AI-assisted Patrolling: Bridging Discord, Pywikibot, and LLMs to patrol small wikis
Speakers
Alongside my almost 20 years (and counting) as a primary school teacher, I joined the team at Wikimedia CH to look after projects for education, and HEP Valais to lend a hand with digital education training.
In 2016, I launched Le Dico des Ados, a collaborative dictionary written by kids that currently hosts over 6,700 definitions. The rest of the time, I tinker with stuff for schools, anything from coding moderation bots to designing gamified webapps. In short, I make tools for the classroom.
Enseignant en primaire depuis bientôt 20 ans (et toujours avec un pied dans l'enseignement) j'ai rejoint l'équipe de Wikimedia CH pour m'occuper des projets éducatifs, et la HEP Valais pour donner un coup de main sur la formation en éducation numérique.
En 2016, j'ai lancé Le Dico des Ados, un dictionnaire collaboratif écrit par les jeunes qui compte aujourd'hui plus de 6 700 définitions. Le reste du temps, je bidouille des trucs pour les écoles : ça va des bots de modération à la conception de webapps. Bref, je fais des outils pour le terrain.
Abstract
Small wikis like Dico des Ados (French Wiktionary for teenagers) face a unique challenge: high vandalism rates but few patrollers.
To solve this, I built Patrouille-IA, a bot that integrates Pywikibot with Discord and the Claude-Haiku LLM. Instead of relying on rigid regex, the bot uses AI to analyze diffs semantically, identifying vandalism, checking tone, and suggesting categories. It reports directly to a Discord channel where human moderators can revert edits with a single click.
This talk demonstrates how combining modern LLMs with legacy MediaWiki tools can reduce patroller burnout and improve reliability on under-resourced wikis.
Additional information
Fiabilité (Reliability): This is the core of the project. By using an LLM to scan recent changes, the bot catches subtle vandalism and "nonsense" edits that standard filters miss, directly increasing the reliability of the dictionary for its young readers. Équité (Equity): Small communities often burn out trying to keep up with vandalism. This tool democratizes advanced patrolling. It allows a small team of volunteers to manage a workload that usually requires a large army of admins, ensuring that smaller wikis can survive and thrive alongside giants like Wikipedia.
This session is tailored for: • Bot developers and tool maintainers who want to see a practical implementation of an LLM API within the Pywikibot framework. • Small wiki admins who struggle with patrolling and need force-multipliers. • The Trust & Safety community interested in how AI can be used as a "co-pilot" for human moderation rather than an autonomous judge.
Everyone can participate in this session