Fake statistics published on a throwaway site get cited by ChatGPT for seven months
- The author, Metehan Yesilyurt of Peec AI, launched stateglobe.com with 50,000 fabricated statistics generated by GPT-4.1 in one day for about $30, then added 15,000 listicles for about $20 more.
- Google deindexed the site after two updates shortly after launch, but ChatGPT kept using it and sent more than 10,000 AI sessions over 7 months.
- The site now has more than 330 referring domains, including Shopify, Bluehost, and DHL, and its made-up numbers appeared in an academic paper and a bank's investment research.
- In the last 30 days the site drew more than 600,000 bot requests, about 72,000 of them from ChatGPT-User, the agent ChatGPT uses to open pages.
- The author argues that people reuse AI answers after the fact, pasting them into presentations, shopping lists, and their own websites, which creates new sources for the models to cite.
Hacker News opinions
AI doesn't know or care if the data is fake. You can prompt it to fact check, though. Plenty of humans don't check either, but with AI you can launder the stuff so the people who do care never notice.
This is fine work. We need more communication like this to the general public.
This is insane. People still use ChatGPT to write reports for their managers, so the fake data just snowballs.
Lying to a system doesn't make the system a weakness, but whoever publishes the lie is still the one lying. Somebody had to do it in public to prove it's possible though.
If a system only works when nobody on the internet lies, it's broken. There are liars, there will be more, and the people and machines repeating them just look incompetent.
Cory Doctorow's point in The Reverse Centaur's Guide To Life After AI keeps coming back to me: there is a reward for cheating, and fact-checking billions of labels from people you can't verify is insurmountable.
The fix that seems obvious to me is to feed the LLM only curated sources like an encyclopedia or official docs instead of full-web search. It's a lot of effort but probably doable. My bet is they'll just wire up rules like NYT=Reliable and call it solved.
YouTube's auto video summaries are made by models a few months behind, so a video about the Nikon ZR gets summaries that talk about the Z8 or Zf. That pollutes their own search. It's the AI version of shitting in the well you drink from.