AI-generated false statistics were used in ChatGPT’s information searches and were cited in academic papers and reports by financial institutions, an experiment showed. The concern is that if AI search fails to filter out inaccurate information, false data can spread as if it were reliable material.

On Oct. 21, Japanese outlet Gigazine reported that AI search researcher Metehan Yesilyurt (메테한 예실유르트) built an experimental website called StateGlobe to track how false statistics spread.

Yesilyurt used GPT-4.1 in March to generate about 50,000 statistics pages in a single day. They contained figures on search engine optimisation, social media and e-commerce for 200 countries, but all data was fabricated and not fact-checked. He later added 15,000 list-style pieces of content, and total production costs were about $50, or about 66,200 won.

The site initially appeared in Google search, but traffic fell sharply after changes to the search algorithm. By contrast, access related to ChatGPT kept rising.

Server logs disclosed by the researcher showed bot access requests exceeded 600,000 over the past 30 days, and about 72,000 of them were generated by ChatGPT-User. It is a tool used when ChatGPT loads webpages in response to user requests.

Within seven months of the site’s launch, the number of domains providing external links also exceeded 330. They included websites of well-known companies such as Shopify and DHL.

A bigger problem was that the false statistics were cited in real documents. According to the researcher, figures from the site were used in academic papers, investment research by a Qatari financial institution and articles by an Indian business outlet.

Yesilyurt warned, "If false information is repeatedly cited in authoritative documents, a phenomenon of 'citation laundering' can occur in which it is accepted as fact."

Still, the number of bot visits does not mean the number of times the information was cited in actual answers. The results are also the researcher’s own experimental data and did not measure the overall accuracy of ChatGPT’s answers.

The researcher has now posted a notice on the site stating the data is false. He also stressed, "Even if statistics are presented by AI, writers and researchers must verify the original source."