How Wikipedia Edits Shape AI's Understanding of Animal Welfare

A team of researchers from the Pro-Animal Wikipedians (PAW) demonstrated that even small edits to Wikipedia can significantly impact how AI models discuss animal welfare. By making 125 edits across 115 pages, they showed that Wikipedia, which is heavily weighted in AI training datasets, can shape the values and information presented by AI models. The study used gradient-based data attribution techniques (Bergson; MAGIC) and TrackStar retrieval attribution to trace how these edits influenced AI behavior, revealing the profound impact of curated content.

This research matters because it shows how a small group of dedicated editors can influence the information that AI models learn and present. Think of it like a librarian carefully selecting which books to place on the front shelf — the choices they make can shape what people read and learn. For everyday users, this means that the information you encounter in AI-generated content is not just a result of random data but is shaped by intentional edits and advocacy.

If you're curious about how AI models learn from Wikipedia, you can explore the Pro-Animal Wikipedians' edits yourself. Visit Wikipedia and search for articles related to animal welfare. You might notice how these edits have subtly shaped the information presented, giving you a glimpse into the power of curated content in AI training.