[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"doc-detail-85636-en":3,"doc-seo-85636-105":30,"detail-sidebar-cat-0-en-105":91},{"code":4,"msg":5,"data":6},0,"success",{"doc_id":7,"user_id":8,"nickname":9,"user_avatar":10,"doc_module":4,"category_id":11,"category_name":12,"doc_title":13,"doc_description":14,"doc_content":15,"file_id":16,"file_url":17,"file_type":18,"file_size":19,"view_count":20,"is_deleted":4,"is_public":21,"is_downloadable":21,"audit_status":21,"page_count":22,"language":23,"language_code":24,"site_id":25,"html_lang":24,"table_of_contents":26,"faqs":27,"seo_title":13,"seo_description":14,"update_tm":28,"read_time":29},85636,4398048949847,"Eliana","https://ap-avatar.wpscdn.com/avatar/400002536579ef2da7f?_k=1778318612642679267",8,"Research & Report","Small edits, large models: How Wikipedia advocacy shapes LLM values","Small edits in Wikipedia can measurably steer how large language models discuss animal welfare. Research traces the impact of Pro-Animal Wikipedians (PAW), who added sourced welfare content across 115 pages via 125 edits. Using gradient-based data attribution and retrieval attribution on Llama models, PAW-edited sections dominate high-attribution results for welfare queries. Causal effects are verified via finetuning ablations, reducing perplexity on held-out welfare text by 32%, showing coordinated advocacy can shape downstream model behavior.","arXiv :2606 .24890v 3 [ cs .CL] 12 Jul 2026  \nSmall edits, large models: How Wikipedia advocacy shapes LLM values  \nJasmine Brazilek∗ Maria Navas∗  \nCompassion Aligned Machine Learning (CaML) Independent researcher  \nAlexa Gnauck  \nPro-Animal Wikipedians (PAW)  \nAbstract  \nCan a small group of volunteers shape how AI systems discuss animal welfare, just by editing Wikipedia? We show that they can. Wikipedia appears in nearly every major language model training corpus and is weighted more heavily than webcrawled text. The Pro-Animal Wikipedians (PAW), advocates who add sourced animal welfare content to relevant articles, have made 125 edits across 115 pages.  \nWe trace the influence of these edits on language models with gradient-based data attribution (Bergson, Lucia and Belrose 2026; MAGIC, Ilyas and Engstrom 2025) .  \nWith TrackStar retrieval attribution on Llama 3.1 8B, PAW-edited sections make up 68% of the highest-attributed documents for animal welfare queries (p \u003C 0.0001), but only 52% for unrelated queries about the same entities (p = 0 .53) . The model links PAW content specifically to animal welfare. MAGIC counterfactual influence estimation on Llama 3.2 1B, repeated across five random training-order seeds, shows the same pattern more sharply: in every seed, the ten most influential documents for animal welfare queries are all PAW edits, while the same ranking for general queries sits at chance. Mean PAW influence exceeds mean control influence at p \u003C 0.0001 in every seed, an effect 6–30 × larger than on general queries, and leave-subset-out validation gives Spearman ρ = 1 .00 across all ten runs. A finetuning ablation confirms the attributions are causal: training on PAW content cuts perplexity on held-out animal welfare text by 32%, while control training only helps on control text. A small, coordinated Wikipedia editing campaign therefore measurably shapes how language models handle the topics those edits address.  \n1 Introduction  \nLanguage models are becoming a primary way people get information. For advocacy organizations, this raises a practical question: can you influence what these models say about your cause? We show that Wikipedia editing is one way to do it. Wikipedia appears in nearly every major training dataset used to build language models (The Pile, RedPajama, Dolma, and others) and is given more weight than web-crawled sources because of its quality and breadth (Gao et al., 2020; Soldaini et al., 2024; Weber et al., 2024) . This means that what Wikipedia says about a topic feeds directly into what language models say about it.  \nThis creates an opportunity. Wikipedia has always been a place where small groups of dedicated editors shape how topics are presented (Yasseri et al., 2012) . Organized editing campaigns by civil society groups, political movements, and PR firms are common, ranging from state-backed takeovers  \nof entire language editions to corporate reputation management (DiStaso, 2013; Kharazian et al.,∗Equal contribution.  \nPreprint.  \n2024) . What has gone mostly unnoticed is that these editing efforts now have a second effect: they shape the training data of language models, which in turn shape how millions of people get information. For advocacy organizations with limited budgets, this raises a concrete question: does editing Wikipedia actually change what AI systems say?  \nWe study this phenomenon in the domain of animal welfare. The Pro-Animal Wikipedians (PAW) have conducted sustained editing of Wikipedia articles related to the use and perceptions of animal exploitation in articles about fast-food, animal sentience and politics. Using revision histories, we isolate the textual changes attributable to this group and quantify their downstream influence on language model behavior using gradient-based data attribution.  \nWe employ Bergson (Lucia and Belrose, 2026), an open-source library implementing TrackStar (Chang et al., 2024), to trace how individual Wikipedia edits influence mod","cbCaibM7le4tuEia","https://ap.wps.com/l/cbCaibM7le4tuEia","pdf",196004,3,1,10,"English","en",105,"# Abstract\n# Introduction\n## Why this matters for advocacy","[{\"question\":\"What question does the study investigate about Wikipedia and language models?\",\"answer\":\"The study asks whether a small group of Wikipedia volunteers can influence how AI systems discuss animal welfare, by editing Wikipedia content that is used in model training.\"},{\"question\":\"Who are the Pro-Animal Wikipedians (PAW) and what did they do?\",\"answer\":\"PAW is a loose coalition of editors who add sourced animal-welfare information to relevant Wikipedia articles. They made 125 edits across 115 pages focused on animal exploitation, sentience, and related topics.\"},{\"question\":\"How does the study test whether PAW edits causally affect model behavior?\",\"answer\":\"It traces influence using gradient-based data attribution and retrieval attribution, then confirms causality with finetuning ablations: training on PAW content reduces perplexity on held-out animal welfare text by 32%, while control training does not.\"}]",1784205193,25,{"code":4,"msg":31,"data":32},"ok",{"site_id":25,"language":24,"slug":33,"title":13,"keywords":34,"description":14,"schema_data":35,"social_meta":86,"head_meta":88,"extra_data":90,"updated_unix":28},"small-edits-large-models-how-wikipedia-advocacy-shapes-llm-values","",{"@graph":36,"@context":85},[37,53,68],{"@type":38,"itemListElement":39},"BreadcrumbList",[40,44,48,50],{"item":41,"name":42,"@type":43,"position":21},"https://docshare.wps.com","Home","ListItem",{"item":45,"name":46,"@type":43,"position":47},"https://docshare.wps.com/document/","Document",2,{"item":49,"name":12,"@type":43,"position":20},"https://docshare.wps.com/document/research-report/",{"item":51,"name":13,"@type":43,"position":52},"https://docshare.wps.com/document/small-edits-large-models-how-wikipedia-advocacy-shapes-llm-values/85636/",4,{"url":51,"name":13,"@type":54,"author":55,"headline":13,"publisher":57,"fileFormat":60,"inLanguage":24,"description":14,"dateModified":61,"datePublished":62,"encodingFormat":60,"isAccessibleForFree":63,"interactionStatistic":64},"DigitalDocument",{"name":9,"@type":56},"Person",{"url":41,"name":58,"@type":59},"DocShare","Organization","application/pdf","2026-07-25","2026-07-16",true,{"@type":65,"interactionType":66,"userInteractionCount":20},"InteractionCounter",{"@type":67},"ViewAction",{"@type":69,"mainEntity":70},"FAQPage",[71,77,81],{"name":72,"@type":73,"acceptedAnswer":74},"What question does the study investigate about Wikipedia and language models?","Question",{"text":75,"@type":76},"The study asks whether a small group of Wikipedia volunteers can influence how AI systems discuss animal welfare, by editing Wikipedia content that is used in model training.","Answer",{"name":78,"@type":73,"acceptedAnswer":79},"Who are the Pro-Animal Wikipedians (PAW) and what did they do?",{"text":80,"@type":76},"PAW is a loose coalition of editors who add sourced animal-welfare information to relevant Wikipedia articles. They made 125 edits across 115 pages focused on animal exploitation, sentience, and related topics.",{"name":82,"@type":73,"acceptedAnswer":83},"How does the study test whether PAW edits causally affect model behavior?",{"text":84,"@type":76},"It traces influence using gradient-based data attribution and retrieval attribution, then confirms causality with finetuning ablations: training on PAW content reduces perplexity on held-out animal welfare text by 32%, while control training does not.","https://schema.org",{"og:url":51,"og:type":87,"og:title":13,"og:site_name":58,"og:description":14},"article",{"robots":89,"canonical":51},"index,follow",{"doc_id":7,"site_id":25},{"code":4,"msg":5,"data":92},[93,97,101,105,110,115,120,123,128,131,134],{"id":21,"doc_module":4,"doc_module_name":46,"category_name":94,"show_sort_weight":95,"slug":96},"Story & Novel",90,"story-novel",{"id":47,"doc_module":4,"doc_module_name":46,"category_name":98,"show_sort_weight":99,"slug":100},"Literature",80,"literature",{"id":52,"doc_module":4,"doc_module_name":46,"category_name":102,"show_sort_weight":103,"slug":104},"Exam",70,"exam",{"id":106,"doc_module":4,"doc_module_name":46,"category_name":107,"show_sort_weight":108,"slug":109},5,"Comic",60,"comic",{"id":111,"doc_module":4,"doc_module_name":46,"category_name":112,"show_sort_weight":113,"slug":114},6,"Technology",50,"technology",{"id":116,"doc_module":4,"doc_module_name":46,"category_name":117,"show_sort_weight":118,"slug":119},7,"Healthcare",40,"healthcare",{"id":11,"doc_module":4,"doc_module_name":46,"category_name":12,"show_sort_weight":121,"slug":122},30,"research-report",{"id":124,"doc_module":4,"doc_module_name":46,"category_name":125,"show_sort_weight":126,"slug":127},9,"Religion & Spirituality",20,"religion-spirituality",{"id":126,"doc_module":4,"doc_module_name":46,"category_name":129,"show_sort_weight":126,"slug":130},"World Cup","world-cup",{"id":22,"doc_module":4,"doc_module_name":46,"category_name":132,"show_sort_weight":22,"slug":133},"Lifestyle","lifestyle",{"id":135,"doc_module":4,"doc_module_name":46,"category_name":136,"show_sort_weight":106,"slug":137},19,"General","general"]