In the latest issue of our monthly newsletter:
* Three papers on how Wikipedians use (or don't use) LLMs
* Experienced editors contribute more to "underproduced" articles than newbies
and more: meta.wikimedia.org/wiki/Research:…

"Specificity Bias and the Wikipedia Paradox: An xAI Analysis of AI Text Detectors under Domain Shift" - purely statistical detection paradigms face systematic limitations in highly structured human domains such as Wikipedia

(Morloher et al, 2026)

arxiv.org/pdf/2611.03280

"What AI Gets to See: Auditing the Loss of Multilingual Community Knowledge in Machine-Facing Metadata on Wikimedia Common" multilingual NLP should include auditing and repairing recovery gaps in public knowledge infrastructures

(Yu and @aylin_cim, 2026)

aclanthology.org/2026.emnlp-mai…

"WPTokens: A Dataset for Studying the Regulation of Wikipedia" Sentence- and node-level change tracking over the full revision history of French Wikipedia

(@leojoubert, 2026)

arxiv.org/pdf/2610.07108

"When the Commons Appropriates a Large Language Model: How WikiVault Reshaped Korean Wikipedia" amplifying production capacity, shifting work to reviewing articles, and importing editorial judgments from enwiki
arxiv.org/pdf/2610.07660

@song_inhwa_, @sohwng, T.Yoo, @manoelribeiro

"Editorial. Wikidata Across the Humanities: Datasets, Methodologies, Reuse" a special collection dedicated to the publication of discussion and data papers that explore the use of Wikidata in the digital humanities.

openhumanitiesdata.metajnl.com/articles/10.53…

@mappingreligion @up_johd

"Pluralistic Annotation of Cultural Weaponization on Wikipedia: A Multi-Annotator Case Study across Ukraine and Armenia" examining cultural heritage weaponization through pluralistic human annotation and LLM-assisted analysis

(Hamdane et al, 2026)

aclanthology.org/2026.pandora-1…

"Can Language Models Play the Wikipedia Game?" The game start swith random source and target Wikipedia articles, and participants navigate between them by only clicking on blue links.

(Hickey et al, 2026)

arxiv.org/pdf/2610.13948

"The Wikipedia Edit-a-thon Companion App: A Crowdsourcing Tool for Reparative Description with Oral History Transcripts"

(Ballmer et al, 2026)

dl.acm.org/doi/full/10.11…

"Beyond Initial Creation: Two Decades of Community-based Chart Revisions from Wikimedia Commons" a novel dataset of 7,786 longitudinal chart revisions across 1,664 images.

(Shih et al, 2026)

arxiv.org/pdf/2610.19337

"WikiFlow: Predicting Wikipedia Editor Dropout at Scale Using Distributed Machine Learning on Google Cloud" doi.org/10.1007/978-3-…
#paywall💰
4 "editor archetypes":
Frustrated burst editors
Casual contributors ("most valuable")
Power/talk-page editors
Focused/bot-like accounts

"A Corpus for Evaluating Neutrality in Political Biographies on Wikipedia: NeutralWikiPol-Fr" dx.doi.org/10.1007/978-3-… #paywall💰
"LLMs identify biased words effectively [but] performance drops when bias depends on the broader context"
Human experts' "low agreement" on neutrality

Dataset: github.com/Vic-SP/NWP-fr ("an exploratory, fine-grained dataset designed to capture implicit, context-dependent semantic biases in French political biographies on Wikipedia")

"Linking Speakers of the German Parliament to Wikidata: Scope and Coverage of Metadata" arxiv.org/pdf/2609.18289
For Bundestag members, "Wikidata is almost complete" for basic properties like gender or DOB, but coverage of "CV" fields like "educated at" is "significantly weaker"

"Can Language Models Play the Wikipedia Game?" (get from one article to another via article links) arxiv.org/html/2610.1394…
"No model significantly outperformed humans, although the best models (GPT-o3 mini, GPT-4o, QwQ:32B, and DeepSeek-R1:70B) achieved comparable success."

An new Pleias paper: in collaboration with AI Bridges and @Wikimedia Deutschland we developed a new harness for complex semantic search over the largest knowledge database, @wikidata

What happens when a Wikipedia language edition starts using LLMs?

Unlike in other languages, Korean Wikipedia embraced LLMs through WikiVault: only 0.1% of editors use it weekly, but it accounts for ~30% of added bytes!

New preprint led by @song_inhwa_ (who's at @ACM_CSCW)

🧵

"WikiCoach: Scaffolding Learning for Novices in Online Apprenticeship Communities" doi.org/10.1145/377493…
"a Q&A–style LLM system designed to scaffold learning-through-doing in Wikipedia"

"WiCleanData: Guaranteeing the Type Consistency of Wikidata by Taxonomy Refinement and Constraint Enforcement" arxiv.org/abs/2609.20057
Pipeline "cleans the taxonomy with [LLM] assistance, then simplifies type constraints by hierarchical aggregation, and finally filters facts"

Web interface for the resulting knowledge graph: wicleandata.r2.enst.fr