I have never been impressed by hasbara, though not for the reason usually given. Israel is subjected to propaganda, some of it crude and some of it highly effective, and no government involved in a prolonged war could reasonably be expected to leave the field uncontested. Hamas has its own communications apparatus, sympathetic governments amplify its case, and much of the Western press approaches Israel with assumptions that are far from neutral. Nor does the fact that a statement comes from the Israeli Government make it false. Governments sometimes tell the truth, just as NGOs and international organisations sometimes publish material that deserves to be challenged. My difficulty with hasbara lies elsewhere. It is the recurrent belief that hostility towards Israeli policy is fundamentally a problem of presentation. If the public remains unconvinced, the explanation must have been insufficiently clear. The solution is therefore another campaign, another consultant, another improved narrative.

The latest campaign takes this assumption into territory that is genuinely new. Israeli government money has been used to finance material designed not simply to persuade human readers, but to influence the sources consulted by large language models. This is an interesting development, though I suspect its practical importance has already been exaggerated. Before considering whether ChatGPT can be nudged towards more favourable answers about Israel, it is worth asking which Americans Israel most needs to persuade. The old evangelical constituency hardly requires assistance from artificial intelligence. Its opinions are still formed largely by obese preachers waving Scofield Bibles, supplemented by whatever the grinning, anorexic old women on Fox News happen to say that evening. This audience settled its view of Israel decades before anybody had heard of generative AI. The people among whom Israeli support has weakened are younger, less religious, generally more suspicious of the traditional foreign-policy establishment. They are also more likely to ask ChatGPT questions instead of reading newspapers. That gives the campaign a certain logic, but it also exposes its central weakness.

The most conspicuous part of the operation appeared in early August. A website calling itself the Hanover Institute for Public Policy began publishing reports at a pace that would have been remarkable for an institution ten times its apparent size. Between 6 and 14 August it produced 124 reports amounting to more than 560,000 words. Much of this output appeared within two days. The reports had the visual appearance of conventional American policy research. They contained footnotes, references to institutional sources and long discussions of disputed questions. What they did not contain were named authors or any clear indication that there was a normal research institute behind the name. Their titles were revealing. Instead of resembling the titles of academic articles, they resembled questions that ordinary users might type directly into ChatGPT: whether Israel is committing genocide in Gaza, whether there is a deliberate policy of starvation, whether anti-Zionism is antisemitism, whether Palestinians were expelled in 1948.

The answer to what Hanover actually is can be found in the disclosure material. The content is distributed by the New York firm Piro Inc. on behalf of Havas Media Germany, which is acting for LaPam, the Israeli Government Advertising Agency. Piro registered the arrangement with the American Department of Justice under the Foreign Agents Registration Act. It is therefore difficult to describe Hanover as a think tank in the ordinary sense. It is a state-funded communications project presented in the form of a policy institute. That distinction should matter, although it does not establish that the material itself is false. A funded argument may still be correct. The point is that the reader is entitled to know who is making the argument and why it has been placed before him.

This matters when the intended intermediary is not really a reader at all. Piro advertises what it calls “AI Story Optimization”, which is just the latest extension of an old commercial practice. For years companies adjusted their websites so that Google would notice them. Search-engine optimisation did not normally involve corrupting Google. It involved learning what kinds of pages Google tended to rank highly and then producing those pages. Generative AI has created an equivalent market. If users increasingly ask ChatGPT or Perplexity questions rather than searching the Web themselves, it becomes worthwhile to know what sorts of documents those systems are likely to retrieve and cite.

Hanover’s structure makes sense once this is understood. The reports are long because long documents look substantial. They are footnoted because footnotes suggest authority. Their headings are framed around common questions because those are the questions the machines are being asked. The site was created on a commercial platform that promotes citation by chatbots and initially included an llms.txt file intended to guide machine access. None of this is sinister in any technical sense. It is simply propaganda adapted to a new information market.

There are two ways this may matter, and they should not be confused. The first involves retrieval. A chatbot searches the current Web, finds one of these pages and incorporates its contents into an answer. A newly created site can influence such an answer immediately if the retrieval system happens to rank it highly. The second involves training. Material is collected by crawlers and may later be used to train or update a model. That process is slower and less predictable. Hanover’s August production burst came too late for the July Common Crawl index, so claims that the site has already poisoned the underlying knowledge of every major chatbot are overstated.

The wider Israeli communications operation is more interesting in this respect. Havas has also worked with Clock Tower X, the firm run by Brad Parscale, Donald Trump’s former campaign manager. Foreign-agent filings described part of its task as the “deployment of websites and content to deliver GPT framing results on GPT conversations”. Some of the sites associated with that network later appeared in Common Crawl. This creates at least the possibility that material created for political purposes may eventually enter datasets used in future AI systems. But even here I would resist the increasingly fashionable expression “LLM poisoning”. That suggests the corruption of the model itself. What we are really seeing is source-seeding: placing large amounts of favourable material into the public information environment in the hope that machines will encounter it.

The objective becomes clearer if one imagines the answer that the Israeli Government would prefer. If a user asks whether Israel is deliberately starving Gaza and the chatbot replies that “the Israeli Government denies this allegation”, the provenance of the claim remains obvious. The user can agree or disagree with it while knowing who has supplied the argument. A more valuable result would be an answer in which the same reasoning appears without its institutional label: that available evidence does not establish deliberate starvation, that aid distribution is contested, that some of the underlying statistics require caution. Each of these propositions may deserve serious consideration. The communications advantage lies in their appearing as the neutral synthesis of a machine rather than as the brief of one of the combatants.

This is what makes the operation worth noticing. It does not show that artificial intelligence has somehow been captured by Israel. It shows that governments have begun to understand how much authority users now grant to machine-generated summaries. A search engine traditionally showed its sources. Even a lazy reader could see whether he was clicking on The Guardian or the Israeli Foreign Ministry. A chatbot smooths the differences into a single answer. Once that happens, the identity of the original speaker becomes easier to lose.

How effective this will be is another matter. Journalists testing Hanover material have obtained mixed results. ChatGPT and Perplexity have sometimes cited Hanover reports. On other occasions they have ignored them, while some answers have mentioned the site together with warnings about its funding. That is about what I would expect. Five hundred and sixty thousand words is an enormous amount of material for a public-relations operation to produce in a week. Set against the scale of the public Web, it is very little.

There is another reason to expect only limited success. The information environment surrounding Israel and Palestine is already saturated. The Israeli Government publishes material continually, as does the IDF. Palestinian authorities publish their own claims, while Hamas has never shown any reluctance to conduct propaganda. UN agencies issue reports, as do human-rights organisations. Newspapers and broadcasters produce further layers of commentary, much of it openly partisan and some of it disguised as neutral analysis. Israel is therefore not entering an empty field. It is adding another organised source to an environment in which almost every participant already understands that information is part of the conflict.

That point is worth stressing because commentary on propaganda often becomes one-sided. There is a temptation to imagine that one faction generates narratives while the others merely report reality. No serious observer should accept that picture. Hamas has every incentive to maximise the political effect of civilian suffering. Israel has every incentive to minimise its responsibility for that suffering. International organisations possess their own institutional cultures and assumptions, while campaigning NGOs are often no less committed to predetermined conclusions than governments are. None of this means their information should be discarded. It means that information should be read with some understanding of where it came from.

The Israeli Government’s difficulty is that its communications problem may no longer be a communications problem. American sympathy for Israel has declined since 2023, most noticeably among younger people, and Israeli officials naturally attribute much of this to hostile coverage. Some of it certainly is hostile. But the public has also watched the war continue. It has seen Gaza reduced on television screens to ruined streets and large numbers of displaced civilians. It has encountered repeated reports of hunger. It has heard statements from Israeli politicians that critics then cite as evidence of wider intentions. One may argue that these things are presented without context. One may argue that Hamas deliberately fights from civilian areas and prolongs the war by refusing to surrender. Such arguments deserve to be heard. What they cannot guarantee is that, once supplied with the context, the audience will reach the conclusion Israel wants.

This has always seemed to me the weakness at the heart of hasbara. It treats disapproval as misunderstanding. Sometimes misunderstanding is the cause. At other times people understand well and remain opposed. No amount of improved messaging can eliminate that possibility. Israeli officials reportedly complained to Ynet that the Government had spent a great deal of money while the situation continued to worsen. The remark suggests that at least some of them understand the problem.

There is also a potential for the entire campaign to work against itself. Hanover was useful only so long as users encountering one of its reports believed that they had found an independent American policy institute. Once journalists explained who had financed it, the informational effect changed. The Web now contains the reports and also the reports about the reports. Future crawlers will collect both. A chatbot searching for Hanover may encounter the Israeli-financed argument, but it may equally encounter an article explaining that the apparently independent institute was created as part of a state communications strategy. The attempt to improve Israel’s reputation in AI-generated answers may therefore increase the amount of online material discussing Israeli attempts to manipulate those answers.

This is one reason I doubt that the campaign is strategically important, though the principle behind it certainly is. Israel happens to be an early and visible example of something many organisations will soon attempt. Governments will produce content designed for AI retrieval. Corporations will do the same when defending products or commercial practices. Political movements will attempt to ensure that their preferred vocabulary becomes the vocabulary through which machines describe contested questions. None of this requires a conspiracy. It follows naturally from the incentives created by the technology.

The more troubling issue concerns the users. Artificial intelligence is a useful research tool. It can summarise arguments and locate relevant material. It can make unfamiliar disputes easier to approach. But it does not possess independent experience of the subjects it discusses. ChatGPT has not visited Gaza. It has not examined the internal deliberations of the Israeli Government. It has not independently audited Palestinian casualty records. When it gives a confident answer about genocide or military proportionality, the confidence comes from the form of the language rather than from any first-hand investigation.

This does not make the answer worthless. It makes the answer derivative. The model is processing claims produced by institutions that have their own access to evidence and their own reasons for presenting that evidence in particular ways. The greater the political controversy, the more important it becomes to recover those underlying sources rather than treating the machine’s synthesis as a final judgement.

I suspect this lesson will be forgotten because artificial intelligence is so convenient. It removes the unpleasant labour of comparing contradictory accounts. Instead of reading an Israeli document and then a Palestinian one, followed perhaps by an NGO report or a court filing, the user can ask for a conclusion in twenty seconds. The machine obliges. The disagreement disappears into a paragraph. The disappearance is an illusion. The disagreement is still there, buried underneath the prose.

This is why the Hanover affair seems to me less a warning about Israel than a warning about how we now acquire information. A source should have a pedigree. If the Israeli Government paid for a report, that fact matters. If a casualty figure originates from an institution operating under Hamas authority, that matters as well. If an NGO has spent years campaigning for a particular legal interpretation, that is relevant when assessing its latest report. None of these facts settles the truth of the claims being made. They tell us how carefully the claims should be examined.

There is nothing novel in this principle. Historians have always worked with interested sources. Governments keep records because governments have interests. Political movements preserve documents because they hope posterity will accept their version of events. Memoirs are written partly to justify the people writing them. The task has never been to find a perfectly disinterested source, since such sources are rare. The task is to understand the source well enough to know how much weight it deserves. Artificial intelligence does not remove this requirement. It only makes it easier to forget.

I therefore doubt that Israel’s expenditure will produce the political result its sponsors hope for. The Americans most sympathetic to Israel already possess settled views and are unlikely to need help from ChatGPT. The younger Americans among whom Israeli support has declined are more likely to distrust the old authorities and more likely to notice attempts at manufactured authority when these are exposed. A successful citation here or a favourable paragraph there may be worth something to a public-relations firm. It is not the same thing as reversing a generational change in opinion.

The broader development is nevertheless important. A market now exists for influencing what artificial-intelligence systems retrieve, and where a market exists someone will sell into it. Israel will not be the last government to try this. The only sensible defence is the old one: read beyond the summary and ask where the claims came from.

That may seem an oddly traditional conclusion to draw from a story about artificial intelligence. I think it is the right one. The technology is new. The temptation is not. Governments have always wanted citizens to receive information in the most favourable possible form. What has changed is that citizens are increasingly asking machines to decide which information deserves to be placed in front of them.

The Israelis have noticed this. So should everyone else.