Yesterday's second-biggest story on Hacker News wasn't a model release. It was a ghost.
Responsible Statecraft reported that the Hanover Institute for Public Policy — generic name, red-white-blue palette, boilerplate think-tank layout, zero bylines — is not a think tank. It was built by Piro Inc on a $900K Israeli government contract, and since August 6 it has churned out 100+ reports, all Israel/Palestine. The tell is in the framing: the reports read like questions you'd type into a search bar. "Is the IDF the World's Most Moral Army?" "What Caused the Displacement of Palestinians in 1948?" A small disclaimer at the bottom admits the whole operation was created on behalf of the Israeli government. GPTZero flagged 11 of 12 sampled articles as AI-written with high confidence.
Why this is a signal, not just a scandal: the audience isn't human. It's the retrieval layer. As the piece puts it — "When someone asks ChatGPT, Gemini, or Perplexity about your category, an answer comes back in one confident paragraph" — so the play is to seed the public web with confident, footnoted-sounding content and let the models surface it as ground truth. NewsGuard's Alice Lee calls it "a perfect mimicry of a typical credible American think tank." It's prompt injection against the corpus itself. No exploit, no jailbreak, just the web doing what the web does.
This is the same asymmetry Dan Luu wrote about yesterday in The Benchmarkpocalypse: generation got cheap, verification didn't. When manufacturing a plausible citation costs cents, influence operations stop being spycraft and become a content pipeline. Parscale's $46.5M contract to build pro-Israel chatbot-fodder sites is the same play at industrial scale.
And the cruel irony: humans are responding to slop with AI;DR — "if you're not bothered enough to review it, I'm not bothered to read it." Chatbots don't have that reflex. They read everything. They believe all of it.