For OSINT practitioners, journalists and researchers, that distinction matters.
Documents filed in the United States under the Foreign Agents Registration Act (FARA) and reported by Politico describe an Israeli-linked communications campaign that appears designed not only to influence conventional audiences, but also to increase the visibility of specific narratives within the information ecosystem accessed by AI systems.
The case provides an unusually useful example of an emerging OSINT problem: source-layer influence.
Instead of attacking or controlling an AI model, an actor can attempt to shape the public information environment surrounding it.
And that environment is something we can investigate.
According to the material reviewed, the campaign involved Havas Media, advertising company Piro Inc., and the Hanover Institute for Public Policy, a website publishing material on Gaza, antisemitism, anti-Zionism, Palestinian prisoners and the Israeli-Palestinian conflict.
FARA filings cited in the reporting describe more than a dozen Hanover Institute articles as part of a $100,000 campaign intended to create and distribute publicly available material about Israel and related issues.
The amount itself is not particularly significant compared with major international influence campaigns.
The methodology is.
Many Hanover Institute pages are structured around questions that closely resemble prompts submitted to search engines and AI assistants. Examples contained in the source material include questions about whether the IDF is “the most moral army in the world” and whether there is a starvation policy in Gaza.
This structure creates content that can potentially match the information needs of both human searchers and automated retrieval systems.
That changes the OSINT question.
Instead of asking only:
Who is trying to influence the audience?
we increasingly need to ask:
Who is trying to influence the sources used by the systems that inform the audience?
This distinction is essential.
The source material reports that Politico tested ChatGPT and Perplexity with neutral queries and found that both systems cited Hanover Institute material in responses concerning Gaza, anti-Zionism and antisemitism.
That is evidence that the material was retrievable or citable in those tests.
It is not evidence that Israel controls either platform.
More importantly, it does not establish that the Hanover Institute material was included in the underlying training datasets of the models.
Those are separate technical questions.
Modern AI systems may retrieve information from the live or indexed web when generating an answer. A page appearing as a citation therefore does not demonstrate that its contents altered the model itself.
For OSINT analysis, we should classify the claims accordingly:
Evidence: documents connect the reported campaign, its actors and the production of public-facing content.
Evidence: according to the reporting described in the source material, Hanover Institute pages appeared as citations in tests conducted with ChatGPT and Perplexity.
Indication: the structure and distribution of the content are compatible with an attempt to improve its discoverability by generative systems.
Not established: that the campaign changed the underlying models, their training datasets or their overall treatment of Israel and Gaza.
That last distinction prevents an investigation from becoming a conspiracy theory.
Influence operations targeting AI create a familiar investigative problem in a new environment.
The visible article is only the first layer.
An OSINT investigation should move backwards through the information chain:
AI answer → cited source → publisher → author → domain → organization → contractor → client → funding → disclosure records.
In the United States, FARA provides an especially valuable starting point because organizations conducting certain political or public-relations activities on behalf of foreign principals are required to disclose information about those relationships.
This means that an apparently independent source can sometimes be connected to a broader communications structure through public records.
That is precisely where OSINT adds value.
The objective is not to decide whether an article is “propaganda” by reading its headline.
The objective is to establish provenance.
When an AI system provides sources on a politically sensitive issue, save the answer before doing anything else. Record the prompt, date, model or service used, exact response and cited URLs. AI answers are dynamic; repeating the same query later may produce different sources.
Then investigate the citations.
Open the cited page and record:
publication name;
URL and domain;
publication date;
named author, if any;
editorial or About page;
sources referenced in the article.
In the Hanover case, the absence of named authors on the described reports is itself an element worth recording, but it is not proof of improper activity.
Move beyond the article.
Check the organization behind the website, domain history where relevant, corporate records, disclosed partnerships and publicly available information about funding or commissioning.
The question is not simply Who wrote this?
It is:
Who made this information exist?
For US-based foreign influence investigations, search relevant FARA filings.
Look for:
foreign principal;
registered agent;
subcontractors;
payments;
statements of work;
campaign descriptions;
deliverables;
dates.
A disclosure document can provide stronger evidence about relationships than dozens of speculative social-media posts.
Test the same neutral question across multiple systems.
Do not ask leading questions such as:
“Why is this website Israeli propaganda?”
Use neutral formulations and document which sources each system retrieves.
Repeat the test in a clean session where possible.
The goal is not to prove bias from one answer. It is to observe source selection.
Search the same claim independently.
Compare:
primary documents;
institutional sources;
established reporting;
academic or specialist research;
sources representing competing interpretations.
Then identify which statements are independently supported and which depend on a single information network.
This is particularly important because ten websites repeating the same originating claim do not constitute ten independent confirmations.
A common approach to AI verification is to fact-check the final output.
That is no longer enough.
Suppose an assistant produces a perfectly plausible paragraph and provides four citations. If two of those sources originate from the same communications operation, the answer may look better corroborated than it actually is.
This creates a source independence problem.
The number of citations is therefore less important than their provenance.
OSINT analysts already understand this principle when investigating news reports, corporate networks or coordinated social-media campaigns. The same methodology now needs to be applied to generative search.
The Israel case should not be treated as evidence of a uniquely Israeli strategy.
It should be treated as a documented case that exposes a broader vulnerability.
Governments, corporations, lobbying organizations, political campaigns and advocacy groups all have incentives to influence what appears online.
Generative AI adds another audience to that competition: the retrieval systems sitting between information producers and human users.
Traditional SEO attempts to make content visible to search engines.
The emerging field often described as Generative Engine Optimization (GEO) attempts to make information more likely to appear in generative search and AI-mediated answers.
The technique itself is not inherently deceptive. Publishers legitimately want AI systems to find accurate information.
The OSINT problem begins when optimization, attribution and influence intersect.
A polished research-looking website may be a genuine independent institution.
It may also be part of a communications campaign.
The interface does not tell us which.
Investigation does.
Generative systems compress information.
That is useful, but compression can remove provenance.
Different sources, different interests and different levels of reliability are transformed into paragraphs written in the same confident voice.
For investigators, this means that the AI response should never be treated as the end of the research process.
It should be treated as another lead.
The workflow becomes:
Prompt → answer → citations → provenance → relationships → independent verification → assessment.
The expected result is not simply determining whether an AI answer is “true” or “false”.
It is understanding how that answer became possible.
That distinction will become increasingly important as political influence campaigns adapt to generative search.
In traditional OSINT we often ask:
Who published this information?
In AI-assisted OSINT, we need to add another question:
Who had an interest in making sure the machine could find it?
That is where the investigation starts.
Methodological note
The available material supports the existence and reported structure of the campaign, the relationships described in the cited FARA reporting, and Politico‘s reported observation that Hanover Institute material appeared in ChatGPT and Perplexity test responses.
It does not establish that these materials altered the underlying models, entered their training datasets, or systematically changed their responses. Those claims would require additional evidence.
If this is useful, share it.
This is the weekly selection. But it’s not the only one.
If you’d like to read more: → full articles on the website
If you’d like to get the latest updates first: → Telegram
Comments
Nothing yet. Say the first thing.
Sign in to join the conversation.