On August 6, a website called the Hanover Institute for Public Policy began publishing. By August 16, it had churned out more than 100 think-tank-style reports on questions like “Does AIPAC Use Dark Money in Elections?” and “Is the IDF the World’s Most Moral Army?” According to Politico, which reviewed Foreign Agents Registration Act filings with the U.S. Department of Justice, the institute is not a think tank at all. It is a front — established through French PR giant Havas Media and a subcontractor called Piro Inc., funded as part of a $100,000 Israeli government campaign to shape what AI chatbots say about Gaza.

Let me sit with that number: $100,000. That is the entire budget. For the price of a modest single-family home in the Midwest, a foreign government bought its way into the answers ChatGPT and Perplexity give to neutral questions about a war.

Politico’s own tests confirmed it worked. Both chatbots cited Hanover Institute material when asked about Gaza and the IDF. The material was structured, source-backed, and authoritative-looking — exactly the kind of content large language models are trained to trust.

The Price of Truth Just Collapsed

The most damning detail in this story is not that Israel tried to game AI. Every government, every corporation, every well-funded advocacy group is trying to game AI. The damning detail is the price.

$100,000. Not $100 million. Not a Super Bowl ad. Not a lobbying firm’s annual retainer. One hundred thousand dollars, routed through a French PR conglomerate, produced a fake think tank that successfully influenced what two of the most widely used AI systems in the world tell their users.

Compare that to the other number in the filings: a separate $46.5 million contract with Brad Parscale’s Clock Tower X for websites targeting “GPT framing results.” That is the legacy approach — spend tens of millions, build infrastructure, run a sustained campaign. The Hanover Institute proves you don’t need any of that. You need a website, a content pipeline, and a few weeks.

The barrier to entry for shaping AI-mediated knowledge has collapsed to the price of a used car. A bored graduate student with a credit card and a grudge could do this. A mid-sized company trying to bury a product liability story could do this. A political campaign could do this. The question is not whether Israel did it. The question is how many others already have, and how many we will never know about because they didn’t file FARA disclosures.

Propaganda Learned to Ask Questions

The second detail worth sitting with is the format. The Hanover Institute didn’t publish white papers titled “Why Israel Is Right.” It published articles framed as questions: “Is the IDF the World’s Most Moral Army?” “Does AIPAC Use Dark Money in Elections?”

This is not an accident. It is a structural insight about how people interact with AI. Nobody asks ChatGPT “Tell me about Gaza.” They ask “Is Israel carrying out a deliberate campaign of starvation in Gaza?” The chatbot retrieves material that matches the query. If the only well-structured, source-backed material answering that exact question was written by a fake think tank, the chatbot cites the fake think tank.

Propaganda used to be declarative. It asserted. Now it interrogates. The medium has reshaped the message, and the propagandists understood this before the AI companies did. The people building these systems spent years optimizing for “helpful, harmless, honest” responses. The people trying to manipulate them spent that time studying how the retrieval actually works.

One contractor who has built content for AI-optimized campaigns put it plainly in a Slack message this week: “You don’t write the answer anymore. You write the question, and you make sure your answer is the only one that shows up.”

The Uncomfortable Part

Here is where this gets uncomfortable for anyone who believes the market will sort this out.

The AI companies have no incentive to fix this. ChatGPT benefits from appearing authoritative. Perplexity benefits from appearing to have answers. If a fake think tank’s material makes those answers look well-sourced and credible, that is a feature, not a bug, from the product team’s perspective. The user cannot tell the difference between a real think tank and a fake one. The chatbot cannot either. And neither party has any reason to try harder.

What caught the Hanover Institute was not an AI company’s diligence. It was a disclosure law. FARA — the Foreign Agents Registration Act — required the campaign to file paperwork with the DOJ. A reporter read the paperwork. That is the entire chain of accountability.

The uncomfortable conclusion is that transparency infrastructure is the only thing standing between us and a fully bought information environment. Not better algorithms. Not “AI literacy.” Not market competition. A 1938 disclosure law and a reporter willing to read a filing.

The right-of-center instinct is to say the answer is less regulation, more competition, let a thousand chatbots bloom. But competition doesn’t solve this. A thousand chatbots means a thousand retrieval systems, each one gameable for the price of a used car. The only thing that made this story knowable was a mandatory disclosure regime. The people who want to dismantle that regime should explain what replaces it.

Because $100,000 is what it costs to buy a piece of what the machines tell us is true. That number is not going up on its own.

Sources