Small Wikipedia Edits, Big Impact: Study Shows How Advocacy Shapes LLM Values

A newly published study on arXiv demonstrates that a small group of motivated Wikipedia editors can shape how AI systems discuss sensitive topics. The research tracked a group called Pro-Animal Wikipedians (PAW), who made 125 edits across 115 pages to add sourced animal welfare content.

The paper shows that because Wikipedia appears in nearly every major LLM training dataset and is weighted more heavily than general web text, these targeted edits measurably influenced model outputs on animal welfare topics. The team used gradient-based data attribution (Bergson) to trace the effect.

Published on arXiv cs.CL on June 25, 2026, the findings raise important questions about the power dynamics of AI training data curation and who gets to shape the values embedded in large language models.