WikiProject AI Cleanup Reports 15-16 Percent AI-Generated Wikipedia Drafts in Early 2026
Wikipedia's volunteer effort tracks a measurable but declining share of AI-generated drafts. Policy bans and detection tooling produce partial effects while open models and subtler generation evade current checks. Sustained verification infrastructure remains necessary to preserve citation integrity.
WikiProject AI Cleanup formed in December 2023 after editors identified repeated phrases and fabricated conclusion sections in new articles. The group documented structural markers such as the rule of three and negative parallelisms that replaced earlier obvious clichés. English Wikipedia then enacted a formal prohibition on LLM use for article generation.
Pangram testing on pre-ChatGPT content produced zero false positives, establishing a conservative lower bound. Volunteers also deployed an AI source verifier that cross-checks claims against references and the Paste Check tool that flags direct chatbot paste. Citation hallucinations persist at lower rates under retrieval-augmented generation.
Open-weight models remain outside commercial refusal policies, limiting the impact of prompt blocks implemented by Claude and similar systems. Detection must therefore shift to behavioral signals such as article velocity and citation mismatch patterns. Operational workload for human reviewers therefore scales with the undetected fraction rather than headline percentages.
Continued adaptation of detectors against newer model releases will determine whether the observed decline continues or reverses.
WikiProject AI Cleanup: AI-flagged share of new drafts drops below 10 percent by Q4 2027 only if commercial model refusal coverage exceeds 80 percent of active users.
Sources (3)
- [1]Primary Source(https://spectrum.ieee.org/wikipedia-wikiproject-ai-writing-cleanup)
- [2]Supporting Source(https://en.wikipedia.org/wiki/Wikipedia:WikiProject_AI_Cleanup)
- [3]Supporting Source(https://meta.wikimedia.org/wiki/Paste_Check)