Small edits, large models: How Wikipedia advocacy shapes LLM values

A research study demonstrates that a small group of Wikipedia editors advocating for animal welfare has measurably shaped how large language models discuss the topic, with their edits appearing in 68% of the most relevant documents for animal welfare queries. Using advanced data attribution techniques, researchers traced the influence of 125 edits across 115 pages and found the effect was specific to animal welfare topics rather than general company discussion, revealing how concentrated editorial efforts on widely-used training sources can influence AI system behavior.

This research exposes a significant but underexplored vulnerability in how large language models develop their training distributions: concentrated editorial campaigns on high-authority sources like Wikipedia can meaningfully shape model outputs without widespread coordination or technical manipulation. The Pro-Animal Wikipedians' 125 edits—a modest intervention—demonstrably increased the salience of animal welfare perspectives in LLM responses, with statistical rigor showing the effect is topic-specific rather than spurious. The finding matters because Wikipedia's disproportionate weighting in training datasets means small groups can punch above their weight in shaping AI values.

This research sits at the intersection of AI governance, content moderation, and soft influence. As LLMs become primary information sources for millions of users, the question of who shapes their training data gains urgency. Wikipedia's open-edit model has long enabled advocacy campaigns, but its influence on AI systems represents a new vector for ideological or commercial shaping of machine learning outputs. The technical depth of this study—using gradient-based attribution and counterfactual estimation across multiple seeds—provides evidence that these effects are real and reproducible, not statistical noise.

For AI developers and policymakers, the implications are dual-edged. On one hand, this demonstrates that coordinated communities can improve AI behavior on topics they care about. On the other hand, it reveals that unvetted advocacy on Wikipedia could subtly bias LLM outputs in ways users don't recognize. The effect size (6-30x larger on targeted topics) suggests this is not marginal. Expect increased scrutiny of Wikipedia's role in LLM training pipelines and potential efforts by various groups to influence content strategically.

• →A small group of 125 Wikipedia edits on animal welfare measurably influenced how Llama models respond to animal welfare queries, with 68% attribution concentration.
• →The editorial influence was topic-specific, affecting animal welfare discussion without biasing general company-related queries, suggesting targeted rather than broad ideological shifts.
• →Wikipedia's heavy weighting in LLM training datasets means concentrated editing campaigns can shape AI outputs without technical intervention or model fine-tuning.
• →Multiple validation methods (TrackStar, MAGIC, leave-subset-out) confirmed the effect size was 6-30 times larger on targeted topics than unrelated queries.
• →This finding raises governance questions about how open-source training data can be influenced by advocacy groups and the role of content platforms in shaping AI values.