Wikipedia Tracks AI: 15 to 20% of Articles Flagged

The Wikipedia community has banned the writing of articles by language models and is gearing up to filter automated contributions. Estimates from a detector tested prior to ChatGPT suggest that around 20% of articles and drafts were flagged by the end of 2025, dropping to 15 to 16% in early 2026, with no certainty about the causes of this decline. Verification tools and a future attribution framework are in preparation, while the status of a "human-written" corpus is now fueling deeper reflection.

Quantitative Signals, but an Unexplained Decline

By the end of 2025, approximately 20% of new articles and drafts were flagged as AI-generated by the Pangram tool. By early 2026, the reported share had fallen to 15 or 16%. Pangram had produced no false positives when tested on content prior to ChatGPT, making it, according to Ellia Lebleu, a very conservative lower limit. The cause of the recent decline remains uncertain: it is not established whether it results from the policies implemented, an evolution of the models, or a lag in detection adaptation. It is deemed too early to determine the effectiveness of the ban. However, possible indicators are emerging, including prompt refusals and an apparent decrease in automated volume. The majority of AI-generated content is believed to come from novice editors, which has led to increased visibility of the rule in welcome messages, editing help, and guidelines.

Homegrown and Institutional Tools for Large-Scale Filtering

Wikipedia has recently prohibited the use of language models for writing articles. In parallel, an editor from the WikiProject AI Cleanup has designed an AI-based source checker that confronts cited sources with the claims on the page. Without a guarantee of systematic accuracy, it aims to operate at scale and eliminate obvious cases. On the institutional side, the Wikimedia Foundation is working on Paste Check, intended to detect text pasted from certain chatbots. Some models, like Claude, sometimes refuse to produce a Wikipedia article in accordance with the encyclopedia's policies. If such refusals were widespread among major AI companies, it would not eliminate the problem due to adjustable open-weight models, but it would assist users of commercial solutions who are less familiar with the technology.

Detecting AI Through Writing Patterns and Editing Behaviors

The issue emerged around December 2023 when editors on a Discord server noted articles with recurring phrases, such as "stands as a testimony" or "richness of cultural heritage," sometimes accompanied by conclusion sections or "Why is this important?"—traits attributed to ChatGPT in 2023 and early 2024. These markers, unusual for Wikipedia, were then easy to spot and led to the creation of an official project. By 2025, the indicators became more subtle and structural: a preference for the rule of three, negative parallelisms, or other figures whose accumulation fuels suspicion. Patrols also observe editing rhythms, such as the publication of five articles of several thousand words in five minutes. Volunteers are still developing tools and debating the role to be assigned to AI.

Sources: The Critical Point of Automated Deviations

False or misleading citations have become a major issue. Models sometimes hallucinate entire references, a phenomenon deemed much rarer today thanks to augmented generation through retrieval. More frequently, according to Ellia Lebleu, is the use of a real source whose content is misinterpreted. Citations produced by AIs that do not actually support the claims are described as extremely common.

Preserving a "Human-Written" Corpus While Testing Assistance

The anti-AI writing directive initially responded to an influx of content and a lack of tools. The reflection has expanded towards the identity of Wikipedia, sometimes referred to as "the last good place on the Internet," and the idea of an encyclopedia written by humans. This characteristic also interests AI companies, as a human corpus could be potentially valuable for training. Ellia Lebleu believes that Wikipedia remains, and will remain in the short term, this unique corpus. It remains to be seen whether readers and contributors would accept a degree of generation if it increased completeness. One proposed avenue would be to preserve the identity of the human corpus while offering an option for AI assistance.

In Response to Google’s "AI Snapshots," an Attribution Framework in Development

The rise of AI-assisted searches raises the question of visibility: Wikipedia is frequently cited in Google’s AI Snapshots, but users may only see a synthesized answer and be unaware that they can edit the encyclopedia. The Wikimedia Foundation is preparing an attribution framework intended to display the date of the last modification, the degree of completeness, any reports from editors, and possibly a call to action. Ultimately, readers might even submit modification requests via AI models by adding a source. Ellia Lebleu anticipates a marked symbiotic relationship between Wikipedia and AI models, with the encyclopedia not disappearing but evolving in form. She summarizes the perceived evolution: AI was to be cleaned up yesterday, it helps with cleaning today, and it could become omnipresent tomorrow.