How Wikipedia’s Editors Keep the Online Encyclopedia Human-Written

WikiProject AI Cleanup

Wikipedia's volunteer community launched the WikiProject AI Cleanup in 2023 to systematically address the influx of AI-generated content. This initiative was born out of observations by editors on Wikipedia's Discord server, who noticed peculiar characteristics in newly added articles. These early signs included repetitive, clichéd phrases such as "stands as a testament" or "rich cultural heritage," alongside unusual structural elements like dedicated conclusion sections, which were atypical for Wikipedia's editorial style.

The project's initial focus was on these easily identifiable patterns, which were common outputs from large language models (LLMs) like ChatGPT in late 2023 and early 2024. As AI writing capabilities evolved, the project adapted its detection methods. The ongoing effort underscores the dynamic nature of content moderation in the age of generative AI, requiring continuous innovation from the volunteer base to keep pace with technological advancements.

Ellia Lebleu

Ellia Lebleu, a founding member of the WikiProject AI Cleanup, has been instrumental in shaping the project's strategies. Lebleu, who is also pursuing a Ph.D. in AI at the Inria Center of the University of Rennes, observed the shift in AI writing patterns by 2025. The tell-tale signs became less about obvious clichés and more about subtle structural elements, such as the AI's preference for the "rule of three" or negative parallelisms.

Beyond textual analysis, Lebleu and her team also consider behavioral indicators. For instance, an editor producing multiple thousands-of-word articles within minutes raises immediate suspicion of AI assistance. This holistic approach, combining linguistic analysis with user behavior, is vital for identifying sophisticated AI contributions that might otherwise evade detection. The project also grapples with the pervasive issue of fake or misleading citations, a significant concern for an encyclopedia that relies heavily on verifiable sources.

Retrieval-Augmented Generation

The challenge of AI-generated content extends significantly to the realm of citations, where the problem has become particularly acute. While early AI models might hallucinate citations entirely, the advent of retrieval-augmented generation (RAG) has changed the landscape. RAG allows AI to find real sources, making outright fabrication rarer. However, the new problem is that AI often misinterprets or misrepresents the content of these real sources.

This results in citations that, while pointing to legitimate documents, do not actually support the claims made in the Wikipedia article. Such misleading citations are now extremely commonplace, posing a more insidious threat to the encyclopedia's accuracy than outright fabricated sources. English Wikipedia has responded by prohibiting editors from using LLMs to generate articles, reflecting the community's concern over the integrity of its content and the difficulty in verifying AI-produced information, even when seemingly sourced.