Can Source Reliability Be Predicted on Wikipedia? Evidence from Feature Analysis
🎥 Session recording: [https://commons.wikimedia.org/wiki/File:Wikimania_2026_%E2%80%93_Can_Source_Reliability_Be_Predicted_on_Wikipedia%3F_Evidence_from_Feature_Analysis.webm](https://commons.wikimedia.org/wiki/File:Wikimania_2026_%E2%80%93_Can_Source_Reliability_Be_Predicted_on_Wikipedia%3F_Evidence_from_Feature_Analysis.webm)
Wikipedia depends on reliable sources, but what kinds of signals actually help predict whether a source will later be judged reliable or problematic?
Drawing on recent work on language-agnostic modeling of source reliability, I will highlight which families of features appear most useful for prediction across Wikipedia language editions. These include signals derived from a source’s editing history, patterns of reuse, persistence over time, and contextual interactions with editors and articles. Rather than presenting a full system overview, the talk will focus on what these features reveal about how reliability is expressed in practice on Wikipedia.
The main contribution of the talk is to connect technical feature design with Wikimedia’s community processes: if some signals are consistently more informative than others, what does that tell us about how editors collectively identify trustworthy or risky references? By focusing on features rather than model outputs alone, the session provides a compact, interpretable entry point for researchers, tool builders, and community members interested in verifiability and knowledge integrity.