[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"doc-detail-83480-en":3,"doc-seo-83480-105":30,"detail-sidebar-cat-0-en-105":90},{"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},83480,687197100911,"Himbo","https://ap-avatar.wpscdn.com/avatar/a000239b6f1da00475?x-image-process=image/resize,m_fixed,w_180,h_180&k=1782698725881665579",8,"Research & Report","Child Safety in Generative AI: An Expert-Guided and Incident-Grounded Evaluation Framework","Generative AI increasingly serves children and adolescents, yet existing safety evaluation frameworks largely target general users and miss harms specific to younger populations. The document proposes a child-centered evaluation framework that combines expert-guided risk factors with real-world AI incident data. It derives child-safety hazard categories from expert guidelines and incident databases, then constructs a synthetic test set for model evaluation. A case study applies the method to education, testing three Llama Guard models and finding difficulty detecting education-related unsafe prompts.","Child Safety in Generative AI: An Expert-Guided and Incident-Grounded Evaluation Framework  \nHaein Kong  \n[haein.kong@rutgers.edu](haein.kong@rutgers.edu)  \nRutgers University  \nNew Brunswick, New Jersey, USA  \narXiv :2607 .00395v 1 [ cs .HC] 1 Jul 2026  \nAbstract  \nAs generative AI is increasingly used by children and adolescents, there is a growing need for risk evaluation frameworks that account for child-specific harms. However, most existing safety evaluation frameworks focus on general user populations, often overlooking risks unique to younger users. To address this gap, we propose an evaluation framework that integrates expert-guided risk factors with real-world AI incident data for child safety. The framework identifies hazard categories from expert guidelines and AI incident databases and uses this information to construct a synthetic test set for model evaluation. Particularly, we apply the framework to the education domain and evaluate three Llama Guard models on their ability to detect unsafe user prompts. Our results show that current Llama Guard models struggle to identify education-related unsafe user prompts. We conclude by discussing how future work can extend the evaluation to additional risk categories and incorporate domain experts throughout the evaluation pipeline.  \nCCS Concepts  \n• Computing methodologies → Natural language processing;  \n• Human-centered computing → Interaction paradigms.  \nKeywords  \nAI Safety, Large Language Model, LLM Evaluation, Child Safety  \n1 Introduction  \nAs generative models have rapidly advanced and been adopted across domains, substantial efforts have been made to evaluate their safety. Prior work has proposed comprehensive risk taxonomies through systematic reviews [17], identifying a large number of AI risk categories and analyzing their causal factors [15] . Researchers have also developed diverse benchmarks such as TrustfulQA [12] or HarmBench [13] to empirically evaluate the safety of large language models (LLMs). However, most existing benchmarks in AI safety primarily focus on general user populations, typically adults, not covering age-specific risks or harms.  \nThis practice could pose safety risks for certain populations, especially youth. According to a recent national survey, 72% of adolescents in the United States have used AI companions [14] . Given the vulnerabilities associated with this developmental stage, it is critical to evaluate generative AI systems to protect children from potential harm. While recent regulatory efforts have begun to address AI and child safety, there is still a lack of unified evaluation guidelines or frameworks.  \nThere have been a few attempts to build a child-centered evaluation framework in the literature. They proposed a comprehensive risk taxonomy for youth safety by analyzing multiple resources [18]  \nor offered a benchmark based on interviews and chat logs [9]. While these efforts contribute to AI evaluation for child safety, several limitations remain. For instance, a proposed taxonomy has not been extended into a benchmark dataset for model evaluation. In addition, existing benchmarks lack the flexibility to adapt to emerging risks, and domain experts are rarely involved in the development of evaluation frameworks.  \nTo address these gaps, we introduce a framework that aims to enhance real-world awareness and incorporate experts’ suggestions. This framework includes a flexible, extensible method for generating a synthetic test set that can respond rapidly to real-world AI incidents. To do this, this framework uses two resources: expertproposed child safety guidelines and AI incident databases. We first construct a taxonomy of child safety risk categories informed by expert guidance, and then leverage incident data to build a synthetic test set that reflects realistic, harmful scenarios. As a case study, we apply the framework to the education domain and evaluate three Llama Guard models, fine-tuned for content safety class","cbCaisiubbB8qSHG","https://ap.wps.com/l/cbCaisiubbB8qSHG","pdf",495368,3,1,4,"English","en",105,"# Introduction\n# Related Work\n## AI Safety Evaluation","[{\"question\":\"Why do existing AI safety evaluation frameworks fall short for child safety?\",\"answer\":\"Most existing benchmarks focus on general user populations, often adults, and do not cover age-specific risks or harms relevant to children and adolescents.\"},{\"question\":\"How does the proposed evaluation framework incorporate expert guidance and incident data?\",\"answer\":\"It builds a taxonomy of child-safety risk categories from expert guidelines, then uses AI incident databases to generate a synthetic test set reflecting realistic harmful scenarios.\"},{\"question\":\"What did the case study in the education domain reveal about Llama Guard models?\",\"answer\":\"When evaluating three Llama Guard models on unsafe education-related user prompts, the results indicate that current models struggle to identify these unsafe prompts effectively.\"}]",1784188310,10,{"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":85,"head_meta":87,"extra_data":89,"updated_unix":28},"child-safety-in-generative-ai-an-expert-guided-and-incident-grounded-evaluation-framework","",{"@graph":36,"@context":84},[37,52,67],{"@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":22},"https://docshare.wps.com/document/child-safety-in-generative-ai-an-expert-guided-and-incident-grounded-evaluation-framework/83480/",{"url":51,"name":13,"@type":53,"author":54,"headline":13,"publisher":56,"fileFormat":59,"inLanguage":24,"description":14,"dateModified":60,"datePublished":61,"encodingFormat":59,"isAccessibleForFree":62,"interactionStatistic":63},"DigitalDocument",{"name":9,"@type":55},"Person",{"url":41,"name":57,"@type":58},"DocShare","Organization","application/pdf","2026-07-26","2026-07-16",true,{"@type":64,"interactionType":65,"userInteractionCount":20},"InteractionCounter",{"@type":66},"ViewAction",{"@type":68,"mainEntity":69},"FAQPage",[70,76,80],{"name":71,"@type":72,"acceptedAnswer":73},"Why do existing AI safety evaluation frameworks fall short for child safety?","Question",{"text":74,"@type":75},"Most existing benchmarks focus on general user populations, often adults, and do not cover age-specific risks or harms relevant to children and adolescents.","Answer",{"name":77,"@type":72,"acceptedAnswer":78},"How does the proposed evaluation framework incorporate expert guidance and incident data?",{"text":79,"@type":75},"It builds a taxonomy of child-safety risk categories from expert guidelines, then uses AI incident databases to generate a synthetic test set reflecting realistic harmful scenarios.",{"name":81,"@type":72,"acceptedAnswer":82},"What did the case study in the education domain reveal about Llama Guard models?",{"text":83,"@type":75},"When evaluating three Llama Guard models on unsafe education-related user prompts, the results indicate that current models struggle to identify these unsafe prompts effectively.","https://schema.org",{"og:url":51,"og:type":86,"og:title":13,"og:site_name":57,"og:description":14},"article",{"robots":88,"canonical":51},"index,follow",{"doc_id":7,"site_id":25},{"code":4,"msg":5,"data":91},[92,96,100,104,109,114,119,122,127,130,133],{"id":21,"doc_module":4,"doc_module_name":46,"category_name":93,"show_sort_weight":94,"slug":95},"Story & Novel",90,"story-novel",{"id":47,"doc_module":4,"doc_module_name":46,"category_name":97,"show_sort_weight":98,"slug":99},"Literature",80,"literature",{"id":22,"doc_module":4,"doc_module_name":46,"category_name":101,"show_sort_weight":102,"slug":103},"Exam",70,"exam",{"id":105,"doc_module":4,"doc_module_name":46,"category_name":106,"show_sort_weight":107,"slug":108},5,"Comic",60,"comic",{"id":110,"doc_module":4,"doc_module_name":46,"category_name":111,"show_sort_weight":112,"slug":113},6,"Technology",50,"technology",{"id":115,"doc_module":4,"doc_module_name":46,"category_name":116,"show_sort_weight":117,"slug":118},7,"Healthcare",40,"healthcare",{"id":11,"doc_module":4,"doc_module_name":46,"category_name":12,"show_sort_weight":120,"slug":121},30,"research-report",{"id":123,"doc_module":4,"doc_module_name":46,"category_name":124,"show_sort_weight":125,"slug":126},9,"Religion & Spirituality",20,"religion-spirituality",{"id":125,"doc_module":4,"doc_module_name":46,"category_name":128,"show_sort_weight":125,"slug":129},"World Cup","world-cup",{"id":29,"doc_module":4,"doc_module_name":46,"category_name":131,"show_sort_weight":29,"slug":132},"Lifestyle","lifestyle",{"id":134,"doc_module":4,"doc_module_name":46,"category_name":135,"show_sort_weight":105,"slug":136},19,"General","general"]