---
id: "lib-2026-10-06-when-the-commons-appropriates-an-llm-how"
type: "library"
title: "When the Commons Appropriates an LLM: How WikiVault Reshaped Korean Wikipedia"
lang: "EN"
org: "Princeton et al."
source_type: "Paper"
date: "2026-10-06"
week: "2026-W41"
url: "https://nosov.ai/library/lib-2026-10-06-when-the-commons-appropriates-an-llm-how/"
original_url: "https://arxiv.org/abs/2610.07660"
tier: "A"
tier_name: "Strong evidence"
tier_reason: "Princeton HCI study (Manoel Horta Ribeiro's group) combining ten interviews with platform-wide edit and pageview logs and matched quasi-experimental comparisons; a preprint on a single community."
summary_basis: "full-text"
original_minutes: "55"
tags: ["enablement", "future of work", "bottom up adoption"]
---

# When the Commons Appropriates an LLM: How WikiVault Reshaped Korean Wikipedia

Tier A (Strong evidence): Princeton HCI study (Manoel Horta Ribeiro's group) combining ten interviews with platform-wide edit and pageview logs and matched quasi-experimental comparisons; a preprint on a single community.

Summary based on the full text.

## In short

An arXiv preprint (v2, 7 October 2026) by Inhwa Song, Sohyeon Hwang and Manoel Horta Ribeiro (Princeton) with Ted Yoo (Wikimedia Korea). It studies WikiVault, an LLM-powered editing gadget built and governed by the Korean Wikipedia community since May 2025, used mainly to translate articles from other language editions. A handful of experienced editors used it to produce a large share of new content, which shifted their work toward review and pushed the community to set usage limits. The tool also imported English Wikipedia's editorial choices along with its text.

## Key numbers

- 55: editors who used WikiVault in its first ten months (0.12% of active editors)
- 11.10%: share of all new Korean Wikipedia articles created with WikiVault
- 27.00%: share of all added content (bytes) produced with WikiVault, from 0.63% of edits
- 68.1%: how much longer WikiVault-created articles were than matched articles at creation

## What it reports

- In ten months, 55 editors used WikiVault (0.12% of 49,080 active editors), yet they created 7,607 of 68,642 new pages (11.10%) and 220,028,216 of 813,680,042 added bytes (27.00%), from 0.63% of edits.
- Pages edited with WikiVault drew 36.6% of all pageviews in the period.
- From July 2025, 98.6% of edits to existing articles and 99.3% of article creations with the tool used its translation feature.
- Nearly 80% of users made their first WikiVault edit only after more than 1,000 edits: the tool amplified veterans.
- WikiVault-created articles were 68.1% longer than matched articles at creation and 64.9% longer twelve weeks later; they drew 28.4% more pageviews in their first week.
- Existing articles edited with WikiVault received 1.42 times the pageviews of matched controls in the week of the edit and 1.99 times by week 4; they were 8.2 percentage points more likely to receive another human edit the following week.
- As production outpaced review, editors organised review drives; on 23 February 2026 the community introduced a daily per-user limit on WikiVault calls after deliberation on the Village Pump.

## How they got it

Mixed methods: ten semi-structured interviews with editors, descriptive analysis of the full Korean Wikipedia edit and pageview logs over the first ten months, and matched comparisons of WikiVault-touched articles against similar untouched articles, with pre-trend tests. Edits were identified by a dedicated WikiVault edit tag added at Wikimedia Korea's request.

## Limits

- One language community with a specific shortage: Korean Wikipedia has about 10% of English Wikipedia's articles but about 2.5% of its active contributors.
- Ten interviews, with one co-author from the community studied.
- Pageview gains may reflect topic selection by experienced editors, which matching only partly controls.

## Why it matters for companies

This is a rare measured account of a group adopting a generative AI tool on its own terms and governing it through use. The pattern will look familiar inside companies: a few experts produce far more, review becomes the bottleneck, and content carries over assumptions from its source. The conditions the authors credit, legitimacy, explicit limits and traceability through a dedicated tag, map onto internal knowledge bases and documentation. Organisations should budget for review capacity and tag AI-assisted output from the start.

Original: https://arxiv.org/abs/2610.07660