Wikimedia tries AI to catch bad edits

The Wikimedia Foundation has a problem with new editors: the tools it created to help with quality control was hostile to newbies, and rejected too much of their work.

A new artificial intelligence (AI) tool created for Wikipedia quality assessments is therefore designed to help participants assess whether their edits are likely to damage the article they're working on – but instead of treating the AI as a black box, the foundation has open-sourced the software under the MIT license.

As Aaron Halfaker and Dario Taraborelli write explaining the Objective Revision Evaluation Service (ORES), the idea is to train models against assessments of edits and articles currently made by humans.

They want to reverse the ill-effects of tools like Huggle, STiki and ClueBot, which were good at helping maintain quality control, but came at a cost. They made it hard for new participants to get edits accepted, because they “encourage the rejection of all new editors’ changes as though they were made in bad faith”.

ORES is presented to editors under a RESTful API. The API lets Wikipedia reviewers and editors ask the service whether a particular edit is going to be damaging, they write: “ORES allows you to specify a project (e.g. English Wikipedia), a model (e.g. the damage detection model), and one or more revisions. The API returns an easily consumable response in JSON format.”

ORES in action (click to embiggen)

The service supports 14 languages at the moment, and claims responses of 50 ms for already-scored revisions and 100 ms for un-scored revisions.

As the ORES page explains, the effects should go far beyond Wikipedia. The API means others can use the quality-bot on their own wikis.

The model currently returns three responses to a query: “damaging”, “goodfaith”, and “reverted” (the latter telling you whether your edit is likely to be reverted in the future, in which case don't bother making it).

Refreshingly, Halfaker and Taraborelli swim somewhat against the tide of bot-worship that pervades the tech sector's AI boosterism. Artificial intelligence doesn't somehow magically remove human bias, they say, but rather hides it.

“While artificial intelligence may prove essential for solving problems at Wikipedia’s scale, algorithms that replicate subjective judgements can also dehumanize and subjugate people and obfuscate inherent biases."

“An algorithm that flags edits as subjectively 'good' or 'bad', with little room for scrutiny or correction, changes the way those contributions and the people who made them are perceived,” they add.

To deal with the risk that ORES carries bias, the post notes that detecting and fixing it will be part of the ongoing development: "We’re developing strategies for detecting bias in the models; our scores and fitness measures are freely available and we’re actively collecting feedback on mistakes."

The source code of ORES is on Github, and the Wikimedia post also points to the system's performance statistics, project documentation, and underlying data.

ORES is part of the foundation's Revision Scoring as a Service research project (here). ®