---
id: "lib-2026-10-21-wikipedia-s-cleanup-of-ai-generated-content"
type: "library"
title: "Wikipedia's cleanup of AI-generated content"
lang: "EN"
org: "IEEE Spectrum"
source_type: "Article"
date: "2026-10-21"
week: "2026-W44"
url: "https://nosov.ai/library/lib-2026-10-21-wikipedia-s-cleanup-of-ai-generated-content/"
original_url: "https://spectrum.ieee.org/wikipedia-wikiproject-ai-writing-cleanup"
tier: "B"
tier_name: "Useful, limited"
tier_reason: "An IEEE Spectrum interview with a founder of WikiProject AI Cleanup gives a practitioner's account and one detector-based estimate of AI-written new articles; useful context, but the numbers are informal and the method is described in two sentences."
summary_basis: "full-text"
original_minutes: "4"
tags: ["ai technology", "future of work", "enablement", "business transformation", "content moderation", "ai detection", "wikipedia"]
---

# Wikipedia's cleanup of AI-generated content

Tier B (Useful, limited): An IEEE Spectrum interview with a founder of WikiProject AI Cleanup gives a practitioner's account and one detector-based estimate of AI-written new articles; useful context, but the numbers are informal and the method is described in two sentences.

Summary based on the full text.

## In short

An interview by Eliza Strickland (IEEE Spectrum, 21 October 2026) with Ellia Lebleu, a founding member of WikiProject AI Cleanup and a PhD student in AI at Inria in Rennes. Volunteers moved from spotting obvious chatbot clichés in 2023 to looking for structural patterns and editing behaviour. Using the Pangram detector, Lebleu estimates that a lower bound of 15 to 16 percent of new articles and drafts were AI-generated in early 2026, down from about 20 percent in late 2025. She cannot say whether the drop comes from the new ban or from models that detectors have not caught up with.

## Key numbers

- 20 percent: share of new Wikipedia articles and drafts flagged as AI-generated by Pangram in late 2025
- 15 or 16 percent: same share in early 2026, after English Wikipedia banned LLM-generated articles
- 2023: year the volunteer WikiProject AI Cleanup was launched

## What it reports

- Early signs (December 2023): stock phrases such as "stands as a testament", conclusion sections and "Why is this important?" headings.
- By 2025 the signs were structural: rule-of-three lists, negative parallelisms, and behaviour such as several long articles written in minutes.
- Citations are the larger worry: invented references are now rarer because of retrieval, but real sources cited for claims they do not support are "extremely commonplace".
- English Wikipedia now prohibits editors from using LLMs to generate articles. Most AI content comes from beginner editors, so the rule is placed in welcome messages and editing help.
- Pangram showed no false positives on pre-ChatGPT Wikipedia text, so its rates are treated as a conservative floor: about 20 percent in late 2025, 15 or 16 percent in early 2026.
- Countermeasures: an editor-built AI source verifier that checks whether cited sources match claims, and a Wikimedia Foundation tool, Paste Check, that flags text pasted from some chatbots.
- The Foundation is working on an attribution framework so AI answers citing Wikipedia can show when an article was last edited and whether editors flagged problems.

## How they got it

Interview. The share estimates come from Lebleu running a commercial detector over new articles and drafts after testing it on pre-ChatGPT content. Sample size, period boundaries and detector settings are not given.

## Limits

- Single-person account; no published dataset or method note behind the percentages.
- A detector that misses newer models would also produce a falling rate, which Lebleu acknowledges.
- The AI source verifier is itself a model, with no reported accuracy.

## Why it matters for companies

Wikipedia is one of the largest open tests of keeping a text corpus human-written, and its experience transfers to internal knowledge bases and wikis. Style cues stop working within a year or two; behaviour (volume per minute) and citation checks last longer. The citation finding matters most: a retrieval-backed answer that cites a real document for a claim the document does not make is the common failure, and it needs claim-to-source checks rather than a hallucination filter.

Original: https://spectrum.ieee.org/wikipedia-wikiproject-ai-writing-cleanup