Supporting patrollers with collaborative auditing. Introducing the Judgment and Dialogue Engine.
We're bringing the Bold-Revert-Discuss cycle to patrolling, and to oversight of emerging AIs. JADE provides a content and a talk namespace for coordinating patrolling judgments, and for meta-curation of patrol actions.
The most exciting and unique aspect of wikis is the ability for contributors to edit and discuss each other's content. We can think of these actions as feedback loops, one between editors and content, and the other between editors and one another. Wikis have many of these loops, knitting together communities and subcommunities.
With increasing AI use in quality control work, it's important that we face a new type of feedback loop, which I'm calling a "conformity cycle". We're assuming the existence of a positive feedback loop between editors and AIs, which as currently designed may be serving to increase the power of an already dominant status quo.
By introducing two new namespaces, "Jade" and "Jade_talk", we're hoping to give more structure to patrolling work, and to AI auditing. Pages under "Jade" will correspond to a specific wiki entity to be reviewed, for example a revision's content, or the diff of a change. The "Jade_talk" pages are reserved for discussion of each judgment.
- As a slide deck: https://docs.google.com/presentation/d/1DJdp98jVg7BRKhfmy-Qfv-eyjhSIhc0p4KXbYhK1QSc/edit
- Position paper on the political economy of AI auditing: https://commons.wikimedia.org/wiki/File:Jade-pol-econ.pdf
- Project home page: JADE