Wikimedia-l, Hello. I am pleased to share four quick ideas for improving Wikipedia with AI systems. Firstly, AI systems could answer questions about potential edits by processing, comparing, and answering questions about two revisions of an article. These questions might involve whether content, e.g., potential edits, had a neutral point-of-view or otherwise conformed with Wikipedia’s content guidelines. Wikipedia develops at a rate of about two edits per second [1]. Accordingly, administrators and moderators should be able to indicate and to select which articles to have their potential revisions or edits processed by one or more AI-powered content-processing pipelines. Administrators and moderators could receive numerically weighted or otherwise prioritized messages or alerts from AI-powered content-processing pipelines to review potential edits and edited articles. Secondly, AI systems could assist in moderating talk pages or discussion forums. Thirdly, new wiki templates could be useful for obtaining and inserting content from interoperating AI systems. Computationally, such templates could be processed either upon: (1) each page view, (2) users pressing a button on such pages, (3) each article edit or revision, (4) each hour, day, week, month, year, or other interval of time, (5) whenever the interoperating AI system were upgraded or switched, (6) combinations of these or other factors. With respect to specifying desired formatting to LLM's with respect to their outputs, one can observe the Guidance project [2]. Interestingly, resultant HTML markup could be semantically distinguished as being AI-generated. <p>The current temperature in Seattle, WA, is <span class="ai">42</span>.</p> Fourthly, if AI-powered templates are to be explored and supported, related user experiences could be considered. These might involve Wikipedia users being able to visually detect, easily annotate, and, potentially, correct such content. Content generated by AI via templates could be visually styled with a symbol or a glyph following it, resembling how external hyperlinks are followed by small visual symbols. Hovering over these symbols or glyphs could visually highlight the entireties of the contents that were AI-generated. In theory, users could (right-)click upon such symbols or glyphs (or upon AI-generated content itself) to view a special context menu. This context menu could provide means for indicating whether the AI-generated content was correct or whether it was incorrect. There could be a textbox on the context menu for users to easily enter corrections into. Users' feedback with respect to AI-generated content could be sent to interoperating AI systems.
Best regards, Adam Sobieski [1] https://en.wikipedia.org/wiki/Wikipedia:Statistics [2] https://github.com/guidance-ai/guidance