From catchphrases to structural clues
Wikipedia volunteers working to remove rule-breaking AI-generated material are adapting as machine-written text becomes harder to identify. WikiProject AI Cleanup, a volunteer effort launched in 2023, initially found articles through conspicuous phrases, formulaic conclusions and sections that looked unlike normal encyclopedia entries.
Project cofounder Ellia Lebleu told IEEE Spectrum that newer AI writing is often less obvious. Editors now look for combinations of structural habits rather than treating one phrase as proof. They also examine editing behavior: an account producing several long articles within minutes can warrant scrutiny even when the prose itself contains fewer recognizable signals.
Citation quality is a major part of the work. Lebleu said fabricated references have become less common as retrieval tools improve, but real sources are still frequently attached to claims they do not support. That problem requires editors to inspect the relationship between an assertion and its cited material, rather than merely checking whether a link or publication exists.
Tools support, rather than replace, review
Volunteers are experimenting with automated assistance. One editor created a source-verification tool that checks whether references appear to support claims. Because the checker is itself based on an AI model, its judgments are not guaranteed, but Lebleu described it as a useful first filter for obvious mismatches at scale.
The Wikimedia Foundation is also developing Paste Check, intended to flag text pasted directly from some AI chatbots. Commercial chatbot refusals can provide another barrier when systems decline requests to generate a Wikipedia article, though open-weight models and configurable tools mean refusals cannot prevent all submissions.
Lebleu said work with detection company Pangram produced a conservative estimate that about 20 percent of new articles and drafts in late 2025 were AI-generated. The measured share fell to roughly 15 or 16 percent in early 2026. She cautioned that the decline could reflect policy effects, improved awareness among newcomers or detection software failing to keep pace with newer models.
English Wikipedia has prohibited using large language models to generate articles. According to Lebleu, beginners account for most of the AI-generated submissions the project encounters, so editors have added clearer warnings to welcome messages, help material and guidance. It remains too early to determine how much those measures have changed behavior.
The effort also raises a longer-term question about Wikipedia's identity. Lebleu argued that a human-written encyclopedia has value both to readers and as a distinctive corpus in an internet increasingly filled with generated text. At the same time, AI tools may eventually help with source checking, attribution and requests from readers who discover errors through AI search products. For now, the project's immediate task remains narrower: identify unsupported machine-written material while avoiding claims that stylistic clues alone can prove how text was produced.
