
What is missing is not technology.
It is scale, persistence, and coordination.
And scale, persistence, and coordination are precisely the properties beginning to emerge in the documents examined.
The criminal disinformation enclave
That is why I characterize Hanover and the architecture surrounding it as a criminal disinformation enclave.
“Criminal” does not mean here that FARA has declared the operation criminal. It has not. Nor does it mean that there is a U.S. criminal judgment establishing that Piro, Clock Tower, Havas, or Hanover constituted a criminal organization.
It is my political and moral characterization of a machinery that, according to my interpretation of the documented evidence, deliberately interferes with the collective right to receive information about a war, an occupation, and allegations involving some of the gravest crimes recognized under international law.
Separating the two does not weaken the judgment.
It makes it intellectually clean.
The documentary evidence establishes who contracted whom, what activities were disclosed, what objectives appear in writing, and what mechanisms were budgeted.
The editorial judgment is mine.
Gaza and the battle for the archive
This apparatus is not working on an innocuous commercial product.
It is working on Gaza.
And that matters.
The International Court of Justice has not yet issued a final judgment on the merits of the case brought by South Africa against Israel under the Convention on the Prevention and Punishment of the Crime of Genocide.
But it is equally incorrect to reduce what happened before the Court to the statement that “there is no genocide judgment” and end the story there.
On January 26, 2024, the Court determined that “at least some of the rights claimed by South Africa […] are plausible.” It specified which rights: “the right of the Palestinians in Gaza to be protected from acts of genocide and related prohibited acts.” Document: International Court of Justice, Order of January 26, 2024, paragraphs 54 and 59.
In paragraph 66, it added that those plausible rights are of such a nature that their violation may cause irreparable harm. The Court ordered binding provisional measures intended to protect them.
And on May 24, 2024, it expressly returned to that determination. In paragraph 32, it recalled its January conclusion and stated that it saw no reason to revise it.
This is legally far more precise than claiming that the ICJ “declared genocide plausible,” which the Court did not say.
It also demonstrates why framing matters.
It is perfectly true to write that there is not yet a final judgment on genocide.
It is also perfectly true to write everything that comes after that sentence.
Systematically choosing the first truth while omitting the second does not necessarily make a sentence false.
It can make the narrative deceptive.
Genocidaires also contest memory
I characterize the State of Israel as a terrorist state. This is my political and moral judgment; I do not seek to attribute it to the International Court of Justice or fraudulently transform it into a judicial category that those institutions have not established.
And I believe this campaign must be understood within a much larger struggle to preserve impunity and contest the memory of crimes.
Genocidaires need to dominate more than territory.
They need to contest the archive.
They need certainty, twenty years from now, to be capable of appearing controversial again; the victim to reappear as a dubious source; and the perpetrator to become merely “one of the parties” to an episode whose causality grows increasingly difficult to reconstruct.
Artificial intelligence introduces a possibility that has never existed on this scale: placing a machine between the citizen and the archive.
A generation that once might have consulted five newspapers, two documents, a book, and an encyclopedia can now ask a system a question and receive ten paragraphs of synthesis.
The source disappears behind the narrator.
Whoever succeeds in influencing the narrator’s materials acquires a new form of political power.
That is the real leap Hanover allows us to observe.
Hanover is not merely a fake website.
It is a prototype.
The right to be informed enters a new era
Defending the right to information can no longer be limited to combating fake news.
We need to talk about the integrity of the information ecosystem.
An artificial intelligence system should be able to distinguish ten genuinely independent sources from ten domains financed by the same actor. It should retain provenance information —who funded it, who published it, who controls it, who registered it— and use that information as a relevant signal when evaluating political material.
FARA records and equivalent systems may become fundamental tools for this work.
Retrieval systems will also need to be audited: which sources they privilege, how much weight they assign to repetition, how they identify coordinated content networks, and to what extent they can be manipulated through generative engine optimization.
And journalism will have to investigate more than texts.
It will have to investigate infrastructures.
Who registered the domain, who pays for the advertising, who shares servers, who signed the contract, who appears in FARA, who built the website, who cites whom, who produces half a million words in nine days, and who expects a machine to read them.
Media literacy will also have to become literacy in algorithmic provenance.
It will no longer be enough to ask:
“Is this true?”
We will have to add:
“Why did this machine learn this?”
“Which sources did it obtain this from?”
“Who paid for those sources to exist?”
The final irony
There is something almost grotesque about all of this.
While campaigns intended to alter perceptions of Israel were being developed, its public image in the United States continued to deteriorate, particularly among younger sectors —precisely one of the audiences explicitly identified as a priority target in the contractual documents examined. The Guardian placed the operation within this reputational deterioration and amid doubts even about the effectiveness of such communications investments.
Perhaps there is an explanation considerably less sophisticated than a marketing problem.
After years of deaths, destruction, displacement, hunger, allegations of war crimes, and proceedings before international courts, it is possible that part of the reputational decline is not the result of Israel having poorly explained what it is doing.
It may be the result of people having seen what it is doing.
And then comes the ultimate temptation of the propaganda apparatus:
if you cannot change what produces rejection, change the narrative about what produces rejection.
Hanover belongs to that moment.
Clock Tower takes the logic even further, because its contract explicitly introduces GPT conversations as a territory for intervention.
That is the new frontier.
For a century, propaganda sought to control what people read.
Then it sought to control what search engines and platforms showed them.
Now it is attempting to influence what machines will answer them.
The old power wanted to write history.
The new one also wants to intervene in the automated historian.
It does not end with Israel
This is perhaps the most serious conclusion.
Nothing described here belongs technologically to Israel alone.
Tomorrow, any state, corporation, political party, intelligence service, lobby, sect, or billionaire with enough money to purchase infrastructure, produce content, and learn which signals machines use to decide what deserves to be read could do the same.
If Hanover were an anomaly, it would be an extraordinary story.
If it is a precursor, we are facing something much larger.
Automated production drastically reduces the cost of manufacturing documents. Generative optimization makes it possible to design them with machines in mind. Web crawlers can incorporate them into enormous archives. Generative engines can find them. Retrieval systems can cite them. And citizens can receive the final product without ever having seen the production chain.
Old propaganda had to reach millions of people.
The new propaganda can try to reach a few machines first.
And let them do the rest.
That is why this investigation does not end with Hanover.
It begins there.
It is no longer enough to defend truth against censorship, lies, and silence.
It must also be defended against saturation.
Because in the age of artificial intelligence, perhaps the most efficient way to hide a fact is no longer to erase it.
It may be enough to manufacture a million documents around it.
References
Common Crawl Foundation. (2026, July 28). July 2026 Crawl Archive Now Available. Primary source on the July archive: 2.14 billion pages collected between July 7 and 25, 2026.
International Court of Justice. (2024, January 26). Application of the Convention on the Prevention and Punishment of the Crime of Genocide in the Gaza Strip (South Africa v. Israel), Order of 26 January 2024. See especially paras. 54, 59, and 66.
International Court of Justice. (2024, May 24). Application of the Convention on the Prevention and Punishment of the Crime of Genocide in the Gaza Strip (South Africa v. Israel), Order of 24 May 2024. See especially para. 32.
Piro, Inc. (2026). AI Story Optimization. Primary corporate source describing the commercial system through which Piro identifies sources consulted by generative engines, creates content adapted to the credibility criteria of large language models, deploys it on its own websites or third-party properties, and subsequently measures its appearance in responses.
Spanish Revolution. (2026, September 3). Hanover Institute: how Israel is hacking ChatGPT to deceive you. Reportajes SR. Presented by Patricia Salvador. Secondary audiovisual investigation that constitutes the starting point for this investigation. (https://youtu.be/p3qv9R0Jol4?utm_source=chatgpt.com)
U.S. Department of Justice, National Security Division. (2025, September 18). Clock Tower X LLC, Foreign Agents Registration Act Registration No. 7649, Exhibit A, Exhibit B, Services Agreement and Initial Statement of Work. Primary source. Contains the disclosed relationship with Havas and the State of Israel, audience research, narrative frameworks, production of 5,000 monthly variants, 50 million impressions, SEO operations, and the clause “Deployment of websites and content to deliver GPT framing results on GPT conversations.” (https://efile.fara.gov/docs/7649-Exhibit-AB-20250918-1.pdf?utm_source=chatgpt.com)
U.S. Department of Justice, National Security Division. (2026, May 18, and subsequent filings). Clock Tower X LLC, Registration No. 7649, Supplemental Statement and Informational Materials. The official index records materials under the identification “State of Israel via Havas Media Germany GmbH.” (https://efile.fara.gov/ords/fara/f?p=1381%3A200%3A%3A%3ANO%3ARP%2C200%3AP200_REG_NUMBER%3A7649&utm_source=chatgpt.com)
U.S. Department of Justice, National Security Division. (2026, June 2). Piro, Inc., Foreign Agents Registration Act Registration No. 7732, Exhibit A and Exhibit B. Primary source. Identifies Havas Media Germany GmbH acting on behalf of the Israel Government Advertising Agency (LaPam), records the engagement related to the State of Israel, and discloses communications activities that may be intended to influence the U.S. public. (https://efile.fara.gov/docs/7732-Exhibit-AB-20260602-0.pdf?utm_source=chatgpt.com)
Wilson, J. (2026, August 26). Fake US thinktank set up and funded by Israel sought to game AI for propaganda. The Guardian. Secondary investigation documenting Hanover, its 124 reports, the more than 560,000 words published in nine days, its framing patterns, and its infrastructure oriented toward generative engines. (https://www.theguardian.com/world/2026/aug/26/fake-thinktank-israel-ai-propaganda?utm_source=chatgpt.com)
