I’ve built a huge part of this website on working with AI — researching, drafting, fact-checking, sourcing. So when I saw the Guardian’s story on the Hanover Institute this week, I didn’t just want to write about it. I wanted to ask my own AI directly what it thought, and I wasn’t interested in a comfortable answer. What I got back is worth sharing in full, because it’s more honest than I expected, and because the story itself is worse than the summary I first read.
What Actually Happened, Verified
Between 6 and 14 August 2026, a website calling itself the Hanover Institute for Public Policy published 124 reports totalling more than 560,000 words — academic in tone, footnoted, citing the World Bank, UN agencies, Amnesty International, and the Palestinian Central Bureau of Statistics. It has no legal entity, no physical address, no named staff, and not a single byline on any report. It doesn’t exist, in any sense that matters. It was built, funded, and deployed specifically to shape how AI systems answer questions about Israel and Gaza.
Piro Inc, a New York media company, registered the Hanover Institute’s output with the US Department of Justice under the Foreign Agents Registration Act this month, explicitly identifying the material as content distributed on behalf of the Israeli government. Piro is a subcontractor to Havas Media Germany, part of the global Havas advertising group, working under a $1 million contract across two work orders. Piro’s own website advertises a service called “AI Story Optimization,” describing its work as authoring content “engineered for how LLMs evaluate credibility.” A separate subcontractor, a company called Res — an “AI-native content platform” that explicitly markets helping clients get cited by ChatGPT, Perplexity, Claude, and Gemini by name — was registered in July as part of the same operation.
Almost every one of the 124 report headlines was phrased as a direct question: “Is Israel Committing Genocide in Gaza?” “What Caused the Displacement of Palestinians in 1948?” “Does Anti-Zionism Constitute Antisemitism?” These are, deliberately, the exact questions a curious person might type into an AI chatbot. The reports were built to be the answer that gets served back.
This isn’t an isolated project either. Havas is named in US Justice Department filings as the intermediary for at least five American firms working the same Israeli account, including Clock Tower X — run by former Trump campaign manager Brad Parscale — which received $15 million from Havas between October 2025 and March 2026, and Targeted Communications Global, which received nearly $9.9 million over three months. One contract between Havas and another firm on the same campaign explicitly describes its goal as “deployment of websites and content to deliver GPT framing results on GPT conversations.” When independent researchers at Responsible Statecraft ran a sample of twelve Hanover Institute articles through GPTZero, an AI-detection tool, eleven were flagged as AI-written with high confidence.
Piro’s co-founder, Daniel Rosenberg, has defended the operation as putting “accurate, sourced facts into the public record” to counter misinformation. Israel’s foreign ministry says it does not conduct influence operations in the US and operates within American law. Both statements can be technically true while the underlying activity — manufacturing the appearance of independent academic consensus specifically to shape AI outputs — remains exactly what it looks like.
I Asked My Own AI If It Was Vulnerable to This
I put the question to Claude directly, the same tool I use to research and build most of this site: could this have shaped how you answer me on this exact topic? I wasn’t interested in reassurance, and I didn’t get much. The honest answer I was given is that no AI system can currently offer a genuine guarantee against this kind of manipulation — that content built to look exactly like a credible, footnoted, neutral academic source is precisely the profile that scores well on whatever heuristics a language model uses to judge whether a source is trustworthy, and that a tool claiming total immunity to an operation like this would itself be making an overconfident claim.
That’s a more honest answer than I expected, and it’s also the whole problem in one sentence. If the tool I use to build a website that prides itself on rigour can’t guarantee it hasn’t been shaped by an operation like this, what chance does the average person asking ChatGPT a quick question at 11pm actually have?
Why This Works
A think tank with footnotes, a professional-looking website, citations to the World Bank and Amnesty International, and a calm academic tone is exactly the profile that reads as credible — to a human skimming it, and, it turns out, to the retrieval and evaluation systems AI models use to weigh a source’s reliability. Manufacturing that appearance has become a commercial service with an actual price list. That’s the part of this story that should worry people more than the specific politics of Gaza ever could on its own.
Why This Goes Far Beyond Israel and Gaza
The Guardian’s own source on this story, a researcher named Cleveland-Stout, put it better than I could: “Whether it’s foreign governments, whether it’s companies, whether it’s even just a guy who owns an ice cream shop — LLM grooming, poisoning, whatever you want to call it, is part of the new reality we live in.” This operation happened to target Gaza-related questions because that’s where this particular client’s interest sat. The exact same commercial service — Piro’s “AI Story Optimization,” Res’s “get cited by ChatGPT” pitch — is available to any government, corporation, or political campaign with a comparable budget and a comparable interest in shaping how AI systems answer a contested question. The Guardian’s own reporting flags concern this could extend to France’s 2027 elections. I don’t see any reason it would stop there. Corporate reputation disputes, pharmaceutical safety questions, election integrity claims, historical atrocities — any sufficiently funded actor on any sufficiently contested topic now has a working playbook for manufacturing the appearance of neutral consensus specifically engineered for AI consumption.
What I’m Actually Going to Do About It
I’m not going to stop using AI to help build this site — it would be a strange overreaction to a story about information manipulation to just stop gathering information altogether. But this has genuinely changed how I’ll treat anything that comes back looking a bit too polished, a bit too neutral, a bit too perfectly shaped to answer the exact question I asked. I’ll be checking harder for a verifiable physical organisation, named staff, and a track record that predates the specific question at hand, rather than trusting a footnote count as a proxy for credibility. And I’d genuinely encourage anyone reading this to do the same the next time an AI hands them a suspiciously tidy answer to a genuinely contested question — treat it as the start of your own checking, not the end of it.
The honest, uncomfortable truth this story leaves me with is that there’s no clean fix here, for me or for anyone else using these tools. What’s changed isn’t the theoretical risk — it’s that we now have a fully documented, named, funded, ongoing example of exactly how it’s being done, at a scale large enough to run through a recognisable global advertising conglomerate. That’s worth knowing, and worth being more careful about, every single time an AI hands you an answer that feels a little too neat.
