In nine days, 124 reports and more than 560,000 words appeared.

This was neither a digitised library nor the release of years of research. The documents came from the Hanover Institute for Public Policy, an organisation with the appearance of an American public-policy think tank which presents itself as dedicated to analysing the factors that fuel antisemitism in the United States. Its pages adopt the recognisable form of contemporary research: methodology, data, charts, technical language, academic references and conclusions expressed with apparent detachment.

An investigation published by The Guardian on 26 August placed these materials within a campaign intended to influence the information that artificial-intelligence systems find, retrieve and eventually incorporate into their answers about Israel and Palestine. The newspaper documented the publication of more than half a million words in just nine days and described the Hanover Institute as a structure that does not exist as an identifiable legal entity and has no publicly identifiable staff or institutional address.

The organisation itself now acknowledges a financial relationship that makes it possible to reconstruct part of this chain. According to its funding page and the Foreign Agents Registration Act filings cited by The Guardian and Responsible Statecraft, its materials are distributed by Piro Inc. on behalf of Havas Media Germany, acting for the Israel Government Advertising Agency, LaPam. Hanover maintains, at the same time, that the funding body neither selects topics nor reviews drafts or approves conclusions before publication.

This analysis also begins from a memory that must not be instrumentalised. We remember especially the victims of the Shoah and, alongside them, the many Spanish Republicans deported to Nazi concentration camps, many of whom were murdered in the Mauthausen–Gusen complex. The shared memory of those who suffered antisemitism, fascism, deportation and extermination obliges us to oppose every form of hatred directed at Jewish people and communities. It also prevents that duty of remembrance from being turned into a means of shutting down critical scrutiny of a state’s policies or its communication campaigns.

The question of the actual independence of these reports matters. But another issue extends far beyond this particular case.

What happens when a campaign no longer seeks only to persuade a reader, but to become one of the sources from which a machine will later construct an answer?

Before the answer comes the source

For decades, much political communication could be imagined as a relatively simple sequence.

Someone produced a message. A media outlet distributed it. An audience received it.

The internet added intermediaries. Search engines began deciding which documents appeared first, and an entire industry emerged around optimising content for Google’s criteria. SEO did not necessarily need to change what a person thought. It was enough to increase the likelihood that a particular page would be found.

Generative systems introduce another layer of mediation.

A query no longer necessarily leads to a list of documents. It may end in a seemingly self-contained paragraph that brings together, summarises and rearranges information from several sources.

The sequence changes:

actor → document → AI system → answer → public

And with it, the point at which intervention can be effective also changes.

There is already a name for part of this practice: Generative Engine Optimisation, or GEO. Research on GEO examines how particular features of a page affect the likelihood that it will be retrieved, cited or incorporated into answers generated by systems such as ChatGPT, Gemini or Perplexity.

It is not simply SEO under another name. Recent research distinguishes between a source being located, being cited and having its contents actually absorbed into an answer. A long, well-structured page that closely matches the query and contains data that can be extracted easily may have more opportunities to enter that chain.

This does not mean that there is an infallible formula for controlling what a model will say. The available literature emphasises variation between platforms, queries and individual runs. Nor does it justify claiming that publishing thousands of pages will automatically alter a model’s training or its live retrieval window.

But it opens a different political possibility.

There is no need to change the model if it is possible to influence the documentary landscape the model consults.

Looking like knowledge

This is where the Hanover case becomes particularly revealing.

Because what is being produced does not have the obvious form of an advertisement.

It has the form of a report.

There are tables. Methods. Citations. Samples. Explanations. Quantitative language. Precisely formulated questions. An institution with a name that could belong to any number of research centres.

Almost all the headlines are written as questions resembling those someone might type into a chatbot: ‘Is Anti-Zionism Antisemitism?’, ‘What Caused the Displacement of Palestinians in 1948?’ or ‘Which Humanitarian Organisations Have Documented Israeli War Crimes?’. The document does not simply wait to be read. It already presents itself in the form of a query and its possible answer.

Nor is the interrogative form a neutral container. The Palestinian-American anthropologist Sa’ed Atshan has shown how an apparently descriptive word such as ‘crisis’ can obscure occupation, normalise oppression and erase the historical structure that makes an event intelligible. Something similar happens with a question: by defining the problem, it has already decided which period matters, which actors appear, which causal relationships seem plausible and which evidence will count as relevant.

Optimising documents for chatbot questions may therefore mean more than competing to supply an answer. It also makes it possible to contest the premises from which the machine will organise the issue. The choice of title, the definition of categories and the timeframe are all part of the intervention. Before it answers, the system may already have accepted the framing proposed by the source.

All these features have a cultural history that predates artificial intelligence. People, too, have learnt to use them as signals of authority.

A table does not prove that something is true.

A written methodology does not guarantee that the method is sound.

An academic reference does not, by itself, turn an interpretation into reliable knowledge.

And yet these forms are constantly involved in our everyday procedures for deciding what deserves trust.

Artificial intelligence inherits this landscape.

Its search and retrieval systems must decide which of millions of pages appear relevant to a particular question. They must then select passages, compare them and use them to produce an answer.

The problem, therefore, cannot be reduced to whether an AI can detect ‘propaganda’.

The prior question is more uncomfortable:

how does a machine recognise authority?

And another follows:

how many of the cues it uses to recognise authority were originally constructed by our own documentary culture?

The potential effectiveness of an operation like this lies precisely there. It does not need to invent a new form of credibility. It can appropriate one we already know.

Piro, the company distributing Hanover’s materials, advertises this activity as AI Story Optimisation. Its own offer promises to produce content ‘engineered for how LLMs evaluate credibility’. The phrase turns the appearance of authority before a machine into an explicit commercial product.

No source exists without conditions

The Society of Palestinian Anthropologists, Insaniyyat, offers an especially useful criterion for reading this case. Its ethical guidelines maintain that producing knowledge about Palestine cannot be separated from the colonial and political conditions in which that knowledge is made. Against the illusion of abstract objectivity, they ask us to examine the relationships between researchers, the institutions that host them, funding bodies, participants and the circuits through which findings will be disseminated.

The proposal is not to replace rigour with the identity of the speaker. It is to expand what we understand by rigour. A source cannot be evaluated solely by the formal correctness of its tables or the number of references it contains. The provenance of the data, the conditions of consent, the position of those who contributed knowledge, mechanisms of accountability and the possibility of reconstructing the path from research to publication also matter.

Seen from this perspective, Hanover’s financial disclosure resolves only one part of the problem. It allows us to identify a chain of funding and distribution, but it does not identify authors, research teams, selection criteria, verification procedures or specific institutional responsibilities. Financial transparency can exist without epistemic traceability.

This distinction is crucial for generative systems. A machine may retrieve the methodology written on a page, but not necessarily the social relationships absent from it. It can extract a figure without knowing who was able to produce it, challenge it or refuse to take part. When authority is calculated from the visible properties of a document, the concealed conditions of its production risk disappearing twice: first from the source and then from the answer.

From SEO to the struggle over the answer

The search engine made the first page of results politically valuable.

Generative AI may make something else politically valuable: becoming part of the set of documents from which an answer is synthesised.

The difference seems small, but it is not.

In a traditional search engine, we can still see, at least in principle, a plurality of sources. We can open one, go back, compare another, recognise a publication or distrust a domain.

In a generative answer, those boundaries can dissolve.

Five documents can become five sentences.

A controversy can become an explanation.

A minority source may disappear, while a carefully optimised source may acquire disproportionate presence.

Authority is no longer expressed only as a position in a ranking. It begins to appear embedded in the very language of the answer.

This turns the documentary infrastructure of the internet into a new terrain of contestation.

And not only for states.

Companies, lobbying groups, political parties, industries, social movements and organisations of every kind have the same potential incentive. If millions of people begin their enquiries at a generative interface, appearing among the sources that feed that interface acquires economic and political value.

The production of content for machines may thus become an activity in its own right.

Not writing so that someone will read.

Writing to be retrieved.

Not persuading directly.

Becoming evidence available to whoever will answer.

What a synthesis can erase

But the problem does not end with which sources get in. What happens to them when they are synthesised also matters.

The Palestinian anthropologist Amahl Bishara has stressed that Palestinian experience is traversed by different forms of settler-colonial violence according to legal status and geography: Israeli citizenship, residency in Jerusalem, military occupation in the West Bank or Gaza, refuge and exile. Speaking of ‘the Palestinian situation’ as a homogeneous unit may conceal precisely those differences and the relations that produce them.

A generative answer can bring together diverse documents and turn them into a fluent paragraph. That fluency is useful, but it is not the same as establishing a just relationship between sources. It may erase the provenance of a claim, place testimonies produced under radically unequal conditions on the same footing, or compress distinct histories into an explanation without friction. Synthesis is not merely an operation of reduction: it also distributes visibility.

The doctoral research of the Palestinian-American anthropologist Hadeel Assali takes us a step further. Her dissertation examines how narratives do not merely describe spaces but help to produce them and confer authority upon them. The collective volume Producing Palestine, edited by Dina Matar and Helga Tawil-Souri, extends that idea to media and infrastructures such as TikTok, digital mapping and drone imagery: Palestine is produced and reproduced through technologies, languages and geographies, not merely represented within them.

Hanover participates in this contest from an industrial position funded by a state agency. It is not merely attempting to add a favourable interpretation to the internet. It is attempting to help manufacture the Palestine made documentarily available to machines: which questions define it, which data appear extractable and which voices arrive already transformed into evidence. Asking about the answer therefore also requires asking which experiences were excluded from the corpus, which differences were compressed and who had the means to fill the archive at speed.

An old practice, a new infrastructure

There is something profoundly familiar in all this.

Archives, censuses, encyclopaedias, libraries, statistics and maps have never been neutral repositories of information. Decisions about what is recorded, under which categories, by what name and with whose authority have always participated in producing social reality.

Generative systems add a new infrastructure to that history.

An infrastructure capable of traversing immense archives and returning a compressed version of them in seconds.

This may be why public discussion about the manipulation of artificial intelligence is looking too late in the chain.

We ask what the model will answer.

We ask what biases it learnt.

We ask who wrote the prompt.

But before any of that, there is a world of documents that someone had to produce, classify, publish, link and make visible.

The Hanover case matters less because it proves that one particular campaign managed to control major models — something the available evidence does not establish — than because it shows with extraordinary clarity what the next frontier may be.

If automated answers acquire social authority, appearing among the sources for those answers acquires power.

The struggle over artificial intelligence therefore begins long before we type a question.

It begins when someone decides what will look like a source.

Sources