Mathias Schindler, a longtime Wikipedia contributor and co-founder of Wikimedia Germany, reports on a troubling discovery at the 39C3 conference in Hamburg. While developing a tool to check ISBN checksums in German Wikipedia, he uncovered a significant problem: articles containing completely fabricated literature references generated by large language models.
The issue emerged when Schindler found ISBNs with incorrect checksums that didn’t match any real books in library catalogs. The book titles and author names appeared plausible, sometimes even using names of actual scholars in relevant fields. However, these references were entirely hallucinated by ChatGPT and similar tools. Users had asked AI systems to write Wikipedia articles and simply pasted the results, including fake citations, into the encyclopedia.
“These literature references were hallucinated by, in most cases, by ChatGPT,” Schindler explains during his presentation. He describes this as “anti-knowledge,” the antithesis of what an encyclopedia project represents.
A problem beyond Wikipedia
The problem extends beyond Wikipedia. The University Library of Hagen reported students requesting books with plausible journal names and issue numbers that don’t exist. The International Committee of the Red Cross issued warnings about hallucinated references, noting that librarians are sometimes accused of hiding truth when they cannot locate the nonexistent sources.
Schindler identifies several possible motivations for users dumping AI content into Wikipedia. Some may genuinely not understand the limitations of large language models. Others appear aware they cannot contribute meaningfully and compensate with AI assistance. A more concerning category involves users weaponizing LLMs to deliberately rewrite history or push agendas.
The irony, Schindler notes, is that Wikipedia serves as prominent training data for these same language models. “The large language model providers are in some way poisoning their own pond from which they are drinking the water from,” he observes. He adds that AI companies now actively seek content proven free from synthetic information, paying premiums for uncontaminated data.
New policies
Wikipedia communities are responding with new policies. English Wikipedia introduced a speedy deletion rule for blatantly obvious LLM-generated content. German Wikipedia has seen success in reducing ISBN errors through community efforts. However, Schindler emphasizes that identifying AI content remains challenging as the technology improves.
His litmus test involves asking users to share their prompts. The excuses he receives range from “I was never logged in” to claims that prompt information is sensitive. Few comply with the request.
Schindler openly acknowledges using Claude Opus to write his ISBN checker tool, releasing the code on GitHub with clear disclosure. He argues this represents acceptable AI use, distinguishing between using LLMs as coding assistants versus content generators.
He calls for wider discussion on AI responsibility, improved communication about limitations, and potential legal or political pathways to pressure companies into better disclosure. The fundamental question remains whether humans can maintain encyclopedic knowledge in an AI-saturated environment.
