The (in)ability of AI to evaluate sources is a bit of a preoccupation on this blog. In a post from last month, Metehan Yesilyurt described an experiment he ran to seed the web with fake statistics. He generated 50,000 fake statistics pages and 15,000 listicles for about $50 USD and added them to a brand new website, StateGlobe. Over seven months, he managed to get over 300 domains linking to his site and even a handful of citations in academic, financial, and other publications. This despite the fact that Google, to its credit, quickly deindexed his site, cutting off a huge source of traffic.
Why did it continue to succeed? AI bots. In the thirty days before the post, the site received 600,000 bot requests, including 72,000 from ChatGPT’s page-fetching bot. Over the course of the experiment, ChatGPT referred over 10,000 sessions to his site.
Part of why this problem is particularly pernicious is because, as the author points out, citation laundering means these stats from a random website can eventually be cited by something more authoritative-looking, like an academic publication, which leads to others becoming much more willing to cite it.
I’ve discussed before the problem with LLMs increasingly intermediating our interactions with information on the internet, making it harder for us to judge the trustworthiness of sources. Experiments like this show that fear is justified.
I’m not sure how ethical this experiment was, given that the author has now unleashed an army of zombie statistics on the internet, and you can’t exactly take that back. The author subsequently joined an AI search optimization company and used the results to better market real content (and posted this article on their blog), so it’s hard to sell this as a noble-if-misguided attempt to draw attention to an important problem. It’s just more marketing.
More troubling, the author claims in the article:
It should go without saying that you should never publish fake content. While my experiment relied on tens of thousands of unverified stats, I ran it in order to share the results with the public, and the website, while still live, now carries a disclaimer.
But as far as I can tell, this isn’t really true. The closest I could find on the site is this methodology statement:
These projections for Afghanistan are derived from cross-referencing publicly available government data, industry reports from organizations such as the ITU, World Bank, and Statista, and historical trend analysis. Figures represent 2026 estimates based on compound growth modeling and may differ from actual outcomes. Last reviewed: September 2026.
And this tiny disclaimer in the footer:
Disclaimer: All statistics presented are 2026 estimates and projections based on industry trend analysis, historical data, and publicly available research. Individual data points may vary from actual figures.
I would not exactly call this a clear disclaimer that the data were, in the author’s own words, “made up”—purely fabricated by an LLM with no human verification.
<sup>Hat tip to robtherobber on Hacker News.</sup>